var bibbase_data = {"data":"\"Loading..\"\n\n
\n\n \n\n \n\n \n \n\n \n\n \n \n\n \n\n \n
\n generated by\n \n \"bibbase.org\"\n\n \n
\n \n\n
\n\n \n\n\n
\n\n Excellent! Next you can\n create a new website with this list, or\n embed it in an existing web page by copying & pasting\n any of the following snippets.\n\n
\n JavaScript\n (easiest)\n
\n \n <script src=\"https://bibbase.org/show?bib=https%3A%2F%2Fbibbase.org%2Fzotero-group%2Fndrmyrvtl%2F5786256&jsonp=1&jsonp=1\"></script>\n \n
\n\n PHP\n
\n \n <?php\n $contents = file_get_contents(\"https://bibbase.org/show?bib=https%3A%2F%2Fbibbase.org%2Fzotero-group%2Fndrmyrvtl%2F5786256&jsonp=1\");\n print_r($contents);\n ?>\n \n
\n\n iFrame\n (not recommended)\n
\n \n <iframe src=\"https://bibbase.org/show?bib=https%3A%2F%2Fbibbase.org%2Fzotero-group%2Fndrmyrvtl%2F5786256&jsonp=1\"></iframe>\n \n
\n\n

\n For more details see the documention.\n

\n
\n
\n\n
\n\n This is a preview! To use this list on your own web site\n or create a new web site from it,\n create a free account. The file will be added\n and you will be able to edit it in the File Manager.\n We will show you instructions once you've created your account.\n
\n\n
\n\n

To the site owner:

\n\n

Action required! Mendeley is changing its\n API. In order to keep using Mendeley with BibBase past April\n 14th, you need to:\n

    \n
  1. renew the authorization for BibBase on Mendeley, and
    2. \n
  2. update the BibBase URL\n in your page the same way you did when you initially set up\n this page.\n
    3. \n
\n

\n\n

\n \n \n Fix it now\n

\n
\n\n
\n\n\n
\n \n \n
\n
\n  \n 2025\n \n \n (18)\n \n \n
\n
\n \n \n
\n \n\n \n \n \n \n \n \n Separation-Based Distance Measures for Causal Graphs.\n \n \n \n \n\n\n \n Wahl, J.; and Runge, J.\n\n\n \n\n\n\n In Li, Y.; Mandt, S.; Agrawal, S.; and Khan, E., editor(s), Proceedings of The 28th International Conference on Artificial Intelligence and Statistics, volume 258, of Proceedings of Machine Learning Research, pages 3412–3420, May 2025. \n \n\n\n\n
\n\n\n\n \n \n \"Separation-BasedPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{wahl_separation-based_2025,\n\tseries = {Proceedings of {Machine} {Learning} {Research}},\n\ttitle = {Separation-{Based} {Distance} {Measures} for {Causal} {Graphs}},\n\tvolume = {258},\n\turl = {https://proceedings.mlr.press/v258/wahl25b.html},\n\tbooktitle = {Proceedings of {The} 28th {International} {Conference} on {Artificial} {Intelligence} and {Statistics}},\n\tauthor = {Wahl, Jonas and Runge, Jakob},\n\teditor = {Li, Yingzhen and Mandt, Stephan and Agrawal, Shipra and Khan, Emtiyaz},\n\tmonth = may,\n\tyear = {2025},\n\tpages = {3412--3420},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Developing Effective and Value-Aligned AI Tools for Journalists: 12 Critical Questions to Reflect upon.\n \n \n \n\n\n \n Cordero, J. M.; Henn, T.; Holtel, F.; Sánchez Gómez, J. Á; Arenas, D.; Šipka, A.; and Vollmer, S.\n\n\n \n\n\n\n Journalism Practice,1–20. 2025.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{cordero_developing_2025,\n\ttitle = {Developing {Effective} and {Value}-{Aligned} {AI} {Tools} for {Journalists}: 12 {Critical} {Questions} to {Reflect} upon},\n\tjournal = {Journalism Practice},\n\tauthor = {Cordero, José Miguel and Henn, Theresa and Holtel, Frederik and Sánchez Gómez, José Á and Arenas, Diego and Šipka, Andrea and Vollmer, Sebastian},\n\tyear = {2025},\n\tpages = {1--20},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n X Hacking: The Threat of Misguided AutoML.\n \n \n \n\n\n \n Sharma, R.; Redyuk, S.; Mukherjee, S.; Sipka, A.; Vollmer, S.; and Selby, D.\n\n\n \n\n\n\n Accepted at ICML. 2025.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{sharma_x_2025,\n\ttitle = {X {Hacking}: {The} {Threat} of {Misguided} {AutoML}},\n\tjournal = {Accepted at ICML},\n\tauthor = {Sharma, Rahul and Redyuk, Sergey and Mukherjee, Sumantrak and Sipka, Andrea and Vollmer, Sebastian and Selby, David},\n\tyear = {2025},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Cross-Refine: Improving Natural Language Explanation Generation by Learning in Tandem.\n \n \n \n \n\n\n \n Wang, Q.; Anikina, T.; Feldhus, N.; Ostermann, S.; Möller, S.; and Schmitt, V.\n\n\n \n\n\n\n In Rambow, O.; Wanner, L.; Apidianaki, M.; Al-Khalifa, H.; Eugenio, B. D.; and Schockaert, S., editor(s), Proceedings of the 31st International Conference on Computational Linguistics, pages 1150–1167, Abu Dhabi, UAE, January 2025. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"Cross-Refine:Paper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{wang_cross-refine_2025,\n\taddress = {Abu Dhabi, UAE},\n\ttitle = {Cross-{Refine}: {Improving} {Natural} {Language} {Explanation} {Generation} by {Learning} in {Tandem}},\n\tshorttitle = {Cross-{Refine}},\n\turl = {https://aclanthology.org/2025.coling-main.77/},\n\tabstract = {Natural language explanations (NLEs) are vital for elucidating the reasoning behind large language model (LLM) decisions. Many techniques have been developed to generate NLEs using LLMs. However, like humans, LLMs might not always produce optimal NLEs on first attempt. Inspired by human learning processes, we introduce Cross-Refine, which employs role modeling by deploying two LLMs as generator and critic, respectively. The generator outputs a first NLE and then refines this initial explanation using feedback and suggestions provided by the critic. Cross-Refine does not require any supervised training data or additional training. We validate Cross-Refine across three NLP tasks using three state-of-the-art open-source LLMs through automatic and human evaluation. We select Self-Refine (Madaan et al., 2023) as the baseline, which only utilizes self-feedback to refine the explanations. Our findings from automatic evaluation and a user study indicate that Cross-Refine outperforms Self-Refine. Meanwhile, Cross-Refine can perform effectively with less powerful LLMs, whereas Self-Refine only yields strong results with ChatGPT. Additionally, we conduct an ablation study to assess the importance of feedback and suggestions. Both of them play an important role in refining explanations. We further evaluate Cross-Refine on a bilingual dataset in English and German.},\n\turldate = {2025-05-22},\n\tbooktitle = {Proceedings of the 31st {International} {Conference} on {Computational} {Linguistics}},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Wang, Qianli and Anikina, Tatiana and Feldhus, Nils and Ostermann, Simon and Möller, Sebastian and Schmitt, Vera},\n\teditor = {Rambow, Owen and Wanner, Leo and Apidianaki, Marianna and Al-Khalifa, Hend and Eugenio, Barbara Di and Schockaert, Steven},\n\tmonth = jan,\n\tyear = {2025},\n\tpages = {1150--1167},\n}\n\n\n\n\n\n\n\n\n\n\n\n
\n
\n\n\n
\n Natural language explanations (NLEs) are vital for elucidating the reasoning behind large language model (LLM) decisions. Many techniques have been developed to generate NLEs using LLMs. However, like humans, LLMs might not always produce optimal NLEs on first attempt. Inspired by human learning processes, we introduce Cross-Refine, which employs role modeling by deploying two LLMs as generator and critic, respectively. The generator outputs a first NLE and then refines this initial explanation using feedback and suggestions provided by the critic. Cross-Refine does not require any supervised training data or additional training. We validate Cross-Refine across three NLP tasks using three state-of-the-art open-source LLMs through automatic and human evaluation. We select Self-Refine (Madaan et al., 2023) as the baseline, which only utilizes self-feedback to refine the explanations. Our findings from automatic evaluation and a user study indicate that Cross-Refine outperforms Self-Refine. Meanwhile, Cross-Refine can perform effectively with less powerful LLMs, whereas Self-Refine only yields strong results with ChatGPT. Additionally, we conduct an ablation study to assess the importance of feedback and suggestions. Both of them play an important role in refining explanations. We further evaluate Cross-Refine on a bilingual dataset in English and German.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n GrEmLIn: A Repository of Green Baseline Embeddings for 87 Low-Resource Languages Injected with Multilingual Graph Knowledge.\n \n \n \n \n\n\n \n Gurgurov, D.; Kumar, R.; and Ostermann, S.\n\n\n \n\n\n\n In Chiruzzo, L.; Ritter, A.; and Wang, L., editor(s), Findings of the Association for Computational Linguistics: NAACL 2025, pages 1204–1221, Albuquerque, New Mexico, April 2025. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"GrEmLIn:Paper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{gurgurov_gremlin_2025,\n\taddress = {Albuquerque, New Mexico},\n\ttitle = {{GrEmLIn}: {A} {Repository} of {Green} {Baseline} {Embeddings} for 87 {Low}-{Resource} {Languages} {Injected} with {Multilingual} {Graph} {Knowledge}},\n\tisbn = {9798891761957},\n\tshorttitle = {{GrEmLIn}},\n\turl = {https://aclanthology.org/2025.findings-naacl.67/},\n\turldate = {2025-05-22},\n\tbooktitle = {Findings of the {Association} for {Computational} {Linguistics}: {NAACL} 2025},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Gurgurov, Daniil and Kumar, Rishu and Ostermann, Simon},\n\teditor = {Chiruzzo, Luis and Ritter, Alan and Wang, Lu},\n\tmonth = apr,\n\tyear = {2025},\n\tpages = {1204--1221},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n FitCF: A Framework for Automatic Feature Importance-guided Counterfactual Example Generation.\n \n \n \n \n\n\n \n Wang, Q.; Feldhus, N.; Ostermann, S.; Villa-Arenas, L. F.; Möller, S.; and Schmitt, V.\n\n\n \n\n\n\n May 2025.\n \n\n\n\n
\n\n\n\n \n \n \"FitCF:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n\n\n\n
\n 
@misc{wang_fitcf_2025,\n\ttitle = {{FitCF}: {A} {Framework} for {Automatic} {Feature} {Importance}-guided {Counterfactual} {Example} {Generation}},\n\tshorttitle = {{FitCF}},\n\turl = {http://arxiv.org/abs/2501.00777},\n\tabstract = {Counterfactual examples are widely used in natural language processing (NLP) as valuable data to improve models, and in explainable artificial intelligence (XAI) to understand model behavior. The automated generation of counterfactual examples remains a challenging task even for large language models (LLMs), despite their impressive performance on many tasks. In this paper, we first introduce ZeroCF, a faithful approach for leveraging important words derived from feature attribution methods to generate counterfactual examples in a zero-shot setting. Second, we present a new framework, FitCF, which further verifies aforementioned counterfactuals by label flip verification and then inserts them as demonstrations for few-shot prompting, outperforming two state-of-the-art baselines. Through ablation studies, we identify the importance of each of FitCF's core components in improving the quality of counterfactuals, as assessed through flip rate, perplexity, and similarity measures. Furthermore, we show the effectiveness of LIME and Integrated Gradients as backbone attribution methods for FitCF and find that the number of demonstrations has the largest effect on performance. Finally, we reveal a strong correlation between the faithfulness of feature attribution scores and the quality of generated counterfactuals.},\n\turldate = {2025-05-22},\n\tpublisher = {arXiv},\n\tauthor = {Wang, Qianli and Feldhus, Nils and Ostermann, Simon and Villa-Arenas, Luis Felipe and Möller, Sebastian and Schmitt, Vera},\n\tmonth = may,\n\tyear = {2025},\n\tdoi = {10.48550/arXiv.2501.00777},\n\tkeywords = {Computer Science - Computation and Language, Computer Science - Machine Learning},\n}\n\n\n\n
\n
\n\n\n
\n Counterfactual examples are widely used in natural language processing (NLP) as valuable data to improve models, and in explainable artificial intelligence (XAI) to understand model behavior. The automated generation of counterfactual examples remains a challenging task even for large language models (LLMs), despite their impressive performance on many tasks. In this paper, we first introduce ZeroCF, a faithful approach for leveraging important words derived from feature attribution methods to generate counterfactual examples in a zero-shot setting. Second, we present a new framework, FitCF, which further verifies aforementioned counterfactuals by label flip verification and then inserts them as demonstrations for few-shot prompting, outperforming two state-of-the-art baselines. Through ablation studies, we identify the importance of each of FitCF's core components in improving the quality of counterfactuals, as assessed through flip rate, perplexity, and similarity measures. Furthermore, we show the effectiveness of LIME and Integrated Gradients as backbone attribution methods for FitCF and find that the number of demonstrations has the largest effect on performance. Finally, we reveal a strong correlation between the faithfulness of feature attribution scores and the quality of generated counterfactuals.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Reverse Probing: Evaluating Knowledge Transfer via Finetuned Task Embeddings for Coreference Resolution.\n \n \n \n \n\n\n \n Anikina, T.; Binder, A.; Harbecke, D.; Varanasi, S.; Hennig, L.; Ostermann, S.; Möller, S.; and Genabith, J. V.\n\n\n \n\n\n\n In Adlakha, V.; Chronopoulou, A.; Li, X. L.; Majumder, B. P.; Shi, F.; and Vernikos, G., editor(s), Proceedings of the 10th Workshop on Representation Learning for NLP (RepL4NLP-2025), pages 108–119, Albuquerque, NM, May 2025. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"ReversePaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n  \n \n 1 download\n \n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{anikina_reverse_2025,\n\taddress = {Albuquerque, NM},\n\ttitle = {Reverse {Probing}: {Evaluating} {Knowledge} {Transfer} via {Finetuned} {Task} {Embeddings} for {Coreference} {Resolution}},\n\tisbn = {9798891762459},\n\tshorttitle = {Reverse {Probing}},\n\turl = {https://aclanthology.org/2025.repl4nlp-1.9/},\n\tabstract = {In this work, we reimagine classical probing to evaluate knowledge transfer from simple source to more complex target tasks. Instead of probing frozen representations from a complex source task on diverse simple target probing tasks (as usually done in probing), we explore the effectiveness of embeddings from multiple simple source tasks on a single target task. We select coreference resolution, a linguistically complex problem requiring contextual understanding, as focus target task, and test the usefulness of embeddings from comparably simpler tasks tasks such as paraphrase detection, named entity recognition, and relation extraction. Through systematic experiments, we evaluate the impact of individual and combined task embeddings. Our findings reveal that task embeddings vary significantly in utility for coreference resolution, with semantic similarity tasks (e.g., paraphrase detection) proving most beneficial. Additionally, representations from intermediate layers of fine-tuned models often outperform those from final layers. Combining embeddings from multiple tasks consistently improves performance, with attention-based aggregation yielding substantial gains. These insights shed light on relationships between task-specific representations and their adaptability to complex downstream tasks, encouraging further exploration of embedding-level task transfer. Our source code is publicly available under https://github.com/Cora4NLP/multi-task-knowledge-transfer.},\n\turldate = {2025-05-22},\n\tbooktitle = {Proceedings of the 10th {Workshop} on {Representation} {Learning} for {NLP} ({RepL4NLP}-2025)},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Anikina, Tatiana and Binder, Arne and Harbecke, David and Varanasi, Stalin and Hennig, Leonhard and Ostermann, Simon and Möller, Sebastian and Genabith, Josef Van},\n\teditor = {Adlakha, Vaibhav and Chronopoulou, Alexandra and Li, Xiang Lorraine and Majumder, Bodhisattwa Prasad and Shi, Freda and Vernikos, Giorgos},\n\tmonth = may,\n\tyear = {2025},\n\tpages = {108--119},\n}\n\n\n\n
\n
\n\n\n
\n In this work, we reimagine classical probing to evaluate knowledge transfer from simple source to more complex target tasks. Instead of probing frozen representations from a complex source task on diverse simple target probing tasks (as usually done in probing), we explore the effectiveness of embeddings from multiple simple source tasks on a single target task. We select coreference resolution, a linguistically complex problem requiring contextual understanding, as focus target task, and test the usefulness of embeddings from comparably simpler tasks tasks such as paraphrase detection, named entity recognition, and relation extraction. Through systematic experiments, we evaluate the impact of individual and combined task embeddings. Our findings reveal that task embeddings vary significantly in utility for coreference resolution, with semantic similarity tasks (e.g., paraphrase detection) proving most beneficial. Additionally, representations from intermediate layers of fine-tuned models often outperform those from final layers. Combining embeddings from multiple tasks consistently improves performance, with attention-based aggregation yielding substantial gains. These insights shed light on relationships between task-specific representations and their adaptability to complex downstream tasks, encouraging further exploration of embedding-level task transfer. Our source code is publicly available under https://github.com/Cora4NLP/multi-task-knowledge-transfer.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Soft Language Prompts for Language Transfer.\n \n \n \n \n\n\n \n Vykopal, I.; Ostermann, S.; and Simko, M.\n\n\n \n\n\n\n In Chiruzzo, L.; Ritter, A.; and Wang, L., editor(s), Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 10294–10313, Albuquerque, New Mexico, April 2025. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"SoftPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{vykopal_soft_2025,\n\taddress = {Albuquerque, New Mexico},\n\ttitle = {Soft {Language} {Prompts} for {Language} {Transfer}},\n\tisbn = {9798891761896},\n\turl = {https://aclanthology.org/2025.naacl-long.517/},\n\tabstract = {Cross-lingual knowledge transfer, especially between high- and low-resource languages, remains challenging in natural language processing (NLP). This study offers insights for improving cross-lingual NLP applications through the combination of parameter-efficient fine-tuning methods. We systematically explore strategies for enhancing cross-lingual transfer through the incorporation of language-specific and task-specific adapters and soft prompts. We present a detailed investigation of various combinations of these methods, exploring their efficiency across 16 languages, focusing on 10 mid- and low-resource languages. We further present to our knowledge the first use of soft prompts for language transfer, a technique we call soft language prompts. Our findings demonstrate that in contrast to claims of previous work, a combination of language and task adapters does not always work best; instead, combining a soft language prompt with a task adapter outperforms most configurations in many cases.},\n\turldate = {2025-05-22},\n\tbooktitle = {Proceedings of the 2025 {Conference} of the {Nations} of the {Americas} {Chapter} of the {Association} for {Computational} {Linguistics}: {Human} {Language} {Technologies} ({Volume} 1: {Long} {Papers})},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Vykopal, Ivan and Ostermann, Simon and Simko, Marian},\n\teditor = {Chiruzzo, Luis and Ritter, Alan and Wang, Lu},\n\tmonth = apr,\n\tyear = {2025},\n\tpages = {10294--10313},\n}\n\n\n\n
\n
\n\n\n
\n Cross-lingual knowledge transfer, especially between high- and low-resource languages, remains challenging in natural language processing (NLP). This study offers insights for improving cross-lingual NLP applications through the combination of parameter-efficient fine-tuning methods. We systematically explore strategies for enhancing cross-lingual transfer through the incorporation of language-specific and task-specific adapters and soft prompts. We present a detailed investigation of various combinations of these methods, exploring their efficiency across 16 languages, focusing on 10 mid- and low-resource languages. We further present to our knowledge the first use of soft prompts for language transfer, a technique we call soft language prompts. Our findings demonstrate that in contrast to claims of previous work, a combination of language and task adapters does not always work best; instead, combining a soft language prompt with a task adapter outperforms most configurations in many cases.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Taming the AI Monster: Monitoring of Individual Fairness for Effective Human Oversight.\n \n \n \n \n\n\n \n Baum, K.; Biewer, S.; Hermanns, H.; Hetmank, S.; Langer, M.; Lauber-Rönsberg, A.; and Sterz, S.\n\n\n \n\n\n\n In Neele, T.; and Wijs, A., editor(s), Model Checking Software, volume 14624, of Lecture Notes in Computer Science (LNCS), pages 3–25, Cham, 2025. Springer Nature Switzerland\n \n\n\n\n
\n\n\n\n \n \n \"TamingPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{baum_taming_2025,\n\taddress = {Cham},\n\tseries = {Lecture {Notes} in {Computer} {Science} ({LNCS})},\n\ttitle = {Taming the {AI} {Monster}: {Monitoring} of {Individual} {Fairness} for {Effective} {Human} {Oversight}},\n\tvolume = {14624},\n\tisbn = {978-3-031-66149-5},\n\turl = {https://spin-web.github.io/SPIN2024/assets/preproceedings/SPIN2024-paper1.pdf},\n\tdoi = {10.1007/978-3-031-66149-5_1},\n\tbooktitle = {Model {Checking} {Software}},\n\tpublisher = {Springer Nature Switzerland},\n\tauthor = {Baum, Kevin and Biewer, Sebastian and Hermanns, Holger and Hetmank, Sven and Langer, Markus and Lauber-Rönsberg, Anne and Sterz, Sarah},\n\teditor = {Neele, Thomas and Wijs, Anton},\n\tyear = {2025},\n\tpages = {3--25},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Transparent Transparency: Developing a Scheme for Understanding Transparency Requirements.\n \n \n \n\n\n \n Baum, D.; Baum, K.; Zamani, S.; Bennoit, C.; and Werth, D.\n\n\n \n\n\n\n In Steffen, B., editor(s), Bridging the Gap Between AI and Reality, volume 15217, of Lecture Notes in Computer Science (LNCS), pages 55–73, Cham, 2025. Springer Nature Switzerland\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{baum_transparent_2025,\n\taddress = {Cham},\n\tseries = {Lecture {Notes} in {Computer} {Science} ({LNCS})},\n\ttitle = {Transparent {Transparency}: {Developing} a {Scheme} for {Understanding} {Transparency} {Requirements}},\n\tvolume = {15217},\n\tisbn = {978-3-031-75434-0},\n\tdoi = {10.1007/978-3-031-75434-0_5},\n\tbooktitle = {Bridging the {Gap} {Between} {AI} and {Reality}},\n\tpublisher = {Springer Nature Switzerland},\n\tauthor = {Baum, Deborah and Baum, Kevin and Zamani, Sasha and Bennoit, Christian and Werth, Dirk},\n\teditor = {Steffen, Bernhard},\n\tyear = {2025},\n\tpages = {55--73},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n LLAVAGUARD: VLM-based Safeguard for Vision Dataset Curation and Safety Assessment.\n \n \n \n\n\n \n Helff, L.; Friedrich, F.; Brack, M.; Schramowski, P.; and Kersting, K.\n\n\n \n\n\n\n In Proceedings of the 41st International Conference on Machine Learning, 2025. \n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{helff_llavaguard_2025,\n\ttitle = {{LLAVAGUARD}: {VLM}-based {Safeguard} for {Vision} {Dataset} {Curation} and {Safety} {Assessment}},\n\tbooktitle = {Proceedings of the 41st {International} {Conference} on {Machine} {Learning}},\n\tauthor = {Helff, Lukas and Friedrich, Felix and Brack, Manuel and Schramowski, Patrick and Kersting, Kristian},\n\tyear = {2025},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Multilingual Text-to-Image Generation Magnifies Gender Stereotypes and Prompt Engineering May Not Help You.\n \n \n \n \n\n\n \n Friedrich, F.; Hämmerl, K.; Schramowski, P.; Brack, M.; Libovicky, J.; Kersting, K.; and Fraser, A.\n\n\n \n\n\n\n In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, 2025. arXiv\n \n\n\n\n
\n\n\n\n \n \n \"MultilingualPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@inproceedings{friedrich_multilingual_2025,\n\ttitle = {Multilingual {Text}-to-{Image} {Generation} {Magnifies} {Gender} {Stereotypes} and {Prompt} {Engineering} {May} {Not} {Help} {You}},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\turl = {https://arxiv.org/abs/2401.16092},\n\tdoi = {10.48550/ARXIV.2401.16092},\n\tabstract = {Text-to-image generation models have recently achieved astonishing results in image quality, flexibility, and text alignment, and are consequently employed in a fast-growing number of applications. Through improvements in multilingual abilities, a larger community now has access to this technology. However, our results show that multilingual models suffer from significant gender biases just as monolingual models do. Furthermore, the natural expectation that multilingual models will provide similar results across languages does not hold up. Instead, there are important differences between languages. We propose a novel benchmark, MAGBIG, intended to foster research on gender bias in multilingual models. We use MAGBIG to investigate the effect of multilingualism on gender bias in T2I models. To this end, we construct multilingual prompts requesting portraits of people with a certain occupation or trait. Our results show that not only do models exhibit strong gender biases but they also behave differently across languages. Furthermore, we investigate prompt engineering strategies, such as indirect, neutral formulations, to mitigate these biases. Unfortunately, these approaches have limited success and result in worse text-to-image alignment. Consequently, we call for more research into diverse representations across languages in image generators, as well as into steerability to address biased model behavior.},\n\turldate = {2025-05-13},\n\tbooktitle = {Proceedings of the 63rd {Annual} {Meeting} of the {Association} for {Computational} {Linguistics}},\n\tpublisher = {arXiv},\n\tauthor = {Friedrich, Felix and Hämmerl, Katharina and Schramowski, Patrick and Brack, Manuel and Libovicky, Jindrich and Kersting, Kristian and Fraser, Alexander},\n\tyear = {2025},\n\tkeywords = {Computation and Language (cs.CL), Computers and Society (cs.CY), FOS: Computer and information sciences, Machine Learning (cs.LG)},\n}\n\n\n\n
\n
\n\n\n
\n Text-to-image generation models have recently achieved astonishing results in image quality, flexibility, and text alignment, and are consequently employed in a fast-growing number of applications. Through improvements in multilingual abilities, a larger community now has access to this technology. However, our results show that multilingual models suffer from significant gender biases just as monolingual models do. Furthermore, the natural expectation that multilingual models will provide similar results across languages does not hold up. Instead, there are important differences between languages. We propose a novel benchmark, MAGBIG, intended to foster research on gender bias in multilingual models. We use MAGBIG to investigate the effect of multilingualism on gender bias in T2I models. To this end, we construct multilingual prompts requesting portraits of people with a certain occupation or trait. Our results show that not only do models exhibit strong gender biases but they also behave differently across languages. Furthermore, we investigate prompt engineering strategies, such as indirect, neutral formulations, to mitigate these biases. Unfortunately, these approaches have limited success and result in worse text-to-image alignment. Consequently, we call for more research into diverse representations across languages in image generators, as well as into steerability to address biased model behavior.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n MSTS: A Multimodal Safety Test Suite for Vision-Language Models.\n \n \n \n \n\n\n \n Röttger, P.; Attanasio, G.; Friedrich, F.; Goldzycher, J.; Parrish, A.; Bhardwaj, R.; Di Bonaventura, C.; Eng, R.; Geagea, G. E. K.; Goswami, S.; Han, J.; Hovy, D.; Jeong, S.; Jeretič, P.; Plaza-del-Arco, F. M.; Rooein, D.; Schramowski, P.; Shaitarova, A.; Shen, X.; Willats, R.; Zugarini, A.; and Vidgen, B.\n\n\n \n\n\n\n 2025.\n \n\n\n\n
\n\n\n\n \n \n \"MSTS:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n\n\n\n
\n 
@misc{rottger_msts_2025,\n\ttitle = {{MSTS}: {A} {Multimodal} {Safety} {Test} {Suite} for {Vision}-{Language} {Models}},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\tshorttitle = {{MSTS}},\n\turl = {https://arxiv.org/abs/2501.10057},\n\tdoi = {10.48550/ARXIV.2501.10057},\n\tabstract = {Vision-language models (VLMs), which process image and text inputs, are increasingly integrated into chat assistants and other consumer AI applications. Without proper safeguards, however, VLMs may give harmful advice (e.g. how to self-harm) or encourage unsafe behaviours (e.g. to consume drugs). Despite these clear hazards, little work so far has evaluated VLM safety and the novel risks created by multimodal inputs. To address this gap, we introduce MSTS, a Multimodal Safety Test Suite for VLMs. MSTS comprises 400 test prompts across 40 fine-grained hazard categories. Each test prompt consists of a text and an image that only in combination reveal their full unsafe meaning. With MSTS, we find clear safety issues in several open VLMs. We also find some VLMs to be safe by accident, meaning that they are safe because they fail to understand even simple test prompts. We translate MSTS into ten languages, showing non-English prompts to increase the rate of unsafe model responses. We also show models to be safer when tested with text only rather than multimodal prompts. Finally, we explore the automation of VLM safety assessments, finding even the best safety classifiers to be lacking.},\n\turldate = {2025-05-13},\n\tpublisher = {arXiv},\n\tauthor = {Röttger, Paul and Attanasio, Giuseppe and Friedrich, Felix and Goldzycher, Janis and Parrish, Alicia and Bhardwaj, Rishabh and Di Bonaventura, Chiara and Eng, Roman and Geagea, Gaia El Khoury and Goswami, Sujata and Han, Jieun and Hovy, Dirk and Jeong, Seogyeong and Jeretič, Paloma and Plaza-del-Arco, Flor Miriam and Rooein, Donya and Schramowski, Patrick and Shaitarova, Anastassia and Shen, Xudong and Willats, Richard and Zugarini, Andrea and Vidgen, Bertie},\n\tyear = {2025},\n\tkeywords = {Computation and Language (cs.CL), FOS: Computer and information sciences},\n}\n\n\n\n
\n
\n\n\n
\n Vision-language models (VLMs), which process image and text inputs, are increasingly integrated into chat assistants and other consumer AI applications. Without proper safeguards, however, VLMs may give harmful advice (e.g. how to self-harm) or encourage unsafe behaviours (e.g. to consume drugs). Despite these clear hazards, little work so far has evaluated VLM safety and the novel risks created by multimodal inputs. To address this gap, we introduce MSTS, a Multimodal Safety Test Suite for VLMs. MSTS comprises 400 test prompts across 40 fine-grained hazard categories. Each test prompt consists of a text and an image that only in combination reveal their full unsafe meaning. With MSTS, we find clear safety issues in several open VLMs. We also find some VLMs to be safe by accident, meaning that they are safe because they fail to understand even simple test prompts. We translate MSTS into ten languages, showing non-English prompts to increase the rate of unsafe model responses. We also show models to be safer when tested with text only rather than multimodal prompts. Finally, we explore the automation of VLM safety assessments, finding even the best safety classifiers to be lacking.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Safe Reinforcement Learning Through Regret and State Restorations in Evaluation Stages.\n \n \n \n \n\n\n \n Gros, T. P.; Müller, N. J.; Höller, D.; and Wolf, V.\n\n\n \n\n\n\n In Jansen, N.; Junges, S.; Kaminski, B. L.; Matheja, C.; Noll, T.; Quatmann, T.; Stoelinga, M.; and Volk, M., editor(s), Principles of Verification: Cycling the Probabilistic Landscape, volume 15262, pages 18–38. Springer Nature Switzerland, Cham, 2025.\n \n\n\n\n
\n\n\n\n \n \n \"SafePaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@incollection{jansen_safe_2025,\n\taddress = {Cham},\n\ttitle = {Safe {Reinforcement} {Learning} {Through} {Regret} and {State} {Restorations} in {Evaluation} {Stages}},\n\tvolume = {15262},\n\tisbn = {9783031757778 9783031757785},\n\turl = {https://link.springer.com/10.1007/978-3-031-75778-5_2},\n\tlanguage = {en},\n\turldate = {2025-05-13},\n\tbooktitle = {Principles of {Verification}: {Cycling} the {Probabilistic} {Landscape}},\n\tpublisher = {Springer Nature Switzerland},\n\tauthor = {Gros, Timo P. and Müller, Nicola J. and Höller, Daniel and Wolf, Verena},\n\teditor = {Jansen, Nils and Junges, Sebastian and Kaminski, Benjamin Lucien and Matheja, Christoph and Noll, Thomas and Quatmann, Tim and Stoelinga, Mariëlle and Volk, Matthias},\n\tyear = {2025},\n\tdoi = {10.1007/978-3-031-75778-5_2},\n\tpages = {18--38},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n How do we assess the trustworthiness of AI? Introducing the trustworthiness assessment model (TrAM).\n \n \n \n \n\n\n \n Schlicker, N.; Baum, K.; Uhde, A.; Sterz, S.; Hirsch, M. C.; and Langer, M.\n\n\n \n\n\n\n Computers in Human Behavior, 170: 108671. September 2025.\n \n\n\n\n
\n\n\n\n \n \n \"HowPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{schlicker_how_2025,\n\ttitle = {How do we assess the trustworthiness of {AI}? {Introducing} the trustworthiness assessment model ({TrAM})},\n\tvolume = {170},\n\tissn = {07475632},\n\tshorttitle = {How do we assess the trustworthiness of {AI}?},\n\turl = {https://linkinghub.elsevier.com/retrieve/pii/S0747563225001189},\n\tdoi = {10.1016/j.chb.2025.108671},\n\tlanguage = {en},\n\turldate = {2025-05-13},\n\tjournal = {Computers in Human Behavior},\n\tauthor = {Schlicker, Nadine and Baum, Kevin and Uhde, Alarith and Sterz, Sarah and Hirsch, Martin C. and Langer, Markus},\n\tmonth = sep,\n\tyear = {2025},\n\tpages = {108671},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Multi-Agent Causal Reinforcement Learning.\n \n \n \n\n\n \n Meyer-Vitali, A.\n\n\n \n\n\n\n In Proceedings of the 13th International Conference on Model-Based Software and Systems Engineering - MBSE-AI Integration, pages 435–442, 2025. SciTePress\n Backup Publisher: INSTICC ISSN: 2184-4348\n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{meyer-vitali_multi-agent_2025,\n\ttitle = {Multi-{Agent} {Causal} {Reinforcement} {Learning}},\n\tisbn = {978-989-758-729-0},\n\tdoi = {10.5220/0013400100003896},\n\tbooktitle = {Proceedings of the 13th {International} {Conference} on {Model}-{Based} {Software} and {Systems} {Engineering} - {MBSE}-{AI} {Integration}},\n\tpublisher = {SciTePress},\n\tauthor = {Meyer-Vitali, André},\n\tyear = {2025},\n\tnote = {Backup Publisher: INSTICC\nISSN: 2184-4348},\n\tpages = {435--442},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n RITSA: Toward a Retrieval-Augmented Generation System for Intelligent Transportation Systems Architecture.\n \n \n \n\n\n \n Awadid, A.; Meyer-Vitali, A.; Vereno, D.; and Gagnant, M.\n\n\n \n\n\n\n In Proceedings of the 13th International Conference on Model-Based Software and Systems Engineering - MBSE-AI Integration, pages 466–473, 2025. SciTePress\n Backup Publisher: INSTICC ISSN: 2184-4348\n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{awadid_ritsa_2025,\n\ttitle = {{RITSA}: {Toward} a {Retrieval}-{Augmented} {Generation} {System} for {Intelligent} {Transportation} {Systems} {Architecture}},\n\tisbn = {978-989-758-729-0},\n\tdoi = {10.5220/0013443300003896},\n\tbooktitle = {Proceedings of the 13th {International} {Conference} on {Model}-{Based} {Software} and {Systems} {Engineering} - {MBSE}-{AI} {Integration}},\n\tpublisher = {SciTePress},\n\tauthor = {Awadid, Afef and Meyer-Vitali, André and Vereno, Dominik and Gagnant, Maxence},\n\tyear = {2025},\n\tnote = {Backup Publisher: INSTICC\nISSN: 2184-4348},\n\tpages = {466--473},\n}\n\n\n\n\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n AI Act / KI-VO und Standardisierung.\n \n \n \n \n\n\n \n Meyer-Vitali, A.\n\n\n \n\n\n\n Datenschutz und Datensicherheit - DuD, 49(4): 241–244. April 2025.\n \n\n\n\n
\n\n\n\n \n \n \"AIPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n  \n \n 1 download\n \n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n\n\n\n
\n 
@article{meyer-vitali_ai_2025,\n\ttitle = {{AI} {Act} / {KI}-{VO} und {Standardisierung}},\n\tvolume = {49},\n\tissn = {1862-2607},\n\turl = {https://doi.org/10.1007/s11623-025-2079-2},\n\tdoi = {10.1007/s11623-025-2079-2},\n\tabstract = {Bei der Einführung der KI-Verordnung müssen noch viele Aspekte definiert werden, damit einesorgfältige Prüfung geschehen kann. Wir betrachten die Landschaft der Standardisierung, definierenRisiken und das damit einhergehende Vertrauen. Einige Methoden und Prüfmetriken werden dargestellt},\n\tlanguage = {de},\n\tnumber = {4},\n\turldate = {2025-04-23},\n\tjournal = {Datenschutz und Datensicherheit - DuD},\n\tauthor = {Meyer-Vitali, André},\n\tmonth = apr,\n\tyear = {2025},\n\tkeywords = {Artificial Intelligence},\n\tpages = {241--244},\n}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n
\n\n\n
\n Bei der Einführung der KI-Verordnung müssen noch viele Aspekte definiert werden, damit einesorgfältige Prüfung geschehen kann. Wir betrachten die Landschaft der Standardisierung, definierenRisiken und das damit einhergehende Vertrauen. Einige Methoden und Prüfmetriken werden dargestellt\n
\n\n\n
\n\n\n\n\n\n
\n
\n\n
\n
\n  \n 2024\n \n \n (35)\n \n \n
\n
\n \n \n
\n \n\n \n \n \n \n \n Invariance & Causal Representation Learning: Prospects and Limitations.\n \n \n \n\n\n \n Bing, S.; Hochsprung, T.; Wahl, J.; Ninad, U.; and Runge, J.\n\n\n \n\n\n\n Transactions of Machine Learning Research (TMLR). ISSN: 2835-8856., (https://openreview.net/forum?id=lpOC6s4B). 2024.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{bing_invariance_2024,\n\ttitle = {Invariance \\& {Causal} {Representation} {Learning}: {Prospects} and {Limitations}},\n\tnumber = {https://openreview.net/forum?id=lpOC6s4B},\n\tjournal = {Transactions of Machine Learning Research (TMLR). ISSN: 2835-8856.},\n\tauthor = {Bing, Simon and Hochsprung, Tom and Wahl, Jonas and Ninad, Urmi and Runge, Jakob},\n\tyear = {2024},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Identifying Linearly-Mixed Causal Representations from Multi-Node Interventions.\n \n \n \n\n\n \n Bing, S.; Ninad, U.; Wahl, J.; and Runge, J.\n\n\n \n\n\n\n In Causal Learning and Reasoning, pages 843–867, 2024. PMLR\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{bing_identifying_2024,\n\ttitle = {Identifying {Linearly}-{Mixed} {Causal} {Representations} from {Multi}-{Node} {Interventions}},\n\tbooktitle = {Causal {Learning} and {Reasoning}},\n\tpublisher = {PMLR},\n\tauthor = {Bing, Simon and Ninad, Urmi and Wahl, Jonas and Runge, Jakob},\n\tyear = {2024},\n\tpages = {843--867},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Foundations of causal discovery on groups of variables.\n \n \n \n\n\n \n Wahl, J.; Ninad, U.; and Runge, J.\n\n\n \n\n\n\n Journal of Causal Inference, 12(1): 20230041. 2024.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{wahl_foundations_2024,\n\ttitle = {Foundations of causal discovery on groups of variables},\n\tvolume = {12},\n\tnumber = {1},\n\tjournal = {Journal of Causal Inference},\n\tauthor = {Wahl, Jonas and Ninad, Urmi and Runge, Jakob},\n\tyear = {2024},\n\tpages = {20230041},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Causal Inference for Spatial Data Analytics (Dagstuhl Seminar 24202).\n \n \n \n \n\n\n \n Tomko, M.; Xin, Y.; and Wahl, J.\n\n\n \n\n\n\n Dagstuhl Reports, 14(5): 25–57. 2024.\n \n\n\n\n
\n\n\n\n \n \n \"CausalPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{tomko_causal_2024,\n\ttitle = {Causal {Inference} for {Spatial} {Data} {Analytics} ({Dagstuhl} {Seminar} 24202)},\n\tvolume = {14},\n\tissn = {2192-5283},\n\turl = {https://drops.dagstuhl.de/entities/document/10.4230/DagRep.14.5.25},\n\tdoi = {10.4230/DagRep.14.5.25},\n\tnumber = {5},\n\tjournal = {Dagstuhl Reports},\n\tauthor = {Tomko, Martin and Xin, Yanan and Wahl, Jonas},\n\teditor = {Tomko, Martin and Xin, Yanan and Wahl, Jonas},\n\tyear = {2024},\n\tpages = {25--57},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n MMAR: Multilingual and Multimodal Anaphora Resolution in Instructional Videos.\n \n \n \n \n\n\n \n Oguz, C.; Denis, P.; Ostermann, S.; Vincent, E.; Skachkova, N.; and Genabith, J. V.\n\n\n \n\n\n\n In Al-Onaizan, Y.; Bansal, M.; and Chen, Y., editor(s), Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1618–1633, Miami, Florida, USA, November 2024. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"MMAR:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{oguz_mmar_2024,\n\taddress = {Miami, Florida, USA},\n\ttitle = {{MMAR}: {Multilingual} and {Multimodal} {Anaphora} {Resolution} in {Instructional} {Videos}},\n\tshorttitle = {{MMAR}},\n\turl = {https://aclanthology.org/2024.findings-emnlp.88/},\n\tdoi = {10.18653/v1/2024.findings-emnlp.88},\n\tabstract = {Multilingual anaphora resolution identifies referring expressions and implicit arguments in texts and links to antecedents that cover several languages. In the most challenging setting, cross-lingual anaphora resolution, training data, and test data are in different languages. As knowledge needs to be transferred across languages, this task is challenging, both in the multilingual and cross-lingual setting. We hypothesize that one way to alleviate some of the difficulty of the task is to include multimodal information in the form of images (i.e. frames extracted from instructional videos). Such visual inputs are by nature language agnostic, therefore cross- and multilingual anaphora resolution should benefit from visual information. In this paper, we provide the first multilingual and multimodal dataset annotated with anaphoric relations and present experimental results for end-to-end multimodal and multilingual anaphora resolution. Given gold mentions, multimodal features improve anaphora resolution results by {\\textbackslash}textbackslashtextasciitilde10 \\% for unseen languages.},\n\turldate = {2025-05-22},\n\tbooktitle = {Findings of the {Association} for {Computational} {Linguistics}: {EMNLP} 2024},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Oguz, Cennet and Denis, Pascal and Ostermann, Simon and Vincent, Emmanuel and Skachkova, Natalia and Genabith, Josef Van},\n\teditor = {Al-Onaizan, Yaser and Bansal, Mohit and Chen, Yun-Nung},\n\tmonth = nov,\n\tyear = {2024},\n\tpages = {1618--1633},\n}\n\n\n\n
\n
\n\n\n
\n Multilingual anaphora resolution identifies referring expressions and implicit arguments in texts and links to antecedents that cover several languages. In the most challenging setting, cross-lingual anaphora resolution, training data, and test data are in different languages. As knowledge needs to be transferred across languages, this task is challenging, both in the multilingual and cross-lingual setting. We hypothesize that one way to alleviate some of the difficulty of the task is to include multimodal information in the form of images (i.e. frames extracted from instructional videos). Such visual inputs are by nature language agnostic, therefore cross- and multilingual anaphora resolution should benefit from visual information. In this paper, we provide the first multilingual and multimodal dataset annotated with anaphoric relations and present experimental results for end-to-end multimodal and multilingual anaphora resolution. Given gold mentions, multimodal features improve anaphora resolution results by \\textbackslashtextasciitilde10 % for unseen languages.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n A Comparison of Different Tokenization Methods for the Georgian Language.\n \n \n \n \n\n\n \n Mikaberidze, B.; Saghinadze, T.; Mikaberidze, G.; Kalandadze, R.; Pkhakadze, K.; van Genabith, J.; Ostermann, S.; van der Plas, L.; and Müller, P.\n\n\n \n\n\n\n In Abbas, M.; and Freihat, A. A., editor(s), Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024), pages 199–208, Trento, October 2024. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"APaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{mikaberidze_comparison_2024,\n\taddress = {Trento},\n\ttitle = {A {Comparison} of {Different} {Tokenization} {Methods} for the {Georgian} {Language}},\n\turl = {https://aclanthology.org/2024.icnlsp-1.22/},\n\turldate = {2025-05-22},\n\tbooktitle = {Proceedings of the 7th {International} {Conference} on {Natural} {Language} and {Speech} {Processing} ({ICNLSP} 2024)},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Mikaberidze, Beso and Saghinadze, Temo and Mikaberidze, Guram and Kalandadze, Raphael and Pkhakadze, Konstantine and van Genabith, Josef and Ostermann, Simon and van der Plas, Lonneke and Müller, Philipp},\n\teditor = {Abbas, Mourad and Freihat, Abed Alhakim},\n\tmonth = oct,\n\tyear = {2024},\n\tpages = {199--208},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Adapting Multilingual LLMs to Low-Resource Languages with Knowledge Graphs via Adapters.\n \n \n \n \n\n\n \n Gurgurov, D.; Hartmann, M.; and Ostermann, S.\n\n\n \n\n\n\n In Biswas, R.; Kaffee, L.; Agarwal, O.; Minervini, P.; Singh, S.; and de Melo, G., editor(s), Proceedings of the 1st Workshop on Knowledge Graphs and Large Language Models (KaLLM 2024), pages 63–74, Bangkok, Thailand, August 2024. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"AdaptingPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{gurgurov_adapting_2024,\n\taddress = {Bangkok, Thailand},\n\ttitle = {Adapting {Multilingual} {LLMs} to {Low}-{Resource} {Languages} with {Knowledge} {Graphs} via {Adapters}},\n\turl = {https://aclanthology.org/2024.kallm-1.7/},\n\tdoi = {10.18653/v1/2024.kallm-1.7},\n\tabstract = {This paper explores the integration of graph knowledge from linguistic ontologies into multilingual Large Language Models (LLMs) using adapters to improve performance for low-resource languages (LRLs) in sentiment analysis (SA) and named entity recognition (NER). Building upon successful parameter-efficient fine-tuning techniques, such as K-ADAPTER and MAD-X, we propose a similar approach for incorporating knowledge from multilingual graphs, connecting concepts in various languages with each other through linguistic relationships, into multilingual LLMs for LRLs. Specifically, we focus on eight LRLs — Maltese, Bulgarian, Indonesian, Nepali, Javanese, Uyghur, Tibetan, and Sinhala — and employ language-specific adapters fine-tuned on data extracted from the language-specific section of ConceptNet, aiming to enable knowledge transfer across the languages covered by the knowledge graph. We compare various fine-tuning objectives, including standard Masked Language Modeling (MLM), MLM with full-word masking, and MLM with targeted masking, to analyze their effectiveness in learning and integrating the extracted graph data. Through empirical evaluation on language-specific tasks, we assess how structured graph knowledge affects the performance of multilingual LLMs for LRLs in SA and NER, providing insights into the potential benefits of adapting language models for low-resource scenarios.},\n\turldate = {2025-05-22},\n\tbooktitle = {Proceedings of the 1st {Workshop} on {Knowledge} {Graphs} and {Large} {Language} {Models} ({KaLLM} 2024)},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Gurgurov, Daniil and Hartmann, Mareike and Ostermann, Simon},\n\teditor = {Biswas, Russa and Kaffee, Lucie-Aimée and Agarwal, Oshin and Minervini, Pasquale and Singh, Sameer and de Melo, Gerard},\n\tmonth = aug,\n\tyear = {2024},\n\tpages = {63--74},\n}\n\n\n\n
\n
\n\n\n
\n This paper explores the integration of graph knowledge from linguistic ontologies into multilingual Large Language Models (LLMs) using adapters to improve performance for low-resource languages (LRLs) in sentiment analysis (SA) and named entity recognition (NER). Building upon successful parameter-efficient fine-tuning techniques, such as K-ADAPTER and MAD-X, we propose a similar approach for incorporating knowledge from multilingual graphs, connecting concepts in various languages with each other through linguistic relationships, into multilingual LLMs for LRLs. Specifically, we focus on eight LRLs — Maltese, Bulgarian, Indonesian, Nepali, Javanese, Uyghur, Tibetan, and Sinhala — and employ language-specific adapters fine-tuned on data extracted from the language-specific section of ConceptNet, aiming to enable knowledge transfer across the languages covered by the knowledge graph. We compare various fine-tuning objectives, including standard Masked Language Modeling (MLM), MLM with full-word masking, and MLM with targeted masking, to analyze their effectiveness in learning and integrating the extracted graph data. Through empirical evaluation on language-specific tasks, we assess how structured graph knowledge affects the performance of multilingual LLMs for LRLs in SA and NER, providing insights into the potential benefits of adapting language models for low-resource scenarios.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n HybridBERT - Making BERT Pretraining More Efficient Through Hybrid Mixture of Attention Mechanisms.\n \n \n \n \n\n\n \n Srinivasagan, G.; and Ostermann, S.\n\n\n \n\n\n\n In Cao, Y. (.; Papadimitriou, I.; Ovalle, A.; Zampieri, M.; Ferraro, F.; and Swayamdipta, S., editor(s), Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 4: Student Research Workshop), pages 285–291, Mexico City, Mexico, June 2024. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"HybridBERTPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{srinivasagan_hybridbert_2024,\n\taddress = {Mexico City, Mexico},\n\ttitle = {{HybridBERT} - {Making} {BERT} {Pretraining} {More} {Efficient} {Through} {Hybrid} {Mixture} of {Attention} {Mechanisms}},\n\turl = {https://aclanthology.org/2024.naacl-srw.30/},\n\tdoi = {10.18653/v1/2024.naacl-srw.30},\n\tabstract = {Pretrained transformer-based language models have produced state-of-the-art performance in most natural language understanding tasks. These models undergo two stages of training: pretraining on a huge corpus of data and fine-tuning on a specific downstream task. The pretraining phase is extremely compute-intensive and requires several high-performance computing devices like GPUs and several days or even months of training, but it is crucial for the model to capture global knowledge and also has a significant impact on the fine-tuning task. This is a major roadblock for researchers without access to sophisticated computing resources. To overcome this challenge, we propose two novel hybrid architectures called HybridBERT (HBERT), which combine self-attention and additive attention mechanisms together with sub-layer normalization. We introduce a computing budget to the pretraining phase, limiting the training time and usage to a single GPU. We show that HBERT attains twice the pretraining accuracy of a vanilla-BERT baseline. We also evaluate our proposed models on two downstream tasks, where we outperform BERT-base while accelerating inference. Moreover, we study the effect of weight initialization with a limited pretraining budget. The code and models are publicly available at: www.github.com/gokulsg/HBERT/.},\n\turldate = {2025-05-22},\n\tbooktitle = {Proceedings of the 2024 {Conference} of the {North} {American} {Chapter} of the {Association} for {Computational} {Linguistics}: {Human} {Language} {Technologies} ({Volume} 4: {Student} {Research} {Workshop})},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Srinivasagan, Gokul and Ostermann, Simon},\n\teditor = {Cao, Yang (Trista) and Papadimitriou, Isabel and Ovalle, Anaelia and Zampieri, Marcos and Ferraro, Francis and Swayamdipta, Swabha},\n\tmonth = jun,\n\tyear = {2024},\n\tpages = {285--291},\n}\n\n\n\n
\n
\n\n\n
\n Pretrained transformer-based language models have produced state-of-the-art performance in most natural language understanding tasks. These models undergo two stages of training: pretraining on a huge corpus of data and fine-tuning on a specific downstream task. The pretraining phase is extremely compute-intensive and requires several high-performance computing devices like GPUs and several days or even months of training, but it is crucial for the model to capture global knowledge and also has a significant impact on the fine-tuning task. This is a major roadblock for researchers without access to sophisticated computing resources. To overcome this challenge, we propose two novel hybrid architectures called HybridBERT (HBERT), which combine self-attention and additive attention mechanisms together with sub-layer normalization. We introduce a computing budget to the pretraining phase, limiting the training time and usage to a single GPU. We show that HBERT attains twice the pretraining accuracy of a vanilla-BERT baseline. We also evaluate our proposed models on two downstream tasks, where we outperform BERT-base while accelerating inference. Moreover, we study the effect of weight initialization with a limited pretraining budget. The code and models are publicly available at: www.github.com/gokulsg/HBERT/.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n CoXQL: A Dataset for Parsing Explanation Requests in Conversational XAI Systems.\n \n \n \n \n\n\n \n Wang, Q.; Anikina, T.; Feldhus, N.; Ostermann, S.; and Möller, S.\n\n\n \n\n\n\n In Al-Onaizan, Y.; Bansal, M.; and Chen, Y., editor(s), Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1410–1422, Miami, Florida, USA, November 2024. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"CoXQL:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{wang_coxql_2024,\n\taddress = {Miami, Florida, USA},\n\ttitle = {{CoXQL}: {A} {Dataset} for {Parsing} {Explanation} {Requests} in {Conversational} {XAI} {Systems}},\n\tshorttitle = {{CoXQL}},\n\turl = {https://aclanthology.org/2024.findings-emnlp.76/},\n\tdoi = {10.18653/v1/2024.findings-emnlp.76},\n\tabstract = {Conversational explainable artificial intelligence (ConvXAI) systems based on large language models (LLMs) have garnered significant interest from the research community in natural language processing (NLP) and human-computer interaction (HCI). Such systems can provide answers to user questions about explanations in dialogues, have the potential to enhance users' comprehension and offer more information about the decision-making and generation processes of LLMs. Currently available ConvXAI systems are based on intent recognition rather than free chat, as this has been found to be more precise and reliable in identifying users' intentions. However, the recognition of intents still presents a challenge in the case of ConvXAI, since little training data exist and the domain is highly specific, as there is a broad range of XAI methods to map requests onto. In order to bridge this gap, we present CoXQL, the first dataset in the NLP domain for user intent recognition in ConvXAI, covering 31 intents, seven of which require filling multiple slots. Subsequently, we enhance an existing parsing approach by incorporating template validations, and conduct an evaluation of several LLMs on CoXQL using different parsing strategies. We conclude that the improved parsing approach (MP+) surpasses the performance of previous approaches. We also discover that intents with multiple slots remain highly challenging for LLMs.},\n\turldate = {2025-05-22},\n\tbooktitle = {Findings of the {Association} for {Computational} {Linguistics}: {EMNLP} 2024},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Wang, Qianli and Anikina, Tatiana and Feldhus, Nils and Ostermann, Simon and Möller, Sebastian},\n\teditor = {Al-Onaizan, Yaser and Bansal, Mohit and Chen, Yun-Nung},\n\tmonth = nov,\n\tyear = {2024},\n\tpages = {1410--1422},\n}\n\n\n\n
\n
\n\n\n
\n Conversational explainable artificial intelligence (ConvXAI) systems based on large language models (LLMs) have garnered significant interest from the research community in natural language processing (NLP) and human-computer interaction (HCI). Such systems can provide answers to user questions about explanations in dialogues, have the potential to enhance users' comprehension and offer more information about the decision-making and generation processes of LLMs. Currently available ConvXAI systems are based on intent recognition rather than free chat, as this has been found to be more precise and reliable in identifying users' intentions. However, the recognition of intents still presents a challenge in the case of ConvXAI, since little training data exist and the domain is highly specific, as there is a broad range of XAI methods to map requests onto. In order to bridge this gap, we present CoXQL, the first dataset in the NLP domain for user intent recognition in ConvXAI, covering 31 intents, seven of which require filling multiple slots. Subsequently, we enhance an existing parsing approach by incorporating template validations, and conduct an evaluation of several LLMs on CoXQL using different parsing strategies. We conclude that the improved parsing approach (MP+) surpasses the performance of previous approaches. We also discover that intents with multiple slots remain highly challenging for LLMs.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Common European Language Data Space.\n \n \n \n \n\n\n \n Rehm, G.; Piperidis, S.; Choukri, K.; Vasiļjevs, A.; Marheinecke, K.; Arranz, V.; Bērziņš, A.; Deligiannis, M.; Galanis, D.; Giagkou, M.; Gkirtzou, K.; Gkoumas, D.; Grützner-Zahn, A.; Kolovou, A.; Labropoulou, P.; Lagzdiņš, A.; Leitner, E.; Mapelli, V.; Mazo, H.; Ostermann, S.; Racioppa, S.; Rigault, M.; and Voukoutis, L.\n\n\n \n\n\n\n In Calzolari, N.; Kan, M.; Hoste, V.; Lenci, A.; Sakti, S.; and Xue, N., editor(s), Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 3579–3586, Torino, Italia, May 2024. ELRA and ICCL\n \n\n\n\n
\n\n\n\n \n \n \"CommonPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{rehm_common_2024,\n\taddress = {Torino, Italia},\n\ttitle = {Common {European} {Language} {Data} {Space}},\n\turl = {https://aclanthology.org/2024.lrec-main.317/},\n\tabstract = {The Common European Language Data Space (LDS) is an integral part of the EU data strategy, which aims at developing a single market for data. Its decentralised technical infrastructure and governance scheme are currently being developed by the LDS project, which also has dedicated tasks for proof-of-concept prototypes, handling legal aspects, raising awareness and promoting the LDS through events and social media channels. The LDS is part of a broader vision for establishing all necessary components to develop European large language models.},\n\turldate = {2025-05-22},\n\tbooktitle = {Proceedings of the 2024 {Joint} {International} {Conference} on {Computational} {Linguistics}, {Language} {Resources} and {Evaluation} ({LREC}-{COLING} 2024)},\n\tpublisher = {ELRA and ICCL},\n\tauthor = {Rehm, Georg and Piperidis, Stelios and Choukri, Khalid and Vasiļjevs, Andrejs and Marheinecke, Katrin and Arranz, Victoria and Bērziņš, Aivars and Deligiannis, Miltos and Galanis, Dimitris and Giagkou, Maria and Gkirtzou, Katerina and Gkoumas, Dimitris and Grützner-Zahn, Annika and Kolovou, Athanasia and Labropoulou, Penny and Lagzdiņš, Andis and Leitner, Elena and Mapelli, Valérie and Mazo, Hélène and Ostermann, Simon and Racioppa, Stefania and Rigault, Mickaël and Voukoutis, Leon},\n\teditor = {Calzolari, Nicoletta and Kan, Min-Yen and Hoste, Veronique and Lenci, Alessandro and Sakti, Sakriani and Xue, Nianwen},\n\tmonth = may,\n\tyear = {2024},\n\tpages = {3579--3586},\n}\n\n\n\n
\n
\n\n\n
\n The Common European Language Data Space (LDS) is an integral part of the EU data strategy, which aims at developing a single market for data. Its decentralised technical infrastructure and governance scheme are currently being developed by the LDS project, which also has dedicated tasks for proof-of-concept prototypes, handling legal aspects, raising awareness and promoting the LDS through events and social media channels. The LDS is part of a broader vision for establishing all necessary components to develop European large language models.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Effective Human Oversight of AI-Based Systems: A Signal Detection Perspective on the Detection of Inaccurate and Unfair Outputs.\n \n \n \n\n\n \n Langer, M.; Baum, K.; and Schlicker, N.\n\n\n \n\n\n\n Minds and Machines, 35(1): 1–30. 2024.\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n  \n \n 1 download\n \n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{langer_effective_2024,\n\ttitle = {Effective {Human} {Oversight} of {AI}-{Based} {Systems}: {A} {Signal} {Detection} {Perspective} on the {Detection} of {Inaccurate} and {Unfair} {Outputs}},\n\tvolume = {35},\n\tdoi = {10.1007/s11023-024-09701-0},\n\tnumber = {1},\n\tjournal = {Minds and Machines},\n\tauthor = {Langer, Markus and Baum, Kevin and Schlicker, Nadine},\n\tyear = {2024},\n\tpages = {1--30},\n}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Acting for the Right Reasons: Creating Reason-Sensitive Artificial Moral Agents.\n \n \n \n \n\n\n \n Baum, K.; Dargasz, L.; Jahn, F.; Gros, T. P.; and Wolf, V.\n\n\n \n\n\n\n In 2024. \n \n\n\n\n
\n\n\n\n \n \n \"ActingPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n  \n \n 1 download\n \n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{baum_acting_2024,\n\ttitle = {Acting for the {Right} {Reasons}: {Creating} {Reason}-{Sensitive} {Artificial} {Moral} {Agents}},\n\turl = {https://openreview.net/forum?id=Mgyr7EY7X9&referrer=%5Bthe%20profile%20of%20Kevin%20Baum%5D(%2Fprofile%3Fid%3D~Kevin_Baum1)},\n\tdoi = {https://doi.org/10.48550/arXiv.2409.15014},\n\tauthor = {Baum, Kevin and Dargasz, Lisa and Jahn, Felix and Gros, Timo P. and Wolf, Verena},\n\tyear = {2024},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Divergent Token Metrics: Measuring degradation to prune away LLM components – and optimize quantization.\n \n \n \n \n\n\n \n Deiseroth, B.; Meuer, M.; Gritsch, N.; Eichenberg, C.; Schramowski, P.; Aßenmacher, M.; and Kersting, K.\n\n\n \n\n\n\n In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics, 2024. arXiv\n \n\n\n\n
\n\n\n\n \n \n \"DivergentPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@inproceedings{deiseroth_divergent_2024,\n\ttitle = {Divergent {Token} {Metrics}: {Measuring} degradation to prune away {LLM} components -- and optimize quantization},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\tshorttitle = {Divergent {Token} {Metrics}},\n\turl = {https://arxiv.org/abs/2311.01544},\n\tdoi = {10.48550/ARXIV.2311.01544},\n\tabstract = {Large Language Models (LLMs) have reshaped natural language processing with their impressive capabilities. However, their ever-increasing size has raised concerns about their effective deployment and the need for LLM compression. This study introduces the Divergent Token Metrics (DTMs), a novel approach to assessing compressed LLMs, addressing the limitations of traditional perplexity or accuracy measures that fail to accurately reflect text generation quality. DTMs measure token divergences that allow deeper insights into the subtleties of model compression, in particular, when evaluating components' impacts individually. Utilizing the First Divergent Token Metric (FDTM) in model sparsification reveals that 25\\% of all attention components can be pruned beyond 90\\% on the Llama-2 model family, still keeping SOTA performance. For quantization, FDTM suggests that more than 80\\% of parameters can be naively transformed to int8 without special outlier management. These evaluations indicate the necessity of choosing appropriate compressions for parameters individually -- and that FDTM can identify those -- while standard metrics result in deteriorated outcomes.},\n\turldate = {2025-05-21},\n\tbooktitle = {Proceedings of the 2024 {Conference} of the {North} {American} {Chapter} of the {Association} for {Computational} {Linguistics}},\n\tpublisher = {arXiv},\n\tauthor = {Deiseroth, Björn and Meuer, Max and Gritsch, Nikolas and Eichenberg, Constantin and Schramowski, Patrick and Aßenmacher, Matthias and Kersting, Kristian},\n\tyear = {2024},\n\tkeywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, Machine Learning (cs.LG)},\n}\n\n\n\n
\n
\n\n\n
\n Large Language Models (LLMs) have reshaped natural language processing with their impressive capabilities. However, their ever-increasing size has raised concerns about their effective deployment and the need for LLM compression. This study introduces the Divergent Token Metrics (DTMs), a novel approach to assessing compressed LLMs, addressing the limitations of traditional perplexity or accuracy measures that fail to accurately reflect text generation quality. DTMs measure token divergences that allow deeper insights into the subtleties of model compression, in particular, when evaluating components' impacts individually. Utilizing the First Divergent Token Metric (FDTM) in model sparsification reveals that 25% of all attention components can be pruned beyond 90% on the Llama-2 model family, still keeping SOTA performance. For quantization, FDTM suggests that more than 80% of parameters can be naively transformed to int8 without special outlier management. These evaluations indicate the necessity of choosing appropriate compressions for parameters individually – and that FDTM can identify those – while standard metrics result in deteriorated outcomes.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n LLMs Lost in Translation: M-ALERT uncovers Cross-Linguistic Safety Gaps.\n \n \n \n \n\n\n \n Friedrich, F.; Tedeschi, S.; Schramowski, P.; Brack, M.; Navigli, R.; Nguyen, H.; Li, B.; and Kersting, K.\n\n\n \n\n\n\n 2024.\n \n\n\n\n
\n\n\n\n \n \n \"LLMsPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n\n\n\n
\n 
@misc{friedrich_llms_2024,\n\ttitle = {{LLMs} {Lost} in {Translation}: {M}-{ALERT} uncovers {Cross}-{Linguistic} {Safety} {Gaps}},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\tshorttitle = {{LLMs} {Lost} in {Translation}},\n\turl = {https://arxiv.org/abs/2412.15035},\n\tdoi = {10.48550/ARXIV.2412.15035},\n\tabstract = {Building safe Large Language Models (LLMs) across multiple languages is essential in ensuring both safe access and linguistic diversity. To this end, we introduce M-ALERT, a multilingual benchmark that evaluates the safety of LLMs in five languages: English, French, German, Italian, and Spanish. M-ALERT includes 15k high-quality prompts per language, totaling 75k, following the detailed ALERT taxonomy. Our extensive experiments on 10 state-of-the-art LLMs highlight the importance of language-specific safety analysis, revealing that models often exhibit significant inconsistencies in safety across languages and categories. For instance, Llama3.2 shows high unsafety in the category crime\\_tax for Italian but remains safe in other languages. Similar differences can be observed across all models. In contrast, certain categories, such as substance\\_cannabis and crime\\_propaganda, consistently trigger unsafe responses across models and languages. These findings underscore the need for robust multilingual safety practices in LLMs to ensure safe and responsible usage across diverse user communities.},\n\turldate = {2025-05-13},\n\tpublisher = {arXiv},\n\tauthor = {Friedrich, Felix and Tedeschi, Simone and Schramowski, Patrick and Brack, Manuel and Navigli, Roberto and Nguyen, Huu and Li, Bo and Kersting, Kristian},\n\tyear = {2024},\n\tkeywords = {Computation and Language (cs.CL), FOS: Computer and information sciences},\n}\n\n\n\n
\n
\n\n\n
\n Building safe Large Language Models (LLMs) across multiple languages is essential in ensuring both safe access and linguistic diversity. To this end, we introduce M-ALERT, a multilingual benchmark that evaluates the safety of LLMs in five languages: English, French, German, Italian, and Spanish. M-ALERT includes 15k high-quality prompts per language, totaling 75k, following the detailed ALERT taxonomy. Our extensive experiments on 10 state-of-the-art LLMs highlight the importance of language-specific safety analysis, revealing that models often exhibit significant inconsistencies in safety across languages and categories. For instance, Llama3.2 shows high unsafety in the category crime_tax for Italian but remains safe in other languages. Similar differences can be observed across all models. In contrast, certain categories, such as substance_cannabis and crime_propaganda, consistently trigger unsafe responses across models and languages. These findings underscore the need for robust multilingual safety practices in LLMs to ensure safe and responsible usage across diverse user communities.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n DeiSAM: Segment Anything with Deictic Prompting.\n \n \n \n\n\n \n Shindo, H.; Brack, M.; Sudhakaran, G.; Dhami, D. S.; Schramowski, P.; and Kersting, K.\n\n\n \n\n\n\n In volume 37, pages 52266–52295, 2024. \n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{shindo_deisam_2024,\n\ttitle = {{DeiSAM}: {Segment} {Anything} with {Deictic} {Prompting}},\n\tvolume = {37},\n\tauthor = {Shindo, Hikaru and Brack, Manuel and Sudhakaran, Gopika and Dhami, Devendra Singh and Schramowski, Patrick and Kersting, Kristian},\n\tyear = {2024},\n\tpages = {52266--52295},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n SCAR: Sparse Conditioned Autoencoders for Concept Detection and Steering in LLMs.\n \n \n \n \n\n\n \n Härle, R.; Friedrich, F.; Brack, M.; Deiseroth, B.; Schramowski, P.; and Kersting, K.\n\n\n \n\n\n\n 2024.\n \n\n\n\n
\n\n\n\n \n \n \"SCAR:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n\n\n\n
\n 
@misc{harle_scar_2024,\n\ttitle = {{SCAR}: {Sparse} {Conditioned} {Autoencoders} for {Concept} {Detection} and {Steering} in {LLMs}},\n\tcopyright = {Creative Commons Attribution Non Commercial Share Alike 4.0 International},\n\tshorttitle = {{SCAR}},\n\turl = {https://arxiv.org/abs/2411.07122},\n\tdoi = {10.48550/ARXIV.2411.07122},\n\tabstract = {Large Language Models (LLMs) have demonstrated remarkable capabilities in generating human-like text, but their output may not be aligned with the user or even produce harmful content. This paper presents a novel approach to detect and steer concepts such as toxicity before generation. We introduce the Sparse Conditioned Autoencoder (SCAR), a single trained module that extends the otherwise untouched LLM. SCAR ensures full steerability, towards and away from concepts (e.g., toxic content), without compromising the quality of the model's text generation on standard evaluation benchmarks. We demonstrate the effective application of our approach through a variety of concepts, including toxicity, safety, and writing style alignment. As such, this work establishes a robust framework for controlling LLM generations, ensuring their ethical and safe deployment in real-world applications.},\n\turldate = {2025-05-13},\n\tpublisher = {arXiv},\n\tauthor = {Härle, Ruben and Friedrich, Felix and Brack, Manuel and Deiseroth, Björn and Schramowski, Patrick and Kersting, Kristian},\n\tyear = {2024},\n\tkeywords = {Computation and Language (cs.CL), FOS: Computer and information sciences},\n}\n\n\n\n
\n
\n\n\n
\n Large Language Models (LLMs) have demonstrated remarkable capabilities in generating human-like text, but their output may not be aligned with the user or even produce harmful content. This paper presents a novel approach to detect and steer concepts such as toxicity before generation. We introduce the Sparse Conditioned Autoencoder (SCAR), a single trained module that extends the otherwise untouched LLM. SCAR ensures full steerability, towards and away from concepts (e.g., toxic content), without compromising the quality of the model's text generation on standard evaluation benchmarks. We demonstrate the effective application of our approach through a variety of concepts, including toxicity, safety, and writing style alignment. As such, this work establishes a robust framework for controlling LLM generations, ensuring their ethical and safe deployment in real-world applications.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Occiglot at WMT24: European Open-source Large Language Models Evaluated on Translation.\n \n \n \n \n\n\n \n Avramidis, E.; Grützner-Zahn, A.; Brack, M.; Schramowski, P.; Ortiz Suarez, P.; Ostendorff, M.; Barth, F.; Manakhimova, S.; Macketanz, V.; Rehm, G.; and Kersting, K.\n\n\n \n\n\n\n In Proceedings of the Ninth Conference on Machine Translation, pages 292–298, Miami, Florida, USA, 2024. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"OcciglotPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{avramidis_occiglot_2024,\n\taddress = {Miami, Florida, USA},\n\ttitle = {Occiglot at {WMT24}: {European} {Open}-source {Large} {Language} {Models} {Evaluated} on {Translation}},\n\tshorttitle = {Occiglot at {WMT24}},\n\turl = {https://aclanthology.org/2024.wmt-1.23},\n\tdoi = {10.18653/v1/2024.wmt-1.23},\n\tlanguage = {en},\n\turldate = {2025-05-13},\n\tbooktitle = {Proceedings of the {Ninth} {Conference} on {Machine} {Translation}},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Avramidis, Eleftherios and Grützner-Zahn, Annika and Brack, Manuel and Schramowski, Patrick and Ortiz Suarez, Pedro and Ostendorff, Malte and Barth, Fabio and Manakhimova, Shushen and Macketanz, Vivien and Rehm, Georg and Kersting, Kristian},\n\tyear = {2024},\n\tpages = {292--298},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n T-FREE: Subword Tokenizer-Free Generative LLMs via Sparse Representations for Memory-Efficient Embeddings.\n \n \n \n \n\n\n \n Deiseroth, B.; Brack, M.; Schramowski, P.; Kersting, K.; and Weinbach, S.\n\n\n \n\n\n\n In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21829–21851, Miami, Florida, USA, 2024. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"T-FREE:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{deiseroth_t-free_2024,\n\taddress = {Miami, Florida, USA},\n\ttitle = {T-{FREE}: {Subword} {Tokenizer}-{Free} {Generative} {LLMs} via {Sparse} {Representations} for {Memory}-{Efficient} {Embeddings}},\n\tshorttitle = {T-{FREE}},\n\turl = {https://aclanthology.org/2024.emnlp-main.1217},\n\tdoi = {10.18653/v1/2024.emnlp-main.1217},\n\tlanguage = {en},\n\turldate = {2025-05-13},\n\tbooktitle = {Proceedings of the 2024 {Conference} on {Empirical} {Methods} in {Natural} {Language} {Processing}},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Deiseroth, Björn and Brack, Manuel and Schramowski, Patrick and Kersting, Kristian and Weinbach, Samuel},\n\tyear = {2024},\n\tpages = {21829--21851},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Community OSCAR: A Community Effort for Multilingual Web Data.\n \n \n \n \n\n\n \n Brack, M.; Ostendorff, M.; Ortiz Suarez, P.; Saiz, J. J.; Castilla, I. L.; Palomar-Giner, J.; Shvets, A.; Schramowski, P.; Rehm, G.; Villegas, M.; and Kersting, K.\n\n\n \n\n\n\n In Proceedings of the Fourth Workshop on Multilingual Representation Learning (MRL 2024), pages 232–235, Miami, Florida, USA, 2024. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"CommunityPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{brack_community_2024,\n\taddress = {Miami, Florida, USA},\n\ttitle = {Community {OSCAR}: {A} {Community} {Effort} for {Multilingual} {Web} {Data}},\n\tshorttitle = {Community {OSCAR}},\n\turl = {https://aclanthology.org/2024.mrl-1.19},\n\tdoi = {10.18653/v1/2024.mrl-1.19},\n\tlanguage = {en},\n\turldate = {2025-05-13},\n\tbooktitle = {Proceedings of the {Fourth} {Workshop} on {Multilingual} {Representation} {Learning} ({MRL} 2024)},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Brack, Manuel and Ostendorff, Malte and Ortiz Suarez, Pedro and Saiz, José Javier and Castilla, Iñaki Lacunza and Palomar-Giner, Jorge and Shvets, Alexander and Schramowski, Patrick and Rehm, Georg and Villegas, Marta and Kersting, Kristian},\n\tyear = {2024},\n\tpages = {232--235},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Core Tokensets for Data-efficient Sequential Training of Transformers.\n \n \n \n \n\n\n \n Paul, S.; Brack, M.; Schramowski, P.; Kersting, K.; and Mundt, M.\n\n\n \n\n\n\n 2024.\n \n\n\n\n
\n\n\n\n \n \n \"CorePaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@misc{paul_core_2024,\n\ttitle = {Core {Tokensets} for {Data}-efficient {Sequential} {Training} of {Transformers}},\n\tcopyright = {Creative Commons Attribution Non Commercial Share Alike 4.0 International},\n\turl = {https://arxiv.org/abs/2410.05800},\n\tdoi = {10.48550/ARXIV.2410.05800},\n\tabstract = {Deep networks are frequently tuned to novel tasks and continue learning from ongoing data streams. Such sequential training requires consolidation of new and past information, a challenge predominantly addressed by retaining the most important data points - formally known as coresets. Traditionally, these coresets consist of entire samples, such as images or sentences. However, recent transformer architectures operate on tokens, leading to the famous assertion that an image is worth 16x16 words. Intuitively, not all of these tokens are equally informative or memorable. Going beyond coresets, we thus propose to construct a deeper-level data summary on the level of tokens. Our respectively named core tokensets both select the most informative data points and leverage feature attribution to store only their most relevant features. We demonstrate that core tokensets yield significant performance retention in incremental image classification, open-ended visual question answering, and continual image captioning with significantly reduced memory. In fact, we empirically find that a core tokenset of 1{\\textbackslash}\\% of the data performs comparably to at least a twice as large and up to 10 times larger coreset.},\n\turldate = {2025-05-13},\n\tpublisher = {arXiv},\n\tauthor = {Paul, Subarnaduti and Brack, Manuel and Schramowski, Patrick and Kersting, Kristian and Mundt, Martin},\n\tyear = {2024},\n\tkeywords = {Artificial Intelligence (cs.AI), Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences},\n}\n\n\n\n
\n
\n\n\n
\n Deep networks are frequently tuned to novel tasks and continue learning from ongoing data streams. Such sequential training requires consolidation of new and past information, a challenge predominantly addressed by retaining the most important data points - formally known as coresets. Traditionally, these coresets consist of entire samples, such as images or sentences. However, recent transformer architectures operate on tokens, leading to the famous assertion that an image is worth 16x16 words. Intuitively, not all of these tokens are equally informative or memorable. Going beyond coresets, we thus propose to construct a deeper-level data summary on the level of tokens. Our respectively named core tokensets both select the most informative data points and leverage feature attribution to store only their most relevant features. We demonstrate that core tokensets yield significant performance retention in incremental image classification, open-ended visual question answering, and continual image captioning with significantly reduced memory. In fact, we empirically find that a core tokenset of 1\\% of the data performs comparably to at least a twice as large and up to 10 times larger coreset.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Auditing and instructing text-to-image generation models on fairness.\n \n \n \n \n\n\n \n Friedrich, F.; Brack, M.; Struppek, L.; Hintersdorf, D.; Schramowski, P.; Luccioni, S.; and Kersting, K.\n\n\n \n\n\n\n AI and Ethics. August 2024.\n \n\n\n\n
\n\n\n\n \n \n \"AuditingPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{friedrich_auditing_2024,\n\ttitle = {Auditing and instructing text-to-image generation models on fairness},\n\tissn = {2730-5953, 2730-5961},\n\turl = {https://link.springer.com/10.1007/s43681-024-00531-5},\n\tdoi = {10.1007/s43681-024-00531-5},\n\tabstract = {Abstract \n             \n              Generative AI models have recently achieved astonishing results in quality and are consequently employed in a fast-growing number of applications. However, since they are highly data-driven, relying on billion-sized datasets randomly scraped from the internet, they also suffer from degenerated and biased human behavior, as we demonstrate. In fact, they may even reinforce such biases. To not only uncover but also combat these undesired effects, we present a novel strategy, called \n              Fair Diffusion \n              , to attenuate biases during the deployment of generative text-to-image models. Specifically, we demonstrate shifting a bias in any direction based on human instructions yielding arbitrary proportions for, e.g., identity groups. As our empirical evaluation demonstrates, this introduced control enables instructing generative image models on fairness, requiring no data filtering nor additional training.},\n\tlanguage = {en},\n\turldate = {2025-05-13},\n\tjournal = {AI and Ethics},\n\tauthor = {Friedrich, Felix and Brack, Manuel and Struppek, Lukas and Hintersdorf, Dominik and Schramowski, Patrick and Luccioni, Sasha and Kersting, Kristian},\n\tmonth = aug,\n\tyear = {2024},\n}\n\n\n\n
\n
\n\n\n
\n Abstract Generative AI models have recently achieved astonishing results in quality and are consequently employed in a fast-growing number of applications. However, since they are highly data-driven, relying on billion-sized datasets randomly scraped from the internet, they also suffer from degenerated and biased human behavior, as we demonstrate. In fact, they may even reinforce such biases. To not only uncover but also combat these undesired effects, we present a novel strategy, called Fair Diffusion , to attenuate biases during the deployment of generative text-to-image models. Specifically, we demonstrate shifting a bias in any direction based on human instructions yielding arbitrary proportions for, e.g., identity groups. As our empirical evaluation demonstrates, this introduced control enables instructing generative image models on fairness, requiring no data filtering nor additional training.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Does CLIP Know My Face?.\n \n \n \n \n\n\n \n Hintersdorf, D.; Struppek, L.; Brack, M.; Friedrich, F.; Schramowski, P.; and Kersting, K.\n\n\n \n\n\n\n Journal of Artificial Intelligence Research, 80: 1033–1062. July 2024.\n \n\n\n\n
\n\n\n\n \n \n \"DoesPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{hintersdorf_does_2024,\n\ttitle = {Does {CLIP} {Know} {My} {Face}?},\n\tvolume = {80},\n\tissn = {1076-9757},\n\turl = {https://www.jair.org/index.php/jair/article/view/15461},\n\tdoi = {10.1613/jair.1.15461},\n\tabstract = {With the rise of deep learning in various applications, privacy concerns around the protection of training data have become a critical area of research. Whereas prior studies have focused on privacy risks in single-modal models, we introduce a novel method to assess privacy for multi-modal models, specifically vision-language models like CLIP. The proposed Identity Inference Attack (IDIA) reveals whether an individual was included in the training data by querying the model with images of the same person. Letting the model choose from a wide variety of possible text labels, the model reveals whether it recognizes the person and, therefore, was used for training. Our large-scale experiments on CLIP demonstrate that individuals used for training can be identified with very high accuracy. We confirm that the model has learned to associate names with depicted individuals, implying the existence of sensitive information that can be extracted by adversaries. Our results highlight the need for stronger privacy protection in large-scale models and suggest that IDIAs can be used to prove the unauthorized use of data for training and to enforce privacy laws. \nThis article appears in the AI \\& Society track.},\n\turldate = {2025-05-13},\n\tjournal = {Journal of Artificial Intelligence Research},\n\tauthor = {Hintersdorf, Dominik and Struppek, Lukas and Brack, Manuel and Friedrich, Felix and Schramowski, Patrick and Kersting, Kristian},\n\tmonth = jul,\n\tyear = {2024},\n\tpages = {1033--1062},\n}\n\n\n\n
\n
\n\n\n
\n With the rise of deep learning in various applications, privacy concerns around the protection of training data have become a critical area of research. Whereas prior studies have focused on privacy risks in single-modal models, we introduce a novel method to assess privacy for multi-modal models, specifically vision-language models like CLIP. The proposed Identity Inference Attack (IDIA) reveals whether an individual was included in the training data by querying the model with images of the same person. Letting the model choose from a wide variety of possible text labels, the model reveals whether it recognizes the person and, therefore, was used for training. Our large-scale experiments on CLIP demonstrate that individuals used for training can be identified with very high accuracy. We confirm that the model has learned to associate names with depicted individuals, implying the existence of sensitive information that can be extracted by adversaries. Our results highlight the need for stronger privacy protection in large-scale models and suggest that IDIAs can be used to prove the unauthorized use of data for training and to enforce privacy laws. This article appears in the AI & Society track.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Introducing v0.5 of the AI Safety Benchmark from MLCommons.\n \n \n \n \n\n\n \n Vidgen, B.; Agrawal, A.; Ahmed, A. M.; Akinwande, V.; Al-Nuaimi, N.; Alfaraj, N.; Alhajjar, E.; Aroyo, L.; Bavalatti, T.; Bartolo, M.; Blili-Hamelin, B.; Bollacker, K.; Bomassani, R.; Boston, M. F.; Campos, S.; Chakra, K.; Chen, C.; Coleman, C.; Coudert, Z. D.; Derczynski, L.; Dutta, D.; Eisenberg, I.; Ezick, J.; Frase, H.; Fuller, B.; Gandikota, R.; Gangavarapu, A.; Gangavarapu, A.; Gealy, J.; Ghosh, R.; Goel, J.; Gohar, U.; Goswami, S.; Hale, S. A.; Hutiri, W.; Imperial, J. M.; Jandial, S.; Judd, N.; Juefei-Xu, F.; Khomh, F.; Kailkhura, B.; Kirk, H. R.; Klyman, K.; Knotz, C.; Kuchnik, M.; Kumar, S. H.; Kumar, S.; Lengerich, C.; Li, B.; Liao, Z.; Long, E. P.; Lu, V.; Luger, S.; Mai, Y.; Mammen, P. M.; Manyeki, K.; McGregor, S.; Mehta, V.; Mohammed, S.; Moss, E.; Nachman, L.; Naganna, D. J.; Nikanjam, A.; Nushi, B.; Oala, L.; Orr, I.; Parrish, A.; Patlak, C.; Pietri, W.; Poursabzi-Sangdeh, F.; Presani, E.; Puletti, F.; Röttger, P.; Sahay, S.; Santos, T.; Scherrer, N.; Sebag, A. S.; Schramowski, P.; Shahbazi, A.; Sharma, V.; Shen, X.; Sistla, V.; Tang, L.; Testuggine, D.; Thangarasa, V.; Watkins, E. A.; Weiss, R.; Welty, C.; Wilbers, T.; Williams, A.; Wu, C.; Yadav, P.; Yang, X.; Zeng, Y.; Zhang, W.; Zhdanov, F.; Zhu, J.; Liang, P.; Mattson, P.; and Vanschoren, J.\n\n\n \n\n\n\n 2024.\n \n\n\n\n
\n\n\n\n \n \n \"IntroducingPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@misc{vidgen_introducing_2024,\n\ttitle = {Introducing v0.5 of the {AI} {Safety} {Benchmark} from {MLCommons}},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\turl = {https://arxiv.org/abs/2404.12241},\n\tdoi = {10.48550/ARXIV.2404.12241},\n\tabstract = {This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.},\n\turldate = {2025-05-13},\n\tpublisher = {arXiv},\n\tauthor = {Vidgen, Bertie and Agrawal, Adarsh and Ahmed, Ahmed M. and Akinwande, Victor and Al-Nuaimi, Namir and Alfaraj, Najla and Alhajjar, Elie and Aroyo, Lora and Bavalatti, Trupti and Bartolo, Max and Blili-Hamelin, Borhane and Bollacker, Kurt and Bomassani, Rishi and Boston, Marisa Ferrara and Campos, Siméon and Chakra, Kal and Chen, Canyu and Coleman, Cody and Coudert, Zacharie Delpierre and Derczynski, Leon and Dutta, Debojyoti and Eisenberg, Ian and Ezick, James and Frase, Heather and Fuller, Brian and Gandikota, Ram and Gangavarapu, Agasthya and Gangavarapu, Ananya and Gealy, James and Ghosh, Rajat and Goel, James and Gohar, Usman and Goswami, Sujata and Hale, Scott A. and Hutiri, Wiebke and Imperial, Joseph Marvin and Jandial, Surgan and Judd, Nick and Juefei-Xu, Felix and Khomh, Foutse and Kailkhura, Bhavya and Kirk, Hannah Rose and Klyman, Kevin and Knotz, Chris and Kuchnik, Michael and Kumar, Shachi H. and Kumar, Srijan and Lengerich, Chris and Li, Bo and Liao, Zeyi and Long, Eileen Peters and Lu, Victor and Luger, Sarah and Mai, Yifan and Mammen, Priyanka Mary and Manyeki, Kelvin and McGregor, Sean and Mehta, Virendra and Mohammed, Shafee and Moss, Emanuel and Nachman, Lama and Naganna, Dinesh Jinenhally and Nikanjam, Amin and Nushi, Besmira and Oala, Luis and Orr, Iftach and Parrish, Alicia and Patlak, Cigdem and Pietri, William and Poursabzi-Sangdeh, Forough and Presani, Eleonora and Puletti, Fabrizio and Röttger, Paul and Sahay, Saurav and Santos, Tim and Scherrer, Nino and Sebag, Alice Schoenauer and Schramowski, Patrick and Shahbazi, Abolfazl and Sharma, Vin and Shen, Xudong and Sistla, Vamsi and Tang, Leonard and Testuggine, Davide and Thangarasa, Vithursan and Watkins, Elizabeth Anne and Weiss, Rebecca and Welty, Chris and Wilbers, Tyler and Williams, Adina and Wu, Carole-Jean and Yadav, Poonam and Yang, Xianjun and Zeng, Yi and Zhang, Wenhui and Zhdanov, Fedor and Zhu, Jiacheng and Liang, Percy and Mattson, Peter and Vanschoren, Joaquin},\n\tyear = {2024},\n\tkeywords = {Artificial Intelligence (cs.AI), Computation and Language (cs.CL), FOS: Computer and information sciences},\n}\n\n\n\n
\n
\n\n\n
\n This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Ledits++: Limitless image editing using text-to-image models.\n \n \n \n\n\n \n Brack, M.; Friedrich, F.; Kornmeier, K.; Tsaban, L.; Schramowski, P.; Kersting, K.; and Passos, A.\n\n\n \n\n\n\n In Proceedings of the IEEE/CVF, pages 8861–8870, 2024. \n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{brack_ledits_2024,\n\ttitle = {Ledits++: {Limitless} image editing using text-to-image models},\n\tbooktitle = {Proceedings of the {IEEE}/{CVF}},\n\tauthor = {Brack, Manuel and Friedrich, Felix and Kornmeier, Katharina and Tsaban, Linoy and Schramowski, Patrick and Kersting, Kristian and Passos, Apolinário},\n\tyear = {2024},\n\tpages = {8861--8870},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Motion Primitives as the Action Space of Deep Q-Learning for Planning in Autonomous Driving.\n \n \n \n \n\n\n \n Schneider, T.; Pedrosa, M. V. A.; Gros, T. P.; Wolf, V.; and Flaßkamp, K.\n\n\n \n\n\n\n IEEE Transactions on Intelligent Transportation Systems, 25(11): 17852–17864. November 2024.\n \n\n\n\n
\n\n\n\n \n \n \"MotionPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{schneider_motion_2024,\n\ttitle = {Motion {Primitives} as the {Action} {Space} of {Deep} {Q}-{Learning} for {Planning} in {Autonomous} {Driving}},\n\tvolume = {25},\n\tcopyright = {https://ieeexplore.ieee.org/Xplorehelp/downloads/license-information/IEEE.html},\n\tissn = {1524-9050, 1558-0016},\n\turl = {https://ieeexplore.ieee.org/document/10693315/},\n\tdoi = {10.1109/TITS.2024.3436530},\n\tnumber = {11},\n\turldate = {2025-05-13},\n\tjournal = {IEEE Transactions on Intelligent Transportation Systems},\n\tauthor = {Schneider, Tristan and Pedrosa, Matheus V. A. and Gros, Timo P. and Wolf, Verena and Flaßkamp, Kathrin},\n\tmonth = nov,\n\tyear = {2024},\n\tpages = {17852--17864},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Comparing State-of-the-art Graph Neural Networks and Transformers for General Policy Learning.\n \n \n \n\n\n \n Müller, N. J.; Sánchez, P.; Hoffmann, J.; Wolf, V.; and Gross, T. P.\n\n\n \n\n\n\n . 2024.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{muller_comparing_2024,\n\ttitle = {Comparing {State}-of-the-art {Graph} {Neural} {Networks} and {Transformers} for {General} {Policy} {Learning}},\n\tauthor = {Müller, Nicola J. and Sánchez, Pablo and Hoffmann, Jörg and Wolf, Verena and Gross, Timo P.},\n\tyear = {2024},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n On the Quest for Effectiveness in Human Oversight: Interdisciplinary Perspectives.\n \n \n \n \n\n\n \n Sterz, S.; Baum, K.; Biewer, S.; Hermanns, H.; Lauber-Rönsberg, A.; Meinel, P.; and Langer, M.\n\n\n \n\n\n\n In The 2024 ACM Conference on Fairness, Accountability, and Transparency, pages 2495–2507, Rio de Janeiro Brazil, June 2024. ACM\n \n\n\n\n
\n\n\n\n \n \n \"OnPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{sterz_quest_2024,\n\taddress = {Rio de Janeiro Brazil},\n\ttitle = {On the {Quest} for {Effectiveness} in {Human} {Oversight}: {Interdisciplinary} {Perspectives}},\n\tisbn = {9798400704505},\n\tshorttitle = {On the {Quest} for {Effectiveness} in {Human} {Oversight}},\n\turl = {https://dl.acm.org/doi/10.1145/3630106.3659051},\n\tdoi = {10.1145/3630106.3659051},\n\tlanguage = {en},\n\turldate = {2025-05-13},\n\tbooktitle = {The 2024 {ACM} {Conference} on {Fairness}, {Accountability}, and {Transparency}},\n\tpublisher = {ACM},\n\tauthor = {Sterz, Sarah and Baum, Kevin and Biewer, Sebastian and Hermanns, Holger and Lauber-Rönsberg, Anne and Meinel, Philip and Langer, Markus},\n\tmonth = jun,\n\tyear = {2024},\n\tpages = {2495--2507},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Software doping analysis for human oversight.\n \n \n \n \n\n\n \n Biewer, S.; Baum, K.; Sterz, S.; Hermanns, H.; Hetmank, S.; Langer, M.; Lauber-Rönsberg, A.; and Lehr, F.\n\n\n \n\n\n\n Formal Methods in System Design. April 2024.\n \n\n\n\n
\n\n\n\n \n \n \"SoftwarePaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n  \n \n 2 downloads\n \n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{biewer_software_2024,\n\ttitle = {Software doping analysis for human oversight},\n\tissn = {0925-9856, 1572-8102},\n\turl = {https://link.springer.com/10.1007/s10703-024-00445-2},\n\tdoi = {10.1007/s10703-024-00445-2},\n\tabstract = {Abstract \n             \n              This article introduces a framework that is meant to assist in mitigating societal risks that software can pose. Concretely, this encompasses facets of software doping as well as unfairness and discrimination in high-risk decision-making systems. The term \n              software doping \n              refers to software that contains surreptitiously added functionality that is against the interest of the user. A prominent example of software doping are the tampered emission cleaning systems that were found in millions of cars around the world when the diesel emissions scandal surfaced. The first part of this article combines the formal foundations of software doping analysis with established probabilistic falsification techniques to arrive at a black-box analysis technique for identifying undesired effects of software. We apply this technique to emission cleaning systems in diesel cars but also to high-risk systems that evaluate humans in a possibly unfair or discriminating way. We demonstrate how our approach can assist humans-in-the-loop to make better informed and more responsible decisions. This is to promote effective human oversight, which will be a central requirement enforced by the European Union’s upcoming AI Act. We complement our technical contribution with a juridically, philosophically, and psychologically informed perspective on the potential problems caused by such systems.},\n\tlanguage = {en},\n\turldate = {2025-05-13},\n\tjournal = {Formal Methods in System Design},\n\tauthor = {Biewer, Sebastian and Baum, Kevin and Sterz, Sarah and Hermanns, Holger and Hetmank, Sven and Langer, Markus and Lauber-Rönsberg, Anne and Lehr, Franz},\n\tmonth = apr,\n\tyear = {2024},\n}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n
\n\n\n
\n Abstract This article introduces a framework that is meant to assist in mitigating societal risks that software can pose. Concretely, this encompasses facets of software doping as well as unfairness and discrimination in high-risk decision-making systems. The term software doping refers to software that contains surreptitiously added functionality that is against the interest of the user. A prominent example of software doping are the tampered emission cleaning systems that were found in millions of cars around the world when the diesel emissions scandal surfaced. The first part of this article combines the formal foundations of software doping analysis with established probabilistic falsification techniques to arrive at a black-box analysis technique for identifying undesired effects of software. We apply this technique to emission cleaning systems in diesel cars but also to high-risk systems that evaluate humans in a possibly unfair or discriminating way. We demonstrate how our approach can assist humans-in-the-loop to make better informed and more responsible decisions. This is to promote effective human oversight, which will be a central requirement enforced by the European Union’s upcoming AI Act. We complement our technical contribution with a juridically, philosophically, and psychologically informed perspective on the potential problems caused by such systems.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n ALERT: A Comprehensive Benchmark for Assessing Large Language Models' Safety through Red Teaming.\n \n \n \n \n\n\n \n Tedeschi, S.; Friedrich, F.; Schramowski, P.; Kersting, K.; Navigli, R.; Nguyen, H.; and Li, B.\n\n\n \n\n\n\n June 2024.\n arXiv:2404.08676 [cs]\n\n\n\n
\n\n\n\n \n \n \"ALERT:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@misc{tedeschi_alert_2024,\n\ttitle = {{ALERT}: {A} {Comprehensive} {Benchmark} for {Assessing} {Large} {Language} {Models}' {Safety} through {Red} {Teaming}},\n\tshorttitle = {{ALERT}},\n\turl = {http://arxiv.org/abs/2404.08676},\n\tdoi = {10.48550/arXiv.2404.08676},\n\tabstract = {When building Large Language Models (LLMs), it is paramount to bear safety in mind and protect them with guardrails. Indeed, LLMs should never generate content promoting or normalizing harmful, illegal, or unethical behavior that may contribute to harm to individuals or society. This principle applies to both normal and adversarial use. In response, we introduce ALERT, a large-scale benchmark to assess safety based on a novel fine-grained risk taxonomy. It is designed to evaluate the safety of LLMs through red teaming methodologies and consists of more than 45k instructions categorized using our novel taxonomy. By subjecting LLMs to adversarial testing scenarios, ALERT aims to identify vulnerabilities, inform improvements, and enhance the overall safety of the language models. Furthermore, the fine-grained taxonomy enables researchers to perform an in-depth evaluation that also helps one to assess the alignment with various policies. In our experiments, we extensively evaluate 10 popular open- and closed-source LLMs and demonstrate that many of them still struggle to attain reasonable levels of safety.},\n\turldate = {2024-12-04},\n\tpublisher = {arXiv},\n\tauthor = {Tedeschi, Simone and Friedrich, Felix and Schramowski, Patrick and Kersting, Kristian and Navigli, Roberto and Nguyen, Huu and Li, Bo},\n\tmonth = jun,\n\tyear = {2024},\n\tnote = {arXiv:2404.08676 [cs]},\n\tkeywords = {Computer Science - Computation and Language, Computer Science - Computers and Society, Computer Science - Machine Learning},\n}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n
\n\n\n
\n When building Large Language Models (LLMs), it is paramount to bear safety in mind and protect them with guardrails. Indeed, LLMs should never generate content promoting or normalizing harmful, illegal, or unethical behavior that may contribute to harm to individuals or society. This principle applies to both normal and adversarial use. In response, we introduce ALERT, a large-scale benchmark to assess safety based on a novel fine-grained risk taxonomy. It is designed to evaluate the safety of LLMs through red teaming methodologies and consists of more than 45k instructions categorized using our novel taxonomy. By subjecting LLMs to adversarial testing scenarios, ALERT aims to identify vulnerabilities, inform improvements, and enhance the overall safety of the language models. Furthermore, the fine-grained taxonomy enables researchers to perform an in-depth evaluation that also helps one to assess the alignment with various policies. In our experiments, we extensively evaluate 10 popular open- and closed-source LLMs and demonstrate that many of them still struggle to attain reasonable levels of safety.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Modular Design Patterns for Generative Neuro-Symbolic Systems.\n \n \n \n \n\n\n \n Boer, M. H. T. d.; Smit, Q. S.; Bekkum, M. v.; Meyer-Vitali, A.; and Schmid, T.\n\n\n \n\n\n\n In Sartini, B.; Raad, J.; Lisena, P.; Peñuela, A. M.; Beetz, M.; Blin, I.; Cimiano, P.; Berardinis, J. d.; Gottschalk, S.; Ilievski, F.; Jain, N.; Kim, J.; Kümpel, M.; Motta, E.; Tiddi, I.; and Töberg, J., editor(s), Joint Proceedings of the ESWC 2024 Workshops and Tutorials, volume 3749, of CEUR Workshop Proceedings, Hersonissos, Greece, May 2024. CEUR\n ISSN: 1613-0073\n\n\n\n
\n\n\n\n \n \n \"ModularPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{boer_modular_2024,\n\taddress = {Hersonissos, Greece},\n\tseries = {{CEUR} {Workshop} {Proceedings}},\n\ttitle = {Modular {Design} {Patterns} for {Generative} {Neuro}-{Symbolic} {Systems}},\n\tvolume = {3749},\n\turl = {https://ceur-ws.org/Vol-3749/#genesy-03},\n\tlanguage = {en},\n\turldate = {2024-09-11},\n\tbooktitle = {Joint {Proceedings} of the {ESWC} 2024 {Workshops} and {Tutorials}},\n\tpublisher = {CEUR},\n\tauthor = {Boer, Maaike H. T. de and Smit, Quirine S. and Bekkum, Michael van and Meyer-Vitali, André and Schmid, Thomas},\n\teditor = {Sartini, Bruno and Raad, Joe and Lisena, Pasquale and Peñuela, Albert Meroño and Beetz, Michael and Blin, Inès and Cimiano, Philipp and Berardinis, Jacopo de and Gottschalk, Simon and Ilievski, Filip and Jain, Nitisha and Kim, Jongmo and Kümpel, Michaela and Motta, Enrico and Tiddi, Ilaria and Töberg, Jan-Philipp},\n\tmonth = may,\n\tyear = {2024},\n\tnote = {ISSN: 1613-0073},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Trusted AI – Vertrauenswürdigkeit und große Sprachmodelle.\n \n \n \n \n\n\n \n Meyer-Vitali, A.; and Ostermann, S.\n\n\n \n\n\n\n dfki ai next, 2024(2): 6–7. September 2024.\n \n\n\n\n
\n\n\n\n \n \n \"TrustedPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{meyer-vitali_trusted_2024,\n\ttitle = {Trusted {AI} – {Vertrauenswürdigkeit} und große  {Sprachmodelle}},\n\tvolume = {2024},\n\turl = {https://www.dfki.de/web/news-media/news/dfki-ai-next},\n\tlanguage = {German},\n\tnumber = {2},\n\tjournal = {dfki ai next},\n\tauthor = {Meyer-Vitali, André and Ostermann, Simon},\n\tmonth = sep,\n\tyear = {2024},\n\tpages = {6--7},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n AI Engineering for Trust by Design.\n \n \n \n \n\n\n \n Meyer-Vitali, A.\n\n\n \n\n\n\n In Proceedings of the 12th International Conference on Model-Based Software and Systems Engineering (MODELSWARD 2024), pages 357–364, Rome, Italy, February 2024. SCITEPRESS – Science and Technology Publications, Lda\n \n\n\n\n
\n\n\n\n \n \n \"AIPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{meyer-vitali_ai_2024,\n\taddress = {Rome, Italy},\n\ttitle = {{AI} {Engineering} for {Trust} by {Design}},\n\tcopyright = {All rights reserved},\n\tisbn = {978-989-758-682-8},\n\turl = {https://www.scitepress.org/PublicationsDetail.aspx?ID=/skO/EwOJr4=&t=1},\n\tdoi = {10.5220/0012622400003645},\n\tabstract = {Digital Library},\n\turldate = {2024-02-29},\n\tbooktitle = {Proceedings of the 12th {International} {Conference} on {Model}-{Based} {Software} and {Systems} {Engineering} ({MODELSWARD} 2024)},\n\tpublisher = {SCITEPRESS – Science and Technology Publications, Lda},\n\tauthor = {Meyer-Vitali, André},\n\tmonth = feb,\n\tyear = {2024},\n\tpages = {357--364},\n}\n\n\n\n
\n
\n\n\n
\n Digital Library\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Human-AI Engineering for Adults.\n \n \n \n \n\n\n \n Meyer-Vitali, A.; and Mulder, W.\n\n\n \n\n\n\n In HHAI 2024: Hybrid Human AI Systems for the Social Good, pages 228–240. IOS Press, 2024.\n \n\n\n\n
\n\n\n\n \n \n \"Human-AIPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@incollection{meyer-vitali_human-ai_2024,\n\ttitle = {Human-{AI} {Engineering} for {Adults}},\n\tcopyright = {All rights reserved},\n\turl = {https://ebooks.iospress.nl/doi/10.3233/FAIA240197},\n\turldate = {2024-06-08},\n\tbooktitle = {{HHAI} 2024: {Hybrid} {Human} {AI} {Systems} for the {Social} {Good}},\n\tpublisher = {IOS Press},\n\tauthor = {Meyer-Vitali, Andr\\&\\#233 and Mulder, Wico},\n\tyear = {2024},\n\tdoi = {10.3233/FAIA240197},\n\tpages = {228--240},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Engineering Principles for Building Trusted Human-AI Systems.\n \n \n \n\n\n \n Meyer-Vitali, A.; and Mulder, W.\n\n\n \n\n\n\n In Arai, K., editor(s), Intelligent Systems and Applications, pages 468–485, Cham, 2024. Springer Nature Switzerland\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{meyer-vitali_engineering_2024,\n\taddress = {Cham},\n\ttitle = {Engineering {Principles} for {Building} {Trusted} {Human}-{AI} {Systems}},\n\tcopyright = {All rights reserved},\n\tisbn = {978-3-031-66428-1},\n\tdoi = {10.1007/978-3-031-66428-1_30},\n\tabstract = {In the process engineering reliable and trustworthy AI systems there is significant wisdom to be gained from traditional engineering domains. Extending on earlier work our attention is on topics that stress the principles of building human-AI systems. We plea for a reinforced attention for engineering methods and processes in order to urge the essence for improved scientific progress and industrial AI applications where one can stand on the shoulders of giants. On the one hand, we see their complexity increase on an individual level, as well as on their connected dependency levels, whilst on the other hand, we see a growing lack of experience on the level of their design and engineering. The complexity of current AI models often limits our understanding. The methods and processes to ensure safety, reliability, and transparency are insufficient. This poses serious risks at the level of trustworthiness, particularly when it comes to critical applications with significant social, economic or even physical impact. Future AI systems must adhere to stringent requirements, as mandated, for instance, by the European AI Act, ensuring meticulous design, validation, and certification based on clearly defined criteria.},\n\tlanguage = {en},\n\tbooktitle = {Intelligent {Systems} and {Applications}},\n\tpublisher = {Springer Nature Switzerland},\n\tauthor = {Meyer-Vitali, André and Mulder, Wico},\n\teditor = {Arai, Kohei},\n\tyear = {2024},\n\tpages = {468--485},\n}\n\n\n\n
\n
\n\n\n
\n In the process engineering reliable and trustworthy AI systems there is significant wisdom to be gained from traditional engineering domains. Extending on earlier work our attention is on topics that stress the principles of building human-AI systems. We plea for a reinforced attention for engineering methods and processes in order to urge the essence for improved scientific progress and industrial AI applications where one can stand on the shoulders of giants. On the one hand, we see their complexity increase on an individual level, as well as on their connected dependency levels, whilst on the other hand, we see a growing lack of experience on the level of their design and engineering. The complexity of current AI models often limits our understanding. The methods and processes to ensure safety, reliability, and transparency are insufficient. This poses serious risks at the level of trustworthiness, particularly when it comes to critical applications with significant social, economic or even physical impact. Future AI systems must adhere to stringent requirements, as mandated, for instance, by the European AI Act, ensuring meticulous design, validation, and certification based on clearly defined criteria.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n AI Engineering for Trust by Design.\n \n \n \n \n\n\n \n Meyer-Vitali, A.\n\n\n \n\n\n\n Survey of Tools for Software Engineering, 24(1): 20–22. 2024.\n \n\n\n\n
\n\n\n\n \n \n \"AIPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{meyer-vitali_ai_2024,\n\ttitle = {{AI} {Engineering} for {Trust} by {Design}},\n\tvolume = {24},\n\turl = {https://www.software-innovations.eu/publikationen/},\n\tnumber = {1},\n\tjournal = {Survey of Tools for Software Engineering},\n\tauthor = {Meyer-Vitali, André},\n\tyear = {2024},\n\tpages = {20--22},\n}\n\n\n\n\n
\n
\n\n\n\n
\n\n\n\n\n\n
\n
\n\n
\n
\n  \n 2023\n \n \n (22)\n \n \n
\n
\n \n \n
\n \n\n \n \n \n \n \n Increasing effect sizes of pairwise conditional independence tests between random vectors.\n \n \n \n\n\n \n Hochsprung, T.; Wahl, J.; Gerhardus, A.; Ninad, U.; and Runge, J.\n\n\n \n\n\n\n In Uncertainty in Artificial Intelligence, pages 879–889, 2023. PMLR\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{hochsprung_increasing_2023,\n\ttitle = {Increasing effect sizes of pairwise conditional independence tests between random vectors},\n\tbooktitle = {Uncertainty in {Artificial} {Intelligence}},\n\tpublisher = {PMLR},\n\tauthor = {Hochsprung, Tom and Wahl, Jonas and Gerhardus, Andreas and Ninad, Urmi and Runge, Jakob},\n\tyear = {2023},\n\tpages = {879--889},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Vector causal inference between two groups of variables.\n \n \n \n\n\n \n Wahl, J.; Ninad, U.; and Runge, J.\n\n\n \n\n\n\n In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 12305–12312, 2023. \n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{wahl_vector_2023,\n\ttitle = {Vector causal inference between two groups of variables},\n\tvolume = {37},\n\tbooktitle = {Proceedings of the {AAAI} {Conference} on {Artificial} {Intelligence}},\n\tauthor = {Wahl, Jonas and Ninad, Urmi and Runge, Jakob},\n\tyear = {2023},\n\tpages = {12305--12312},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Investigating the Encoding of Words in BERT`s Neurons Using Feature Textualization.\n \n \n \n \n\n\n \n Baeumel, T.; Vijayakumar, S.; van Genabith, J.; Neumann, G.; and Ostermann, S.\n\n\n \n\n\n\n In Belinkov, Y.; Hao, S.; Jumelet, J.; Kim, N.; McCarthy, A.; and Mohebbi, H., editor(s), Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 261–270, Singapore, December 2023. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"InvestigatingPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{baeumel_investigating_2023,\n\taddress = {Singapore},\n\ttitle = {Investigating the {Encoding} of {Words} in {BERT}`s {Neurons} {Using} {Feature} {Textualization}},\n\turl = {https://aclanthology.org/2023.blackboxnlp-1.20/},\n\tdoi = {10.18653/v1/2023.blackboxnlp-1.20},\n\tabstract = {Pretrained language models (PLMs) form the basis of most state-of-the-art NLP technologies. Nevertheless, they are essentially black boxes: Humans do not have a clear understanding of what knowledge is encoded in different parts of the models, especially in individual neurons. A contrast is in computer vision, where feature visualization provides a decompositional interpretability technique for neurons of vision models. Activation maximization is used to synthesize inherently interpretable visual representations of the information encoded in individual neurons. Our work is inspired by this but presents a cautionary tale on the interpretability of single neurons, based on the first large-scale attempt to adapt activation maximization to NLP, and, more specifically, large PLMs. We propose feature textualization, a technique to produce dense representations of neurons in the PLM word embedding space. We apply feature textualization to the BERT model to investigate whether the knowledge encoded in individual neurons can be interpreted and symbolized. We find that the produced representations can provide insights about the knowledge encoded in individual neurons, but that individual neurons do not represent clear-cut symbolic units of language such as words. Additionally, we use feature textualization to investigate how many neurons are needed to encode words in BERT.},\n\turldate = {2025-05-22},\n\tbooktitle = {Proceedings of the 6th {BlackboxNLP} {Workshop}: {Analyzing} and {Interpreting} {Neural} {Networks} for {NLP}},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Baeumel, Tanja and Vijayakumar, Soniya and van Genabith, Josef and Neumann, Guenter and Ostermann, Simon},\n\teditor = {Belinkov, Yonatan and Hao, Sophie and Jumelet, Jaap and Kim, Najoung and McCarthy, Arya and Mohebbi, Hosein},\n\tmonth = dec,\n\tyear = {2023},\n\tpages = {261--270},\n}\n\n\n\n
\n
\n\n\n
\n Pretrained language models (PLMs) form the basis of most state-of-the-art NLP technologies. Nevertheless, they are essentially black boxes: Humans do not have a clear understanding of what knowledge is encoded in different parts of the models, especially in individual neurons. A contrast is in computer vision, where feature visualization provides a decompositional interpretability technique for neurons of vision models. Activation maximization is used to synthesize inherently interpretable visual representations of the information encoded in individual neurons. Our work is inspired by this but presents a cautionary tale on the interpretability of single neurons, based on the first large-scale attempt to adapt activation maximization to NLP, and, more specifically, large PLMs. We propose feature textualization, a technique to produce dense representations of neurons in the PLM word embedding space. We apply feature textualization to the BERT model to investigate whether the knowledge encoded in individual neurons can be interpreted and symbolized. We find that the produced representations can provide insights about the knowledge encoded in individual neurons, but that individual neurons do not represent clear-cut symbolic units of language such as words. Additionally, we use feature textualization to investigate how many neurons are needed to encode words in BERT.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Find-2-Find: Multitask Learning for Anaphora Resolution and Object Localization.\n \n \n \n \n\n\n \n Oguz, C.; Denis, P.; Vincent, E.; Ostermann, S.; and van Genabith, J.\n\n\n \n\n\n\n In Bouamor, H.; Pino, J.; and Bali, K., editor(s), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8099–8110, Singapore, December 2023. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"Find-2-Find:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n  \n \n 1 download\n \n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{oguz_find-2-find_2023,\n\taddress = {Singapore},\n\ttitle = {Find-2-{Find}: {Multitask} {Learning} for {Anaphora} {Resolution} and {Object} {Localization}},\n\tshorttitle = {Find-2-{Find}},\n\turl = {https://aclanthology.org/2023.emnlp-main.504/},\n\tdoi = {10.18653/v1/2023.emnlp-main.504},\n\tabstract = {In multimodal understanding tasks, visual and linguistic ambiguities can arise. Visual ambiguity can occur when visual objects require a model to ground a referring expression in a video without strong supervision, while linguistic ambiguity can occur from changes in entities in action flows. As an example from the cooking domain, “oil” mixed with “salt” and “pepper” could later be referred to as a “mixture”. Without a clear visual-linguistic alignment, we cannot know which among several objects shown is referred to by the language expression “mixture”, and without resolved antecedents, we cannot pinpoint what the mixture is. We define this chicken-and-egg problem as Visual-linguistic Ambiguity. In this paper, we present Find2Find, a joint anaphora resolution and object localization dataset targeting the problem of visual-linguistic ambiguity, consisting of 500 anaphora-annotated recipes with corresponding videos. We present experimental results of a novel end-to-end joint multitask learning framework for Find2Find that fuses visual and textual information and shows improvements both for anaphora resolution and object localization with one joint model in multitask learning, as compared to a strong single-task baseline.},\n\turldate = {2025-05-22},\n\tbooktitle = {Proceedings of the 2023 {Conference} on {Empirical} {Methods} in {Natural} {Language} {Processing}},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Oguz, Cennet and Denis, Pascal and Vincent, Emmanuel and Ostermann, Simon and van Genabith, Josef},\n\teditor = {Bouamor, Houda and Pino, Juan and Bali, Kalika},\n\tmonth = dec,\n\tyear = {2023},\n\tpages = {8099--8110},\n}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n
\n\n\n
\n In multimodal understanding tasks, visual and linguistic ambiguities can arise. Visual ambiguity can occur when visual objects require a model to ground a referring expression in a video without strong supervision, while linguistic ambiguity can occur from changes in entities in action flows. As an example from the cooking domain, “oil” mixed with “salt” and “pepper” could later be referred to as a “mixture”. Without a clear visual-linguistic alignment, we cannot know which among several objects shown is referred to by the language expression “mixture”, and without resolved antecedents, we cannot pinpoint what the mixture is. We define this chicken-and-egg problem as Visual-linguistic Ambiguity. In this paper, we present Find2Find, a joint anaphora resolution and object localization dataset targeting the problem of visual-linguistic ambiguity, consisting of 500 anaphora-annotated recipes with corresponding videos. We present experimental results of a novel end-to-end joint multitask learning framework for Find2Find that fuses visual and textual information and shows improvements both for anaphora resolution and object localization with one joint model in multitask learning, as compared to a strong single-task baseline.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n XAI Requirements in Smart Production Processes: A Case Study.\n \n \n \n\n\n \n Baum, D.; Baum, K.; Gros, T. P.; and Wolf, V.\n\n\n \n\n\n\n In Longo, L., editor(s), Explainable Artificial Intelligence. Proceedings of the World Conference on eXplainable Artificial Intelligence (xAI 2023), volume 1901, of Communications in Computer and Information Science (CCIS), pages 3–24, Cham, 2023. Springer Nature Switzerland\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{baum_xai_2023,\n\taddress = {Cham},\n\tseries = {Communications in {Computer} and {Information} {Science} ({CCIS})},\n\ttitle = {{XAI} {Requirements} in {Smart} {Production} {Processes}: {A} {Case} {Study}},\n\tvolume = {1901},\n\tdoi = {10.1007/978-3-031-44064-9_1},\n\tabstract = {The increasing prevalence of artificial intelligence (AI) systems has led to a growing consensus on the importance of the explainability of such systems. This is often emphasized with respect to societal and developmental contexts, but it is also crucial within the context of business processes, including manufacturing and production. While this is widely recognized, there is a notable lack of practical examples that demonstrate how to take explainability into account in the latter contexts. This paper presents a real-world use case in which we employed AI to optimize an Industry 4.0 production process without considering explainable AI (XAI) requirements. Building on previous work on models of the relationship between XAI methods and various associated expectations, as well as non-functional explainability requirements, we show how business-oriented XAI requirements can be formulated and prepared for integration into process design. This case study is a valuable resource for researchers and practitioners seeking better to understand the role of explainable AI in practice.},\n\tbooktitle = {Explainable {Artificial} {Intelligence}. {Proceedings} of the {World} {Conference} on {eXplainable} {Artificial} {Intelligence} ({xAI} 2023)},\n\tpublisher = {Springer Nature Switzerland},\n\tauthor = {Baum, Deborah and Baum, Kevin and Gros, Timo P. and Wolf, Verena},\n\teditor = {Longo, Luca},\n\tyear = {2023},\n\tpages = {3--24},\n}\n\n\n\n
\n
\n\n\n
\n The increasing prevalence of artificial intelligence (AI) systems has led to a growing consensus on the importance of the explainability of such systems. This is often emphasized with respect to societal and developmental contexts, but it is also crucial within the context of business processes, including manufacturing and production. While this is widely recognized, there is a notable lack of practical examples that demonstrate how to take explainability into account in the latter contexts. This paper presents a real-world use case in which we employed AI to optimize an Industry 4.0 production process without considering explainable AI (XAI) requirements. Building on previous work on models of the relationship between XAI methods and various associated expectations, as well as non-functional explainability requirements, we show how business-oriented XAI requirements can be formulated and prepared for integration into process design. This case study is a valuable resource for researchers and practitioners seeking better to understand the role of explainable AI in practice.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n From Fear to Action: AI Governance and Opportunities for All.\n \n \n \n\n\n \n Baum, K.; Bryson, J.; Dignum, F.; Dignum, V.; Grobelnik, M.; Hoos, H.; Irgens, M.; Lukowicz, P.; Muller, C.; Rossi, F.; Shawe-Taylor, J.; Theodorou, A.; and Vinuesa, R.\n\n\n \n\n\n\n Frontiers in Computer Science, 5. 2023.\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{baum_fear_2023,\n\ttitle = {From {Fear} to {Action}: {AI} {Governance} and {Opportunities} for {All}},\n\tvolume = {5},\n\tdoi = {10.3389/fcomp.2023.1210421},\n\tjournal = {Frontiers in Computer Science},\n\tauthor = {Baum, Kevin and Bryson, Joanna and Dignum, Frank and Dignum, Virginia and Grobelnik, Marko and Hoos, Holger and Irgens, Morten and Lukowicz, Paul and Muller, Catelijne and Rossi, Francesca and Shawe-Taylor, John and Theodorou, Andreas and Vinuesa, Ricardo},\n\tyear = {2023},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n A typology for exploring the mitigation of shortcut behaviour.\n \n \n \n \n\n\n \n Friedrich, F.; Stammer, W.; Schramowski, P.; and Kersting, K.\n\n\n \n\n\n\n Nature Machine Intelligence, 5(3): 319–330. March 2023.\n \n\n\n\n
\n\n\n\n \n \n \"APaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{friedrich_typology_2023,\n\ttitle = {A typology for exploring the mitigation of shortcut behaviour},\n\tvolume = {5},\n\tissn = {2522-5839},\n\turl = {https://www.nature.com/articles/s42256-023-00612-w},\n\tdoi = {10.1038/s42256-023-00612-w},\n\tlanguage = {en},\n\tnumber = {3},\n\turldate = {2025-05-21},\n\tjournal = {Nature Machine Intelligence},\n\tauthor = {Friedrich, Felix and Stammer, Wolfgang and Schramowski, Patrick and Kersting, Kristian},\n\tmonth = mar,\n\tyear = {2023},\n\tpages = {319--330},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Hyperspectral Imaging in the UV Range Allows for Differentiation of Sugar Beet Diseases Based on Changes in Secondary Plant Metabolites.\n \n \n \n \n\n\n \n Brugger, A.; Yamati, F. I.; Barreto, A.; Paulus, S.; Schramowsk, P.; Kersting, K.; Steiner, U.; Neugart, S.; and Mahlein, A.\n\n\n \n\n\n\n Phytopathology®, 113(1): 44–54. January 2023.\n \n\n\n\n
\n\n\n\n \n \n \"HyperspectralPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{brugger_hyperspectral_2023,\n\ttitle = {Hyperspectral {Imaging} in the {UV} {Range} {Allows} for {Differentiation} of {Sugar} {Beet} {Diseases} {Based} on {Changes} in {Secondary} {Plant} {Metabolites}},\n\tvolume = {113},\n\tissn = {0031-949X, 1943-7684},\n\turl = {https://apsjournals.apsnet.org/doi/10.1094/PHYTO-03-22-0086-R},\n\tdoi = {10.1094/PHYTO-03-22-0086-R},\n\tabstract = {Fungal infections trigger defense or signaling responses in plants, leading to various changes in plant metabolites. The changes in metabolites, for example chlorophyll or flavonoids, have long been detectable using time-consuming destructive analytical methods including high-performance liquid chromatography or photometric determination. Recent plant phenotyping studies have revealed that hyperspectral imaging (HSI) in the UV range can be used to link spectral changes with changes in plant metabolites. To compare established destructive analytical methods with new nondestructive hyperspectral measurements, the interaction between sugar beet leaves and the pathogens Cercospora beticola, which causes Cercospora leaf spot disease (CLS), and Uromyces betae, which causes sugar beet rust (BR), was investigated. With the help of destructive analyses, we showed that both diseases have different effects on chlorophylls, carotenoids, flavonoids, and several phenols. Nondestructive hyperspectral measurements in the UV range revealed different effects of CLS and BR on plant metabolites resulting in distinct reflectance patterns. Both diseases resulted in specific spectral changes that allowed differentiation between the two diseases. Machine learning algorithms enabled the differentiation between the symptom classes and recognition of the two sugar beet diseases. Feature importance analysis identified specific wavelengths important to the classification, highlighting the utility of the UV range. The study demonstrates that HSI in the UV range is a promising, nondestructive tool to investigate the influence of plant diseases on plant physiology and biochemistry.},\n\tlanguage = {en},\n\tnumber = {1},\n\turldate = {2025-05-21},\n\tjournal = {Phytopathology®},\n\tauthor = {Brugger, Anna and Yamati, Facundo Ispizua and Barreto, Abel and Paulus, Stefan and Schramowsk, Patrick and Kersting, Kristian and Steiner, Ulrike and Neugart, Susanne and Mahlein, Anne-Katrin},\n\tmonth = jan,\n\tyear = {2023},\n\tpages = {44--54},\n}\n\n\n\n
\n
\n\n\n
\n Fungal infections trigger defense or signaling responses in plants, leading to various changes in plant metabolites. The changes in metabolites, for example chlorophyll or flavonoids, have long been detectable using time-consuming destructive analytical methods including high-performance liquid chromatography or photometric determination. Recent plant phenotyping studies have revealed that hyperspectral imaging (HSI) in the UV range can be used to link spectral changes with changes in plant metabolites. To compare established destructive analytical methods with new nondestructive hyperspectral measurements, the interaction between sugar beet leaves and the pathogens Cercospora beticola, which causes Cercospora leaf spot disease (CLS), and Uromyces betae, which causes sugar beet rust (BR), was investigated. With the help of destructive analyses, we showed that both diseases have different effects on chlorophylls, carotenoids, flavonoids, and several phenols. Nondestructive hyperspectral measurements in the UV range revealed different effects of CLS and BR on plant metabolites resulting in distinct reflectance patterns. Both diseases resulted in specific spectral changes that allowed differentiation between the two diseases. Machine learning algorithms enabled the differentiation between the symptom classes and recognition of the two sugar beet diseases. Feature importance analysis identified specific wavelengths important to the classification, highlighting the utility of the UV range. The study demonstrates that HSI in the UV range is a promising, nondestructive tool to investigate the influence of plant diseases on plant physiology and biochemistry.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n MultiFusion: Fusing Pre-Trained Models for Multi-Lingual, Multi-Modal Image Generation.\n \n \n \n \n\n\n \n Bellagente, M.; Brack, M.; Teufel, H.; Friedrich, F.; Deiseroth, B.; Eichenberg, C.; Dai, A.; Baldock, R.; Nanda, S.; Oostermeijer, K.; Cruz-Salinas, A. F.; Schramowski, P.; Kersting, K.; and Weinbach, S.\n\n\n \n\n\n\n In Proceedings of the 37th Conference on Neural Information Processing Systems, 2023. arXiv\n \n\n\n\n
\n\n\n\n \n \n \"MultiFusion:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@inproceedings{bellagente_multifusion_2023,\n\ttitle = {{MultiFusion}: {Fusing} {Pre}-{Trained} {Models} for {Multi}-{Lingual}, {Multi}-{Modal} {Image} {Generation}},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\tshorttitle = {{MultiFusion}},\n\turl = {https://arxiv.org/abs/2305.15296},\n\tdoi = {10.48550/ARXIV.2305.15296},\n\tabstract = {The recent popularity of text-to-image diffusion models (DM) can largely be attributed to the intuitive interface they provide to users. The intended generation can be expressed in natural language, with the model producing faithful interpretations of text prompts. However, expressing complex or nuanced ideas in text alone can be difficult. To ease image generation, we propose MultiFusion that allows one to express complex and nuanced concepts with arbitrarily interleaved inputs of multiple modalities and languages. MutliFusion leverages pre-trained models and aligns them for integration into a cohesive system, thereby avoiding the need for extensive training from scratch. Our experimental results demonstrate the efficient transfer of capabilities from individual modules to the downstream model. Specifically, the fusion of all independent components allows the image generation module to utilize multilingual, interleaved multimodal inputs despite being trained solely on monomodal data in a single language.},\n\turldate = {2025-05-21},\n\tbooktitle = {Proceedings of the 37th {Conference} on {Neural} {Information} {Processing} {Systems}},\n\tpublisher = {arXiv},\n\tauthor = {Bellagente, Marco and Brack, Manuel and Teufel, Hannah and Friedrich, Felix and Deiseroth, Björn and Eichenberg, Constantin and Dai, Andrew and Baldock, Robert and Nanda, Souradeep and Oostermeijer, Koen and Cruz-Salinas, Andres Felipe and Schramowski, Patrick and Kersting, Kristian and Weinbach, Samuel},\n\tyear = {2023},\n\tkeywords = {Artificial Intelligence (cs.AI), Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, Machine Learning (cs.LG)},\n}\n\n\n\n
\n
\n\n\n
\n The recent popularity of text-to-image diffusion models (DM) can largely be attributed to the intuitive interface they provide to users. The intended generation can be expressed in natural language, with the model producing faithful interpretations of text prompts. However, expressing complex or nuanced ideas in text alone can be difficult. To ease image generation, we propose MultiFusion that allows one to express complex and nuanced concepts with arbitrarily interleaved inputs of multiple modalities and languages. MutliFusion leverages pre-trained models and aligns them for integration into a cohesive system, thereby avoiding the need for extensive training from scratch. Our experimental results demonstrate the efficient transfer of capabilities from individual modules to the downstream model. Specifically, the fusion of all independent components allows the image generation module to utilize multilingual, interleaved multimodal inputs despite being trained solely on monomodal data in a single language.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n AtMan: Understanding Transformer Predictions Through Memory Efficient Attention Manipulation.\n \n \n \n \n\n\n \n Deiseroth, B.; Deb, M.; Weinbach, S.; Brack, M.; Schramowski, P.; and Kersting, K.\n\n\n \n\n\n\n In Proceedings of the 37th Conference on Neural Information Processing Systems, 2023. \n \n\n\n\n
\n\n\n\n \n \n \"AtMan:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@inproceedings{deiseroth_atman_2023,\n\ttitle = {{AtMan}: {Understanding} {Transformer} {Predictions} {Through} {Memory} {Efficient} {Attention} {Manipulation}},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\tshorttitle = {{AtMan}},\n\turl = {https://arxiv.org/abs/2301.08110},\n\tdoi = {10.48550/ARXIV.2301.08110},\n\tabstract = {Generative transformer models have become increasingly complex, with large numbers of parameters and the ability to process multiple input modalities. Current methods for explaining their predictions are resource-intensive. Most crucially, they require prohibitively large amounts of extra memory, since they rely on backpropagation which allocates almost twice as much GPU memory as the forward pass. This makes it difficult, if not impossible, to use them in production. We present AtMan that provides explanations of generative transformer models at almost no extra cost. Specifically, AtMan is a modality-agnostic perturbation method that manipulates the attention mechanisms of transformers to produce relevance maps for the input with respect to the output prediction. Instead of using backpropagation, AtMan applies a parallelizable token-based search method based on cosine similarity neighborhood in the embedding space. Our exhaustive experiments on text and image-text benchmarks demonstrate that AtMan outperforms current state-of-the-art gradient-based methods on several metrics while being computationally efficient. As such, AtMan is suitable for use in large model inference deployments.},\n\turldate = {2025-05-21},\n\tbooktitle = {Proceedings of the 37th {Conference} on {Neural} {Information} {Processing} {Systems}},\n\tauthor = {Deiseroth, Björn and Deb, Mayukh and Weinbach, Samuel and Brack, Manuel and Schramowski, Patrick and Kersting, Kristian},\n\tyear = {2023},\n\tkeywords = {Artificial Intelligence (cs.AI), FOS: Computer and information sciences, Machine Learning (cs.LG)},\n}\n\n\n\n
\n
\n\n\n
\n Generative transformer models have become increasingly complex, with large numbers of parameters and the ability to process multiple input modalities. Current methods for explaining their predictions are resource-intensive. Most crucially, they require prohibitively large amounts of extra memory, since they rely on backpropagation which allocates almost twice as much GPU memory as the forward pass. This makes it difficult, if not impossible, to use them in production. We present AtMan that provides explanations of generative transformer models at almost no extra cost. Specifically, AtMan is a modality-agnostic perturbation method that manipulates the attention mechanisms of transformers to produce relevance maps for the input with respect to the output prediction. Instead of using backpropagation, AtMan applies a parallelizable token-based search method based on cosine similarity neighborhood in the embedding space. Our exhaustive experiments on text and image-text benchmarks demonstrate that AtMan outperforms current state-of-the-art gradient-based methods on several metrics while being computationally efficient. As such, AtMan is suitable for use in large model inference deployments.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n SEGA: Instructing Text-to-Image Models using Semantic Guidance.\n \n \n \n \n\n\n \n Brack, M.; Friedrich, F.; Hintersdorf, D.; Struppek, L.; Schramowski, P.; and Kersting, K.\n\n\n \n\n\n\n In In Proceedings of the 37th Conference on Neural Information Processing Systems, 2023. arXiv\n \n\n\n\n
\n\n\n\n \n \n \"SEGA:Paper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@inproceedings{brack_sega_2023,\n\ttitle = {{SEGA}: {Instructing} {Text}-to-{Image} {Models} using {Semantic} {Guidance}},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\tshorttitle = {{SEGA}},\n\turl = {https://arxiv.org/abs/2301.12247},\n\tdoi = {10.48550/ARXIV.2301.12247},\n\tabstract = {Text-to-image diffusion models have recently received a lot of interest for their astonishing ability to produce high-fidelity images from text only. However, achieving one-shot generation that aligns with the user's intent is nearly impossible, yet small changes to the input prompt often result in very different images. This leaves the user with little semantic control. To put the user in control, we show how to interact with the diffusion process to flexibly steer it along semantic directions. This semantic guidance (SEGA) generalizes to any generative architecture using classifier-free guidance. More importantly, it allows for subtle and extensive edits, changes in composition and style, as well as optimizing the overall artistic conception. We demonstrate SEGA's effectiveness on both latent and pixel-based diffusion models such as Stable Diffusion, Paella, and DeepFloyd-IF using a variety of tasks, thus providing strong evidence for its versatility, flexibility, and improvements over existing methods.},\n\turldate = {2025-05-21},\n\tbooktitle = {In {Proceedings} of the 37th {Conference} on {Neural} {Information} {Processing} {Systems}},\n\tpublisher = {arXiv},\n\tauthor = {Brack, Manuel and Friedrich, Felix and Hintersdorf, Dominik and Struppek, Lukas and Schramowski, Patrick and Kersting, Kristian},\n\tyear = {2023},\n\tkeywords = {Artificial Intelligence (cs.AI), Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, Machine Learning (cs.LG)},\n}\n\n\n\n
\n
\n\n\n
\n Text-to-image diffusion models have recently received a lot of interest for their astonishing ability to produce high-fidelity images from text only. However, achieving one-shot generation that aligns with the user's intent is nearly impossible, yet small changes to the input prompt often result in very different images. This leaves the user with little semantic control. To put the user in control, we show how to interact with the diffusion process to flexibly steer it along semantic directions. This semantic guidance (SEGA) generalizes to any generative architecture using classifier-free guidance. More importantly, it allows for subtle and extensive edits, changes in composition and style, as well as optimizing the overall artistic conception. We demonstrate SEGA's effectiveness on both latent and pixel-based diffusion models such as Stable Diffusion, Paella, and DeepFloyd-IF using a variety of tasks, thus providing strong evidence for its versatility, flexibility, and improvements over existing methods.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Revision Transformers: Instructing Language Models to Change Their Values.\n \n \n \n \n\n\n \n Friedrich, F.; Stammer, W.; Schramowski, P.; and Kersting, K.\n\n\n \n\n\n\n In ECAI 2023, pages 756–763. IOS Press, 2023.\n \n\n\n\n
\n\n\n\n \n \n \"RevisionPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@incollection{friedrich_revision_2023,\n\ttitle = {Revision {Transformers}: {Instructing} {Language} {Models} to {Change} {Their} {Values}},\n\tshorttitle = {Revision {Transformers}},\n\turl = {https://ebooks.iospress.nl/doi/10.3233/FAIA230341},\n\tlanguage = {en},\n\turldate = {2025-05-20},\n\tbooktitle = {{ECAI} 2023},\n\tpublisher = {IOS Press},\n\tauthor = {Friedrich, Felix and Stammer, Wolfgang and Schramowski, Patrick and Kersting, Kristian},\n\tyear = {2023},\n\tdoi = {10.3233/FAIA230341},\n\tpages = {756--763},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Speaking Multiple Languages Affects the Moral Bias of Language Models.\n \n \n \n \n\n\n \n Haemmerl, K.; Deiseroth, B.; Schramowski, P.; Libovický, J.; Rothkopf, C.; Fraser, A.; and Kersting, K.\n\n\n \n\n\n\n In Findings of the Association for Computational Linguistics: ACL 2023, pages 2137–2156, Toronto, Canada, 2023. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n \n \"SpeakingPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{haemmerl_speaking_2023,\n\taddress = {Toronto, Canada},\n\ttitle = {Speaking {Multiple} {Languages} {Affects} the {Moral} {Bias} of {Language} {Models}},\n\turl = {https://aclanthology.org/2023.findings-acl.134},\n\tdoi = {10.18653/v1/2023.findings-acl.134},\n\tlanguage = {en},\n\turldate = {2025-05-20},\n\tbooktitle = {Findings of the {Association} for {Computational} {Linguistics}: {ACL} 2023},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Haemmerl, Katharina and Deiseroth, Bjoern and Schramowski, Patrick and Libovický, Jindřich and Rothkopf, Constantin and Fraser, Alexander and Kersting, Kristian},\n\tyear = {2023},\n\tpages = {2137--2156},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Exploiting Cultural Biases via Homoglyphs in Text-to-Image Synthesis.\n \n \n \n \n\n\n \n Struppek, L.; Hintersdorf, D.; Friedrich, F.; Br, M.; Schramowski, P.; and Kersting, K.\n\n\n \n\n\n\n Journal of Artificial Intelligence Research, 78: 1017–1068. December 2023.\n \n\n\n\n
\n\n\n\n \n \n \"ExploitingPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{struppek_exploiting_2023,\n\ttitle = {Exploiting {Cultural} {Biases} via {Homoglyphs} in {Text}-to-{Image} {Synthesis}},\n\tvolume = {78},\n\tissn = {1076-9757},\n\turl = {http://www.jair.org/index.php/jair/article/view/15388},\n\tdoi = {10.1613/jair.1.15388},\n\tabstract = {Models for text-to-image synthesis, such as DALL-E 2 and Stable Diffusion, have recently drawn a lot of interest from academia and the general public. These models are capable of producing high-quality images that depict a variety of concepts and styles when conditioned on textual descriptions. However, these models adopt cultural characteristics associated with specific Unicode scripts from their vast amount of training data, which may not be immediately apparent. We show that by simply inserting single non-Latin characters in the textual description, common models reflect cultural biases in their generated images. We analyze this behavior both qualitatively and quantitatively and identify a model’s text encoder as the root cause of the phenomenon. Such behavior can be interpreted as a model feature, offering users a simple way to customize the image generation and reflect their own cultural background. Yet, malicious users or service providers may also try to intentionally bias the image generation. One goal might be to create racist stereotypes by replacing Latin characters with similarly-looking characters from non-Latin scripts, so-called homoglyphs. To mitigate such unnoticed script attacks, we propose a novel homoglyph unlearning method to fine-tune a text encoder, making it robust against homoglyph manipulations.},\n\turldate = {2025-05-13},\n\tjournal = {Journal of Artificial Intelligence Research},\n\tauthor = {Struppek, Lukas and Hintersdorf, Dom and Friedrich, Felix and Br, Manuel and Schramowski, Patrick and Kersting, Kristian},\n\tmonth = dec,\n\tyear = {2023},\n\tpages = {1017--1068},\n}\n\n\n\n
\n
\n\n\n
\n Models for text-to-image synthesis, such as DALL-E 2 and Stable Diffusion, have recently drawn a lot of interest from academia and the general public. These models are capable of producing high-quality images that depict a variety of concepts and styles when conditioned on textual descriptions. However, these models adopt cultural characteristics associated with specific Unicode scripts from their vast amount of training data, which may not be immediately apparent. We show that by simply inserting single non-Latin characters in the textual description, common models reflect cultural biases in their generated images. We analyze this behavior both qualitatively and quantitatively and identify a model’s text encoder as the root cause of the phenomenon. Such behavior can be interpreted as a model feature, offering users a simple way to customize the image generation and reflect their own cultural background. Yet, malicious users or service providers may also try to intentionally bias the image generation. One goal might be to create racist stereotypes by replacing Latin characters with similarly-looking characters from non-Latin scripts, so-called homoglyphs. To mitigate such unnoticed script attacks, we propose a novel homoglyph unlearning method to fine-tune a text encoder, making it robust against homoglyph manipulations.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n ILLUME: Rationalizing Vision-Language Models through Human Interactions.\n \n \n \n\n\n \n Brack, M.; Schramowski, P.; Deiseroth, B.; and Kersting, K.\n\n\n \n\n\n\n In Proceedings of the 40th International Conference on Machine Learning, volume 202, pages 3021–3037, 2023. \n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{brack_illume_2023,\n\ttitle = {{ILLUME}: {Rationalizing} {Vision}-{Language} {Models} through {Human} {Interactions}},\n\tvolume = {202},\n\tbooktitle = {Proceedings of the 40th {International} {Conference} on {Machine} {Learning}},\n\tauthor = {Brack, Manuel and Schramowski, Patrick and Deiseroth, Björn and Kersting, Kristian},\n\tyear = {2023},\n\tpages = {3021--3037},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Self-Supervised Learning of Machine Ethics.\n \n \n \n \n\n\n \n Schramowski, P.\n\n\n \n\n\n\n . 2023.\n \n\n\n\n
\n\n\n\n \n \n \"Self-SupervisedPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{schramowski_self-supervised_2023,\n\ttitle = {Self-{Supervised} {Learning} of {Machine} {Ethics}},\n\tcopyright = {Creative Commons Attribution Share Alike 4.0 International},\n\turl = {https://tuprints.ulb.tu-darmstadt.de/id/eprint/23090},\n\tdoi = {10.26083/TUPRINTS-00023090},\n\tabstract = {In recent years Artificial Intelligence (AI), especially deep learning, has proven to be a technology driver in industry. However, while advancing existing and creating novel technologies, automatizing processes, and assisting humans in essential areas such as drug discovery, they raise many concerns, like other groundbreaking novel technologies before. In this case, these concerns include, for instance, models producing stereotypical and derogatory content as well as gender and racial biases. Since AI technologies will permeate more of our lives in the coming years, these concerns need to be addressed. This thesis examines recent data-driven approaches, which often suffer from degenerated and biased behavior through their self-supervised training on large-scale noisy web data, containing potential inappropriate data. While this is well-established, we will investigate and demonstrate the promises of deep models’ acquired knowledge and capabilities through the provision of this very particular potentially inappropriate data. Importantly, we present the first approaches for learning ethics from data. Our findings suggest that if we build an AI system that learns an improved representation of data and that is able to better understand and produce it, in the process, it will also acquire more accurate societal knowledge, in this case, historical cultural associations to make human-like "right" and "wrong" choices. Furthermore, based on these findings, we consequently ask the arguably "circular" question of whether a machine can help us mitigate their associated concerns. Importantly, we demonstrate the importance of their ability to distinguish between "right" and "wrong" and how utilizing them can mitigate associated risks surrounding large-scale models themselves. However, we also highlight the role of human-machine interaction to explore and reinforce AI systems’ properties, including their flaws and merits, and present how human feedback on explanations can align deep learning based models with our precepts. We present these algorithms and corresponding findings, providing important insights for the goal of putting human values into AI systems, which, summarized, may not be insurmountable in the long run.},\n\turldate = {2025-05-13},\n\tauthor = {Schramowski, Patrick},\n\tyear = {2023},\n}\n\n\n\n
\n
\n\n\n
\n In recent years Artificial Intelligence (AI), especially deep learning, has proven to be a technology driver in industry. However, while advancing existing and creating novel technologies, automatizing processes, and assisting humans in essential areas such as drug discovery, they raise many concerns, like other groundbreaking novel technologies before. In this case, these concerns include, for instance, models producing stereotypical and derogatory content as well as gender and racial biases. Since AI technologies will permeate more of our lives in the coming years, these concerns need to be addressed. This thesis examines recent data-driven approaches, which often suffer from degenerated and biased behavior through their self-supervised training on large-scale noisy web data, containing potential inappropriate data. While this is well-established, we will investigate and demonstrate the promises of deep models’ acquired knowledge and capabilities through the provision of this very particular potentially inappropriate data. Importantly, we present the first approaches for learning ethics from data. Our findings suggest that if we build an AI system that learns an improved representation of data and that is able to better understand and produce it, in the process, it will also acquire more accurate societal knowledge, in this case, historical cultural associations to make human-like \"right\" and \"wrong\" choices. Furthermore, based on these findings, we consequently ask the arguably \"circular\" question of whether a machine can help us mitigate their associated concerns. Importantly, we demonstrate the importance of their ability to distinguish between \"right\" and \"wrong\" and how utilizing them can mitigate associated risks surrounding large-scale models themselves. However, we also highlight the role of human-machine interaction to explore and reinforce AI systems’ properties, including their flaws and merits, and present how human feedback on explanations can align deep learning based models with our precepts. We present these algorithms and corresponding findings, providing important insights for the goal of putting human values into AI systems, which, summarized, may not be insurmountable in the long run.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n DSMC Evaluation Stages: Fostering Robust and Safe Behavior in Deep Reinforcement Learning – Extended Version.\n \n \n \n \n\n\n \n Gros, T. P.; Groß, J.; Höller, D.; Hoffmann, J.; Klauck, M.; Meerkamp, H.; Müller, N. J.; Schaller, L.; and Wolf, V.\n\n\n \n\n\n\n ACM Transactions on Modeling and Computer Simulation, 33(4): 1–28. October 2023.\n \n\n\n\n
\n\n\n\n \n \n \"DSMCPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n  \n \n 1 download\n \n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{gros_dsmc_2023,\n\ttitle = {{DSMC} {Evaluation} {Stages}: {Fostering} {Robust} and {Safe} {Behavior} in {Deep} {Reinforcement} {Learning} – {Extended} {Version}},\n\tvolume = {33},\n\tissn = {1049-3301, 1558-1195},\n\tshorttitle = {{DSMC} {Evaluation} {Stages}},\n\turl = {https://dl.acm.org/doi/10.1145/3607198},\n\tdoi = {10.1145/3607198},\n\tabstract = {Neural networks (NN) are gaining importance in sequential decision-making. Deep reinforcement learning (DRL), in particular, is extremely successful in learning action policies in complex and dynamic environments. Despite this success, however, DRL technology is not without its failures, especially in safety-critical applications: (i) the training objective maximizes \n              average \n              rewards, which may disregard rare but critical situations and hence lack local robustness; (ii) optimization objectives targeting safety typically yield degenerated reward structures, which, for DRL to work, must be replaced with proxy objectives. Here, we introduce a methodology that can help to address both deficiencies. We incorporate \n              evaluation stages \n              (ES) into DRL, leveraging recent work on deep statistical model checking (DSMC), which verifies NN policies in Markov decision processes. Our ES apply DSMC at regular intervals to determine state space regions with weak performance. We adapt the subsequent DRL training priorities based on the outcome, (i) focusing DRL on critical situations and (ii) allowing to foster arbitrary objectives. \n             \n            We run case studies on two benchmarks. One of them is the Racetrack, an abstraction of autonomous driving that requires navigating a map without crashing into a wall. The other is MiniGrid, a widely used benchmark in the AI community. Our results show that DSMC-based ES can significantly improve both (i) and (ii).},\n\tlanguage = {en},\n\tnumber = {4},\n\turldate = {2025-05-13},\n\tjournal = {ACM Transactions on Modeling and Computer Simulation},\n\tauthor = {Gros, Timo P. and Groß, Joschka and Höller, Daniel and Hoffmann, Jörg and Klauck, Michaela and Meerkamp, Hendrik and Müller, Nicola J. and Schaller, Lukas and Wolf, Verena},\n\tmonth = oct,\n\tyear = {2023},\n\tpages = {1--28},\n}\n\n\n\n
\n
\n\n\n
\n Neural networks (NN) are gaining importance in sequential decision-making. Deep reinforcement learning (DRL), in particular, is extremely successful in learning action policies in complex and dynamic environments. Despite this success, however, DRL technology is not without its failures, especially in safety-critical applications: (i) the training objective maximizes average rewards, which may disregard rare but critical situations and hence lack local robustness; (ii) optimization objectives targeting safety typically yield degenerated reward structures, which, for DRL to work, must be replaced with proxy objectives. Here, we introduce a methodology that can help to address both deficiencies. We incorporate evaluation stages (ES) into DRL, leveraging recent work on deep statistical model checking (DSMC), which verifies NN policies in Markov decision processes. Our ES apply DSMC at regular intervals to determine state space regions with weak performance. We adapt the subsequent DRL training priorities based on the outcome, (i) focusing DRL on critical situations and (ii) allowing to foster arbitrary objectives. We run case studies on two benchmarks. One of them is the Racetrack, an abstraction of autonomous driving that requires navigating a map without crashing into a wall. The other is MiniGrid, a widely used benchmark in the AI community. Our results show that DSMC-based ES can significantly improve both (i) and (ii).\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Analyzing neural network behavior through deep statistical model checking.\n \n \n \n \n\n\n \n Gros, T. P.; Hermanns, H.; Hoffmann, J.; Klauck, M.; and Steinmetz, M.\n\n\n \n\n\n\n International Journal on Software Tools for Technology Transfer, 25(3): 407–426. June 2023.\n \n\n\n\n
\n\n\n\n \n \n \"AnalyzingPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{gros_analyzing_2023,\n\ttitle = {Analyzing neural network behavior through deep statistical model checking},\n\tvolume = {25},\n\tissn = {1433-2779, 1433-2787},\n\turl = {https://link.springer.com/10.1007/s10009-022-00685-9},\n\tdoi = {10.1007/s10009-022-00685-9},\n\tabstract = {Abstract \n             \n              Neural networks (NN) are taking over ever more decisions thus far taken by humans, even though verifiable system-level guarantees are far out of reach. Neither is the verification technology available, nor is it even understood what a formal, meaningful, extensible, and scalable testbed might look like for such a technology. The present paper is an attempt to improve on both the above aspects. We present a family of formal models that contain basic features of automated decision-making contexts and which can be extended with further orthogonal features, ultimately encompassing the scope of autonomous driving. Due to the possibility to model random noise in the decision actuation, each model instance induces a Markov decision process (MDP) as verification object. The NN in this context has the duty to actuate (near-optimal) decisions. From the verification perspective, the externally learnt NN serves as a determinizer of the MDP, the result being a Markov chain which as such is amenable to statistical model checking. The combination of an MDP and an NN encoding the action policy is central to what we call “deep statistical model checking” (DSMC). While being a straightforward extension of statistical model checking, it enables to gain deep insight into questions like “how high is the NN-induced safety risk?”, “how good is the NN compared to the optimal policy?” (obtained by model checking the MDP), or “does further training improve the NN?”. We report on an implementation of DSMC inside the \n              Modest \n              Toolset \n              in combination with externally learnt NNs, demonstrating the potential of DSMC on various instances of the model family, and illustrating its scalability as a function of instance size as well as other factors like the degree of NN training.},\n\tlanguage = {en},\n\tnumber = {3},\n\turldate = {2025-05-13},\n\tjournal = {International Journal on Software Tools for Technology Transfer},\n\tauthor = {Gros, Timo P. and Hermanns, Holger and Hoffmann, Jörg and Klauck, Michaela and Steinmetz, Marcel},\n\tmonth = jun,\n\tyear = {2023},\n\tpages = {407--426},\n}\n\n\n\n
\n
\n\n\n
\n Abstract Neural networks (NN) are taking over ever more decisions thus far taken by humans, even though verifiable system-level guarantees are far out of reach. Neither is the verification technology available, nor is it even understood what a formal, meaningful, extensible, and scalable testbed might look like for such a technology. The present paper is an attempt to improve on both the above aspects. We present a family of formal models that contain basic features of automated decision-making contexts and which can be extended with further orthogonal features, ultimately encompassing the scope of autonomous driving. Due to the possibility to model random noise in the decision actuation, each model instance induces a Markov decision process (MDP) as verification object. The NN in this context has the duty to actuate (near-optimal) decisions. From the verification perspective, the externally learnt NN serves as a determinizer of the MDP, the result being a Markov chain which as such is amenable to statistical model checking. The combination of an MDP and an NN encoding the action policy is central to what we call “deep statistical model checking” (DSMC). While being a straightforward extension of statistical model checking, it enables to gain deep insight into questions like “how high is the NN-induced safety risk?”, “how good is the NN compared to the optimal policy?” (obtained by model checking the MDP), or “does further training improve the NN?”. We report on an implementation of DSMC inside the Modest Toolset in combination with externally learnt NNs, demonstrating the potential of DSMC on various instances of the model family, and illustrating its scalability as a function of instance size as well as other factors like the degree of NN training.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Causing Intended Effects in Collaborative Decision-Making.\n \n \n \n \n\n\n \n Meyer-Vitali, A.; and Mulder, W.\n\n\n \n\n\n\n In Murukannaiah, P. K.; and Hirzle, T., editor(s), Proceedings of the Workshops at the Second International Conference on Hybrid Human-Artificial Intelligence, volume 3456, of CEUR Workshop Proceedings, pages 137–144, Munich, Germany, June 2023. CEUR\n ISSN: 1613-0073\n\n\n\n
\n\n\n\n \n \n \"CausingPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{meyer-vitali_causing_2023,\n\taddress = {Munich, Germany},\n\tseries = {{CEUR} {Workshop} {Proceedings}},\n\ttitle = {Causing {Intended} {Effects} in {Collaborative} {Decision}-{Making}},\n\tvolume = {3456},\n\tcopyright = {All rights reserved},\n\turl = {https://ceur-ws.org/Vol-3456/#short4-1},\n\tlanguage = {en},\n\turldate = {2023-08-16},\n\tbooktitle = {Proceedings of the {Workshops} at the {Second} {International} {Conference} on {Hybrid} {Human}-{Artificial} {Intelligence}},\n\tpublisher = {CEUR},\n\tauthor = {Meyer-Vitali, André and Mulder, Wico},\n\teditor = {Murukannaiah, Pradeep K. and Hirzle, Teresa},\n\tmonth = jun,\n\tyear = {2023},\n\tnote = {ISSN: 1613-0073},\n\tpages = {137--144},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Knowledge Engineering for Hybrid Intelligence.\n \n \n \n \n\n\n \n Tiddi, I.; De Boer, V.; Schlobach, S.; and Meyer-Vitali, A.\n\n\n \n\n\n\n In Proceedings of the 12th Knowledge Capture Conference 2023, of K-CAP '23, pages 75–82, New York, NY, USA, December 2023. Association for Computing Machinery\n \n\n\n\n
\n\n\n\n \n \n \"KnowledgePaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@inproceedings{tiddi_knowledge_2023,\n\taddress = {New York, NY, USA},\n\tseries = {K-{CAP} '23},\n\ttitle = {Knowledge {Engineering} for {Hybrid} {Intelligence}},\n\tcopyright = {All rights reserved},\n\tisbn = {9798400701412},\n\turl = {https://doi.org/10.1145/3587259.3627541},\n\tdoi = {10.1145/3587259.3627541},\n\tabstract = {Hybrid Intelligence (HI) is a rapidly growing field aiming at creating collaborative systems where humans and intelligent machines cooperate in mixed teams towards shared goals. A clear characterization of the tasks and knowledge exchanged by the agents in HI applications is still missing, hampering both standardization and reuse when designing new HI systems. Knowledge Engineering (KE) methods have been used to solve such issue through the formalization of tasks and roles in knowledge-intensive processes. We investigate whether KE methods can be applied to HI scenarios, and specifically whether common, reusable elements such as knowledge roles, tasks and subtasks can be identified in contexts where symbolic, subsymbolic and human-in-the-loop components are involved. We first adapt the well-known CommonKADS methodology to HI, and then use it to analyze several HI projects and identify common tasks. The results are (i) a high-level ontology of HI knowledge roles, (ii) a set of novel, HI-specific tasks and (iii) an open repository to store scenarios1 – allowing reuse, validation and design of existing and new HI applications.},\n\turldate = {2023-11-29},\n\tbooktitle = {Proceedings of the 12th {Knowledge} {Capture} {Conference} 2023},\n\tpublisher = {Association for Computing Machinery},\n\tauthor = {Tiddi, Ilaria and De Boer, Victor and Schlobach, Stefan and Meyer-Vitali, André},\n\tmonth = dec,\n\tyear = {2023},\n\tkeywords = {CommonKADS, Hybrid Intelligence, Knowledge Engineering},\n\tpages = {75--82},\n}\n\n\n\n
\n
\n\n\n
\n Hybrid Intelligence (HI) is a rapidly growing field aiming at creating collaborative systems where humans and intelligent machines cooperate in mixed teams towards shared goals. A clear characterization of the tasks and knowledge exchanged by the agents in HI applications is still missing, hampering both standardization and reuse when designing new HI systems. Knowledge Engineering (KE) methods have been used to solve such issue through the formalization of tasks and roles in knowledge-intensive processes. We investigate whether KE methods can be applied to HI scenarios, and specifically whether common, reusable elements such as knowledge roles, tasks and subtasks can be identified in contexts where symbolic, subsymbolic and human-in-the-loop components are involved. We first adapt the well-known CommonKADS methodology to HI, and then use it to analyze several HI projects and identify common tasks. The results are (i) a high-level ontology of HI knowledge roles, (ii) a set of novel, HI-specific tasks and (iii) an open repository to store scenarios1 – allowing reuse, validation and design of existing and new HI applications.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Security of AI-Systems: Fundamentals - Security Considerations for Symbolic and Hybrid AI.\n \n \n \n \n\n\n \n Müller, C.; Vogt, R.; Nonnengart, A.; Klusch, M.; and Meyer-Vitali, A.\n\n\n \n\n\n\n Technical Report Bundesamt für Sicherheit in der Informationstechnik (BSI), June 2023.\n https://www.bsi.bund.de/DE/Service-Navi/Publikationen/Studien/Projekt_P464/Projekt_P464_node.html https://www.bsi.bund.de/DE/Service-Navi/Presse/Alle-Meldungen-News/Meldungen/Studien_Cyber-Sicherheit_KI-Systeme_230202.html\n\n\n\n
\n\n\n\n \n \n \"SecurityPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@techreport{muller_security_2023,\n\ttitle = {Security of {AI}-{Systems}: {Fundamentals} - {Security} {Considerations} for {Symbolic} and {Hybrid} {AI}},\n\tcopyright = {All rights reserved},\n\tshorttitle = {Security of {AI}-{Systems}},\n\turl = {https://www.bsi.bund.de/SharedDocs/Downloads/EN/BSI/KI/Security-of-AI-systems_fundamentals_considerations_symbolic_hybrid.html},\n\tabstract = {Until recent years, cybersecurity was mostly concerned with an arms race of attackers and defence\nmechanisms on the level of admission control, protection of data transmission, cryptography, and so forth.\nOn the other side, functional safety of software-based systems dealt with faults in the system behaviour,\nwhich are caused by electro-mechanical errors in one of its sub-components or systematic errors (like bugs).\nWith AI-based systems becoming ever more wide-spread and complex, both paradigms need to be extended\nand, in a way, they are growing closer together. AI security and AI safety have a large overlap. Part of what\nAI safety tries to cope with are perturbations (or distribution shifts) that occur “naturally”, for example\nbecause the environment changes (day to night, summer to winter, Europe to Asia, simulation to reality,\netc.) or because the domain gradually evolves (demographic changes, generation changes, etc.). Seldom,\nevents can occur that have never been considered in training, causing an undesired emergent behaviour\n(misclassification, wrong decision, etc.) at inference time. Couldn’t we also say that the unexpected event\ncaused an “evasion”? AI security, aside from assuming an adversary, deals with similar problems. Data\npoisoning is the attempt to smuggle-in examples to the training set that decrease the accuracy of the system\n(or increase test error), thereby trying to be as efficient and subtle as possible. Evasion attacks alter the\ninference situation, either by manipulating the environment or otherwise making sure that the system\nreceives input that leads to misclassifications. In a sense, they are trying to create an event that was not\nexpected during training. It is possible that poisoning and evasion attacks are combined in a sense that the\npoisoning attack introduces a trigger for the later evasion. The proximity between the two problem domains\nexists on all levels: in highly automated driving, for example, it is plausible to describe a case in which the\ncar with ego-centric vision is tricked by the behaviour of another vehicle (agent) exhibiting a strange\nmanoeuvre. If we had reasons to assume that the agent’s “adversarial driving” was based on knowledge\nabout the inner working of the ego car, we would call it a security breach – otherwise a safety issue. It\nbecomes apparent that the distinction is somewhat arbitrary. Moreover, if we look at the body of literature\nin AI security, the game of finding new attacks, on the one side, and inventing new ways of defending them,\non the other, could also be framed under the umbrella of research on robustness.},\n\tinstitution = {Bundesamt für Sicherheit in der Informationstechnik (BSI)},\n\tauthor = {Müller, Christian and Vogt, Roland and Nonnengart, Andreas and Klusch, Matthias and Meyer-Vitali, André},\n\tmonth = jun,\n\tyear = {2023},\n\tnote = {https://www.bsi.bund.de/DE/Service-Navi/Publikationen/Studien/Projekt\\_P464/Projekt\\_P464\\_node.html\nhttps://www.bsi.bund.de/DE/Service-Navi/Presse/Alle-Meldungen-News/Meldungen/Studien\\_Cyber-Sicherheit\\_KI-Systeme\\_230202.html},\n}\n\n\n\n
\n
\n\n\n
\n Until recent years, cybersecurity was mostly concerned with an arms race of attackers and defence mechanisms on the level of admission control, protection of data transmission, cryptography, and so forth. On the other side, functional safety of software-based systems dealt with faults in the system behaviour, which are caused by electro-mechanical errors in one of its sub-components or systematic errors (like bugs). With AI-based systems becoming ever more wide-spread and complex, both paradigms need to be extended and, in a way, they are growing closer together. AI security and AI safety have a large overlap. Part of what AI safety tries to cope with are perturbations (or distribution shifts) that occur “naturally”, for example because the environment changes (day to night, summer to winter, Europe to Asia, simulation to reality, etc.) or because the domain gradually evolves (demographic changes, generation changes, etc.). Seldom, events can occur that have never been considered in training, causing an undesired emergent behaviour (misclassification, wrong decision, etc.) at inference time. Couldn’t we also say that the unexpected event caused an “evasion”? AI security, aside from assuming an adversary, deals with similar problems. Data poisoning is the attempt to smuggle-in examples to the training set that decrease the accuracy of the system (or increase test error), thereby trying to be as efficient and subtle as possible. Evasion attacks alter the inference situation, either by manipulating the environment or otherwise making sure that the system receives input that leads to misclassifications. In a sense, they are trying to create an event that was not expected during training. It is possible that poisoning and evasion attacks are combined in a sense that the poisoning attack introduces a trigger for the later evasion. The proximity between the two problem domains exists on all levels: in highly automated driving, for example, it is plausible to describe a case in which the car with ego-centric vision is tricked by the behaviour of another vehicle (agent) exhibiting a strange manoeuvre. If we had reasons to assume that the agent’s “adversarial driving” was based on knowledge about the inner working of the ego car, we would call it a security breach – otherwise a safety issue. It becomes apparent that the distinction is somewhat arbitrary. Moreover, if we look at the body of literature in AI security, the game of finding new attacks, on the one side, and inventing new ways of defending them, on the other, could also be framed under the umbrella of research on robustness.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n A Maturity Model for Collaborative Agents in Human-AI Ecosystems.\n \n \n \n\n\n \n Mulder, W.; and Meyer-Vitali, A.\n\n\n \n\n\n\n In Camarinha-Matos, L. M.; Boucher, X.; and Ortiz, A., editor(s), Collaborative Networks in Digitalization and Society 5.0, of IFIP Advances in Information and Communication Technology, pages 328–335, Cham, 2023. Springer Nature Switzerland\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@inproceedings{mulder_maturity_2023,\n\taddress = {Cham},\n\tseries = {{IFIP} {Advances} in {Information} and {Communication} {Technology}},\n\ttitle = {A {Maturity} {Model} for {Collaborative} {Agents} in {Human}-{AI} {Ecosystems}},\n\tcopyright = {All rights reserved},\n\tisbn = {978-3-031-42622-3},\n\tdoi = {10.1007/978-3-031-42622-3_23},\n\tabstract = {AI entities lean on the aspects of their autonomy to carry out their tasks and perform intelligently. But when these entities collaborate in human-AI teams, their levels of autonomy and collaboration have to be balanced out. We present a maturity model for agents regarding this aspect of balancing. Whereas simple AI systems use pre-designed mechanisms, more advanced systems are able to learn this from experience. The maturity model is a two-dimensional matrix in which the degree of agency forms the horizontal axis, and the level of interaction the vertical axis. We validate the use of this maturity model with use-cases in the field of urban energy efficiency.},\n\tlanguage = {en},\n\tbooktitle = {Collaborative {Networks} in {Digitalization} and {Society} 5.0},\n\tpublisher = {Springer Nature Switzerland},\n\tauthor = {Mulder, Wico and Meyer-Vitali, André},\n\teditor = {Camarinha-Matos, Luis M. and Boucher, Xavier and Ortiz, Angel},\n\tyear = {2023},\n\tkeywords = {Agency, Collaborative networks, Human-AI teaming},\n\tpages = {328--335},\n}\n\n\n\n
\n
\n\n\n
\n AI entities lean on the aspects of their autonomy to carry out their tasks and perform intelligently. But when these entities collaborate in human-AI teams, their levels of autonomy and collaboration have to be balanced out. We present a maturity model for agents regarding this aspect of balancing. Whereas simple AI systems use pre-designed mechanisms, more advanced systems are able to learn this from experience. The maturity model is a two-dimensional matrix in which the degree of agency forms the horizontal axis, and the level of interaction the vertical axis. We validate the use of this maturity model with use-cases in the field of urban energy efficiency.\n
\n\n\n
\n\n\n\n\n\n
\n
\n\n
\n
\n  \n 2022\n \n \n (6)\n \n \n
\n
\n \n \n
\n \n\n \n \n \n \n \n Conditional independence testing with heteroskedastic data and applications to causal discovery.\n \n \n \n\n\n \n Günther, W.; Ninad, U.; Wahl, J.; and Runge, J.\n\n\n \n\n\n\n Advances in Neural Information Processing Systems, 35: 16191–16202. 2022.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{gunther_conditional_2022,\n\ttitle = {Conditional independence testing with heteroskedastic data and applications to causal discovery},\n\tvolume = {35},\n\tjournal = {Advances in Neural Information Processing Systems},\n\tauthor = {Günther, Wiebke and Ninad, Urmi and Wahl, Jonas and Runge, Jakob},\n\tyear = {2022},\n\tpages = {16191--16202},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Comparing the Language of QAnon-related content on Parler, Gab, and Twitter.\n \n \n \n\n\n \n Sipka, A.; Hannak, A.; and Urman, A.\n\n\n \n\n\n\n In Proceedings of the 14th ACM Web Science Conference 2022, pages 411–421, 2022. \n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{sipka_comparing_2022,\n\ttitle = {Comparing the {Language} of {QAnon}-related content on {Parler}, {Gab}, and {Twitter}},\n\tbooktitle = {Proceedings of the 14th {ACM} {Web} {Science} {Conference} 2022},\n\tauthor = {Sipka, Andrea and Hannak, Aniko and Urman, Aleksandra},\n\tyear = {2022},\n\tpages = {411--421},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n From Responsibility to Reason-Giving Explainable Artificial Intelligence.\n \n \n \n\n\n \n Baum, K.; Mantel, S.; Schmidt, E.; and Speith, T.\n\n\n \n\n\n\n Philosophy & Technology, 35(1). 2022.\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{baum_responsibility_2022,\n\ttitle = {From {Responsibility} to {Reason}-{Giving} {Explainable} {Artificial} {Intelligence}},\n\tvolume = {35},\n\tdoi = {10.1007/s13347-022-00510-w},\n\tnumber = {1},\n\tjournal = {Philosophy \\& Technology},\n\tauthor = {Baum, Kevin and Mantel, Susanne and Schmidt, Eva and Speith, Timo},\n\tyear = {2022},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Ethics for Nerds.\n \n \n \n\n\n \n Baum, K.; and Sterz, S.\n\n\n \n\n\n\n The International Review of Information Ethics, 31(1). 2022.\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{baum_ethics_2022,\n\ttitle = {Ethics for {Nerds}},\n\tvolume = {31},\n\tdoi = {10.29173/irie484},\n\tnumber = {1},\n\tjournal = {The International Review of Information Ethics},\n\tauthor = {Baum, Kevin and Sterz, Sarah},\n\tyear = {2022},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models.\n \n \n \n \n\n\n \n Schramowski, P.; Brack, M.; Deiseroth, B.; and Kersting, K.\n\n\n \n\n\n\n 2022.\n \n\n\n\n
\n\n\n\n \n \n \"SafePaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@misc{schramowski_safe_2022,\n\ttitle = {Safe {Latent} {Diffusion}: {Mitigating} {Inappropriate} {Degeneration} in {Diffusion} {Models}},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\tshorttitle = {Safe {Latent} {Diffusion}},\n\turl = {https://arxiv.org/abs/2211.05105},\n\tdoi = {10.48550/ARXIV.2211.05105},\n\tabstract = {Text-conditioned image generation models have recently achieved astonishing results in image quality and text alignment and are consequently employed in a fast-growing number of applications. Since they are highly data-driven, relying on billion-sized datasets randomly scraped from the internet, they also suffer, as we demonstrate, from degenerated and biased human behavior. In turn, they may even reinforce such biases. To help combat these undesired side effects, we present safe latent diffusion (SLD). Specifically, to measure the inappropriate degeneration due to unfiltered and imbalanced training sets, we establish a novel image generation test bed-inappropriate image prompts (I2P)-containing dedicated, real-world image-to-text prompts covering concepts such as nudity and violence. As our exhaustive empirical evaluation demonstrates, the introduced SLD removes and suppresses inappropriate image parts during the diffusion process, with no additional training required and no adverse effect on overall image quality or text alignment.},\n\turldate = {2025-05-20},\n\tpublisher = {arXiv},\n\tauthor = {Schramowski, Patrick and Brack, Manuel and Deiseroth, Björn and Kersting, Kristian},\n\tyear = {2022},\n\tkeywords = {Artificial Intelligence (cs.AI), Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, Machine Learning (cs.LG)},\n}\n\n\n\n
\n
\n\n\n
\n Text-conditioned image generation models have recently achieved astonishing results in image quality and text alignment and are consequently employed in a fast-growing number of applications. Since they are highly data-driven, relying on billion-sized datasets randomly scraped from the internet, they also suffer, as we demonstrate, from degenerated and biased human behavior. In turn, they may even reinforce such biases. To help combat these undesired side effects, we present safe latent diffusion (SLD). Specifically, to measure the inappropriate degeneration due to unfiltered and imbalanced training sets, we establish a novel image generation test bed-inappropriate image prompts (I2P)-containing dedicated, real-world image-to-text prompts covering concepts such as nudity and violence. As our exhaustive empirical evaluation demonstrates, the introduced SLD removes and suppresses inappropriate image parts during the diffusion process, with no additional training required and no adverse effect on overall image quality or text alignment.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Trustworthy Hybrid Team Decision-Support Systems.\n \n \n \n \n\n\n \n Meyer-Vitali, A.; and Mulder, W.\n\n\n \n\n\n\n . March 2022.\n Number: 7607 Publisher: EasyChair\n\n\n\n
\n\n\n\n \n \n \"TrustworthyPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{meyer-vitali_trustworthy_2022,\n\ttitle = {Trustworthy {Hybrid} {Team} {Decision}-{Support} {Systems}},\n\tcopyright = {All rights reserved},\n\tissn = {2516-2314},\n\turl = {https://easychair.org/publications/preprint/jRqf},\n\tlanguage = {en-US},\n\turldate = {2023-08-18},\n\tauthor = {Meyer-Vitali, André and Mulder, Wico},\n\tmonth = mar,\n\tyear = {2022},\n\tnote = {Number: 7607\nPublisher: EasyChair},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n\n\n\n
\n
\n\n
\n
\n  \n 2021\n \n \n (9)\n \n \n
\n
\n \n \n
\n \n\n \n \n \n \n \n Explainability auditing for intelligent systems: a rationale for multi-disciplinary perspectives.\n \n \n \n\n\n \n Langer, M.; Baum, K.; Hartmann, K.; Hessel, S.; Speith, T.; and Wahl, J.\n\n\n \n\n\n\n In 2021 IEEE 29th international requirements engineering conference workshops (REW), pages 164–168, 2021. IEEE\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{langer_explainability_2021,\n\ttitle = {Explainability auditing for intelligent systems: a rationale for multi-disciplinary perspectives},\n\tbooktitle = {2021 {IEEE} 29th international requirements engineering conference workshops ({REW})},\n\tpublisher = {IEEE},\n\tauthor = {Langer, Markus and Baum, Kevin and Hartmann, Kathrin and Hessel, Stefan and Speith, Timo and Wahl, Jonas},\n\tyear = {2021},\n\tpages = {164--168},\n}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Optimized locking for replication solutions.\n \n \n \n\n\n \n Wilkinson, J. P; and Sipka, A.\n\n\n \n\n\n\n October 2021.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@misc{wilkinson_optimized_2021,\n\ttitle = {Optimized locking for replication solutions},\n\tauthor = {Wilkinson, John P and Sipka, Andrea},\n\tmonth = oct,\n\tyear = {2021},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Towards Perspicuity Requirements.\n \n \n \n\n\n \n Sterz, S.; Baum, K.; Lauber-Rönsberg, A.; and Hermanns, H.\n\n\n \n\n\n\n In Yue, T.; and Mirakhorli, M., editor(s), 29th IEEE International Requirements Engineering Conference Workshops (RE 2021 Workshops), Notre Dame, Indiana, USA, pages 159–163, September 2021. IEEE\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{sterz_towards_2021,\n\ttitle = {Towards {Perspicuity} {Requirements}},\n\tdoi = {10.1109/REW53955.2021.00029},\n\tbooktitle = {29th {IEEE} {International} {Requirements} {Engineering} {Conference} {Workshops} ({RE} 2021 {Workshops}), {Notre} {Dame}, {Indiana}, {USA}},\n\tpublisher = {IEEE},\n\tauthor = {Sterz, Sarah and Baum, Kevin and Lauber-Rönsberg, Anne and Hermanns, Holger},\n\teditor = {Yue, Tao and Mirakhorli, Mehdi},\n\tmonth = sep,\n\tyear = {2021},\n\tpages = {159--163},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Spare Me the Details: How the Type of Information About Automated Interviews Influences Applicant Reactions.\n \n \n \n\n\n \n Langer, M.; Baum, K.; König, C. J; Hähne, V.; Oster, D.; and Speith, T.\n\n\n \n\n\n\n International Journal of Selection and Assessment, 29(2): 154–169. 2021.\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{langer_spare_2021,\n\ttitle = {Spare {Me} the {Details}: {How} the {Type} of {Information} {About} {Automated} {Interviews} {Influences} {Applicant} {Reactions}},\n\tvolume = {29},\n\tdoi = {10.1111/ijsa.12325},\n\tnumber = {2},\n\tjournal = {International Journal of Selection and Assessment},\n\tauthor = {Langer, Markus and Baum, Kevin and König, Cornelius J and Hähne, Viviane and Oster, Daniel and Speith, Timo},\n\tyear = {2021},\n\tpages = {154--169},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n What to Expect from Opening Up ‘Black Boxes’? Comparing Perceptions of Justice Between Human and Automated Agents.\n \n \n \n\n\n \n Schlicker, N.; Langer, M.; Ötting, S. K; Baum, K.; König, C. J; and Wallach, D.\n\n\n \n\n\n\n Computers in Human Behavior, 122. 2021.\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{schlicker_what_2021,\n\ttitle = {What to {Expect} from {Opening} {Up} ‘{Black} {Boxes}’? {Comparing} {Perceptions} of {Justice} {Between} {Human} and {Automated} {Agents}},\n\tvolume = {122},\n\tdoi = {10.1016/j.chb.2021.106837},\n\tjournal = {Computers in Human Behavior},\n\tauthor = {Schlicker, Nadine and Langer, Markus and Ötting, Sonja K and Baum, Kevin and König, Cornelius J and Wallach, Dieter},\n\tyear = {2021},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n What Do We Want from Explainable Artificial Intelligence (XAI)? – A Stakeholder Perspective on XAI and a Conceptual Model Guiding Interdisciplinary XAI Research.\n \n \n \n\n\n \n Langer, M.; Oster, D.; Speith, T.; Hermanns, H.; Kästner, L.; Schmidt, E.; Sesing, A.; and Baum, K.\n\n\n \n\n\n\n Artificial Intelligence, 296. 2021.\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{langer_what_2021,\n\ttitle = {What {Do} {We} {Want} from {Explainable} {Artificial} {Intelligence} ({XAI})? – {A} {Stakeholder} {Perspective} on {XAI} and a {Conceptual} {Model} {Guiding} {Interdisciplinary} {XAI} {Research}},\n\tvolume = {296},\n\tdoi = {10.1016/j.artint.2021.103473},\n\tjournal = {Artificial Intelligence},\n\tauthor = {Langer, Markus and Oster, Daniel and Speith, Timo and Hermanns, Holger and Kästner, Lena and Schmidt, Eva and Sesing, Andreas and Baum, Kevin},\n\tyear = {2021},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Adaptive Rational Activations to Boost Deep Reinforcement Learning.\n \n \n \n \n\n\n \n Delfosse, Q.; Schramowski, P.; Mundt, M.; Molina, A.; and Kersting, K.\n\n\n \n\n\n\n In Proceedings of the International Conference on Representation Learning, 2021. \n \n\n\n\n
\n\n\n\n \n \n \"AdaptivePaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n\n\n\n
\n 
@inproceedings{delfosse_adaptive_2021,\n\ttitle = {Adaptive {Rational} {Activations} to {Boost} {Deep} {Reinforcement} {Learning}},\n\tcopyright = {Creative Commons Attribution 4.0 International},\n\turl = {https://arxiv.org/abs/2102.09407},\n\tdoi = {10.48550/ARXIV.2102.09407},\n\tabstract = {Latest insights from biology show that intelligence not only emerges from the connections between neurons but that individual neurons shoulder more computational responsibility than previously anticipated. This perspective should be critical in the context of constantly changing distinct reinforcement learning environments, yet current approaches still primarily employ static activation functions. In this work, we motivate why rationals are suitable for adaptable activation functions and why their inclusion into neural networks is crucial. Inspired by recurrence in residual networks, we derive a condition under which rational units are closed under residual connections and formulate a naturally regularised version: the recurrent-rational. We demonstrate that equipping popular algorithms with (recurrent-)rational activations leads to consistent improvements on Atari games, especially turning simple DQN into a solid approach, competitive to DDQN and Rainbow.},\n\turldate = {2025-05-21},\n\tbooktitle = {Proceedings of the {International} {Conference} on {Representation} {Learning}},\n\tauthor = {Delfosse, Quentin and Schramowski, Patrick and Mundt, Martin and Molina, Alejandro and Kersting, Kristian},\n\tyear = {2021},\n\tkeywords = {FOS: Computer and information sciences, Machine Learning (cs.LG)},\n}\n\n\n\n
\n
\n\n\n
\n Latest insights from biology show that intelligence not only emerges from the connections between neurons but that individual neurons shoulder more computational responsibility than previously anticipated. This perspective should be critical in the context of constantly changing distinct reinforcement learning environments, yet current approaches still primarily employ static activation functions. In this work, we motivate why rationals are suitable for adaptable activation functions and why their inclusion into neural networks is crucial. Inspired by recurrence in residual networks, we derive a condition under which rational units are closed under residual connections and formulate a naturally regularised version: the recurrent-rational. We demonstrate that equipping popular algorithms with (recurrent-)rational activations leads to consistent improvements on Atari games, especially turning simple DQN into a solid approach, competitive to DDQN and Rainbow.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Modular design patterns for hybrid learning and reasoning systems.\n \n \n \n \n\n\n \n van Bekkum, M.; de Boer, M.; van Harmelen, F.; Meyer-Vitali, A.; and Teije, A. t.\n\n\n \n\n\n\n Applied Intelligence, 51(9): 6528–6546. September 2021.\n \n\n\n\n
\n\n\n\n \n \n \"ModularPaper\n  \n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@article{van_bekkum_modular_2021,\n\ttitle = {Modular design patterns for hybrid learning and reasoning systems},\n\tvolume = {51},\n\tcopyright = {All rights reserved},\n\tissn = {1573-7497},\n\turl = {https://doi.org/10.1007/s10489-021-02394-3},\n\tdoi = {10.1007/s10489-021-02394-3},\n\tabstract = {The unification of statistical (data-driven) and symbolic (knowledge-driven) methods is widely recognized as one of the key challenges of modern AI. Recent years have seen a large number of publications on such hybrid neuro-symbolic AI systems. That rapidly growing literature is highly diverse, mostly empirical, and is lacking a unifying view of the large variety of these hybrid systems. In this paper, we analyze a large body of recent literature and we propose a set of modular design patterns for such hybrid, neuro-symbolic systems. We are able to describe the architecture of a very large number of hybrid systems by composing only a small set of elementary patterns as building blocks. The main contributions of this paper are: 1) a taxonomically organised vocabulary to describe both processes and data structures used in hybrid systems; 2) a set of 15+ design patterns for hybrid AI systems organized in a set of elementary patterns and a set of compositional patterns; 3) an application of these design patterns in two realistic use-cases for hybrid AI systems. Our patterns reveal similarities between systems that were not recognized until now. Finally, our design patterns extend and refine Kautz’s earlier attempt at categorizing neuro-symbolic architectures.},\n\tlanguage = {en},\n\tnumber = {9},\n\turldate = {2022-03-29},\n\tjournal = {Applied Intelligence},\n\tauthor = {van Bekkum, Michael and de Boer, Maaike and van Harmelen, Frank and Meyer-Vitali, André and Teije, Annette ten},\n\tmonth = sep,\n\tyear = {2021},\n\tpages = {6528--6546},\n}\n\n\n\n
\n
\n\n\n
\n The unification of statistical (data-driven) and symbolic (knowledge-driven) methods is widely recognized as one of the key challenges of modern AI. Recent years have seen a large number of publications on such hybrid neuro-symbolic AI systems. That rapidly growing literature is highly diverse, mostly empirical, and is lacking a unifying view of the large variety of these hybrid systems. In this paper, we analyze a large body of recent literature and we propose a set of modular design patterns for such hybrid, neuro-symbolic systems. We are able to describe the architecture of a very large number of hybrid systems by composing only a small set of elementary patterns as building blocks. The main contributions of this paper are: 1) a taxonomically organised vocabulary to describe both processes and data structures used in hybrid systems; 2) a set of 15+ design patterns for hybrid AI systems organized in a set of elementary patterns and a set of compositional patterns; 3) an application of these design patterns in two realistic use-cases for hybrid AI systems. Our patterns reveal similarities between systems that were not recognized until now. Finally, our design patterns extend and refine Kautz’s earlier attempt at categorizing neuro-symbolic architectures.\n
\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Modular Design Patterns for Hybrid Actors.\n \n \n \n \n\n\n \n Meyer-Vitali, A.; Mulder, W.; and de Boer, M. H. T.\n\n\n \n\n\n\n In Cooperative AI Workshop, volume 2021, of NeurIPS, December 2021. \n arXiv: 2109.09331\n\n\n\n
\n\n\n\n \n \n \"ModularPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n  \n \n abstract \n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n \n \n \n \n \n \n\n\n\n
\n 
@inproceedings{meyer-vitali_modular_2021,\n\tseries = {{NeurIPS}},\n\ttitle = {Modular {Design} {Patterns} for {Hybrid} {Actors}},\n\tvolume = {2021},\n\tcopyright = {Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International Licence (CC-BY-NC-ND)},\n\turl = {http://arxiv.org/abs/2109.09331},\n\tabstract = {Recently, a boxology (graphical language) with design patterns for hybrid AI was proposed, combining symbolic and sub-symbolic learning and reasoning. In this paper, we extend this boxology with actors and their interactions. The main contributions of this paper are: 1) an extension of the taxonomy to describe distributed hybrid AI systems with actors and interactions; and 2) showing examples using a few design patterns relevant in multi-agent systems and human-agent interaction.},\n\turldate = {2021-11-17},\n\tbooktitle = {Cooperative {AI} {Workshop}},\n\tauthor = {Meyer-Vitali, André and Mulder, Wico and de Boer, Maaike H. T.},\n\tmonth = dec,\n\tyear = {2021},\n\tnote = {arXiv: 2109.09331},\n\tkeywords = {Computer Science - Artificial Intelligence, Computer Science - Multiagent Systems, Computer Science - Software Engineering},\n}\n\n\n\n
\n
\n\n\n
\n Recently, a boxology (graphical language) with design patterns for hybrid AI was proposed, combining symbolic and sub-symbolic learning and reasoning. In this paper, we extend this boxology with actors and their interactions. The main contributions of this paper are: 1) an extension of the taxonomy to describe distributed hybrid AI systems with actors and interactions; and 2) showing examples using a few design patterns relevant in multi-agent systems and human-agent interaction.\n
\n\n\n
\n\n\n\n\n\n
\n
\n\n
\n
\n  \n 2020\n \n \n (4)\n \n \n
\n
\n \n \n
\n \n\n \n \n \n \n \n Using AI to help healthcare professionals stay up-to-date with medical research.\n \n \n \n\n\n \n de Bie, K.; Kishore, N.; Rentsch, A.; Rosado, P.; and Sipka, A.\n\n\n \n\n\n\n In AI for Social Good Workshop, 2020. \n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{de_bie_using_2020,\n\ttitle = {Using {AI} to help healthcare professionals stay up-to-date with medical research},\n\tbooktitle = {{AI} for {Social} {Good} {Workshop}},\n\tauthor = {de Bie, Kim and Kishore, Nishant and Rentsch, Anthony and Rosado, Pablo and Sipka, Andrea},\n\tyear = {2020},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Checking data integrity of data storage systems.\n \n \n \n\n\n \n Wilkinson, J.; and Sipka, A.\n\n\n \n\n\n\n January 2020.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@misc{wilkinson_checking_2020,\n\ttitle = {Checking data integrity of data storage systems},\n\tpublisher = {Google Patents},\n\tauthor = {Wilkinson, John and Sipka, Andrea},\n\tmonth = jan,\n\tyear = {2020},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Multiple point-in-time copies on a remote system.\n \n \n \n\n\n \n Sipka, A.; and Wilkinson, J. P\n\n\n \n\n\n\n October 2020.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@misc{sipka_multiple_2020,\n\ttitle = {Multiple point-in-time copies on a remote system},\n\tauthor = {Sipka, Andrea and Wilkinson, John P},\n\tmonth = oct,\n\tyear = {2020},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Point-in-time copy on a remote system.\n \n \n \n\n\n \n Sipka, A.; and Wilkinson, J. P\n\n\n \n\n\n\n March 2020.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@misc{sipka_point--time_2020,\n\ttitle = {Point-in-time copy on a remote system},\n\tpublisher = {Google Patents},\n\tauthor = {Sipka, Andrea and Wilkinson, John P},\n\tmonth = mar,\n\tyear = {2020},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n\n\n\n
\n
\n\n
\n
\n  \n 2019\n \n \n (6)\n \n \n
\n
\n \n \n
\n \n\n \n \n \n \n \n Collision detection at multi-node storage sites.\n \n \n \n\n\n \n Rostagni, F. C; Sipka, A.; and Wilkinson, J. P\n\n\n \n\n\n\n January 2019.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@misc{rostagni_collision_2019,\n\ttitle = {Collision detection at multi-node storage sites},\n\tpublisher = {Google Patents},\n\tauthor = {Rostagni, Florent C and Sipka, Andrea and Wilkinson, John P},\n\tmonth = jan,\n\tyear = {2019},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Handling node failure in multi-node data storage systems.\n \n \n \n\n\n \n Rostagni, F.; Sipka, A.; and Wilkinson, J.\n\n\n \n\n\n\n October 2019.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@misc{rostagni_handling_2019,\n\ttitle = {Handling node failure in multi-node data storage systems},\n\tpublisher = {Google Patents},\n\tauthor = {Rostagni, Florent and Sipka, Andrea and Wilkinson, John},\n\tmonth = oct,\n\tyear = {2019},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Self-describing volume ancestry for data synchronization.\n \n \n \n\n\n \n Sipka, A.; and Wilkinson, J. P\n\n\n \n\n\n\n December 2019.\n \n\n\n\n
\n\n\n\n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@misc{sipka_self-describing_2019,\n\ttitle = {Self-describing volume ancestry for data synchronization},\n\tpublisher = {Google Patents},\n\tauthor = {Sipka, Andrea and Wilkinson, John P},\n\tmonth = dec,\n\tyear = {2019},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n \n Anforderungen an die Erklärbarkeit maschinengestützter Entscheidungen.\n \n \n \n \n\n\n \n Sesing, A.; and Baum, K.\n\n\n \n\n\n\n In Taeger, J., editor(s), Die Macht der Daten und der Algorithmen – Regulierung von IT, IoT und KI. Tagungsband DSRI-Herbstakademie 2019, pages 435–449, 2019. \n \n\n\n\n
\n\n\n\n \n \n \"AnforderungenPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{sesing_anforderungen_2019,\n\ttitle = {Anforderungen an die {Erklärbarkeit} maschinengestützter {Entscheidungen}},\n\turl = {http://olwir.de/?content=reihen/uebersicht&sort=tb&isbn=978-3-95599-061-9},\n\tbooktitle = {Die {Macht} der {Daten} und der {Algorithmen} – {Regulierung} von {IT}, {IoT} und {KI}. {Tagungsband} {DSRI}-{Herbstakademie} 2019},\n\tauthor = {Sesing, Andreas and Baum, Kevin},\n\teditor = {Taeger, Jürgen},\n\tyear = {2019},\n\tpages = {435--449},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Towards a framework combining machine ethics and machine explainability.\n \n \n \n\n\n \n Baum, K.; Hermanns, H.; and Speith, T.\n\n\n \n\n\n\n In Finkbeiner, B.; and Kleinberg, S., editor(s), Proceedings of the 3rd Workshop on Formal Reasoning about Causation, Responsibility, and Explanations in Science and Technology (CREST 2018), Thessaloniki, Greece, 21st April 2018, 2019. \n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{baum_towards_2019,\n\ttitle = {Towards a framework combining machine ethics and machine explainability},\n\tdoi = {10.4204/EPTCS.286.4},\n\tbooktitle = {Proceedings of the 3rd {Workshop} on {Formal} {Reasoning} about {Causation}, {Responsibility}, and {Explanations} in {Science} and {Technology} ({CREST} 2018), {Thessaloniki}, {Greece}, 21st {April} 2018},\n\tauthor = {Baum, Kevin and Hermanns, Holger and Speith, Timo},\n\teditor = {Finkbeiner, Bernd and Kleinberg, Samantha},\n\tyear = {2019},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n
\n \n\n \n \n \n \n \n Explainability as a Non-Functional Requirement.\n \n \n \n\n\n \n Köhl, M. A; Baum, K.; Langer, M.; Oster, D.; Speith, T.; and Bohlender, D.\n\n\n \n\n\n\n In 27th IEEE International Requirements Engineering Conference (RE 2019), Jeju Island, South Korea, pages 363–368, 2019. IEEE\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{kohl_explainability_2019,\n\ttitle = {Explainability as a {Non}-{Functional} {Requirement}},\n\tdoi = {10.1109/RE.2019.00046},\n\tbooktitle = {27th {IEEE} {International} {Requirements} {Engineering} {Conference} ({RE} 2019), {Jeju} {Island}, {South} {Korea}},\n\tpublisher = {IEEE},\n\tauthor = {Köhl, Maximilian A and Baum, Kevin and Langer, Markus and Oster, Daniel and Speith, Timo and Bohlender, Dimitri},\n\tyear = {2019},\n\tpages = {363--368},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n\n\n\n
\n
\n\n
\n
\n  \n 2018\n \n \n (1)\n \n \n
\n
\n \n \n
\n \n\n \n \n \n \n \n \n From Machine Ethics To Machine Explainability and Back.\n \n \n \n \n\n\n \n Baum, K.; Hermanns, H.; and Speith, T.\n\n\n \n\n\n\n In International Symposium on Artificial Intelligence and Mathematics (ISAIM 2018), Fort Lauderdale, Florida, USA, January 2018. \n \n\n\n\n
\n\n\n\n \n \n \"FromPaper\n  \n \n\n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{baum_machine_2018,\n\ttitle = {From {Machine} {Ethics} {To} {Machine} {Explainability} and {Back}},\n\turl = {https://isaim2018.cs.ou.edu/papers/ISAIM2018_Ethics_Baum_etal.pdf},\n\tbooktitle = {International {Symposium} on {Artificial} {Intelligence} and {Mathematics} ({ISAIM} 2018), {Fort} {Lauderdale}, {Florida}, {USA}},\n\tauthor = {Baum, Kevin and Hermanns, Holger and Speith, Timo},\n\tmonth = jan,\n\tyear = {2018},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n\n\n\n
\n
\n\n
\n
\n  \n 2017\n \n \n (1)\n \n \n
\n
\n \n \n
\n \n\n \n \n \n \n \n Two Challenges for CI Trustworthiness and How to Address Them.\n \n \n \n\n\n \n Baum, K.; Köhl, M.; and Schmidt, E.\n\n\n \n\n\n\n In Pereira-Fariña, M.; and Reed, C., editor(s), Proceedings of the 1st Workshop on Explainable Computational Intelligence (XCI 2017), Dundee, United Kingdom, September 2017. Association for Computational Linguistics\n \n\n\n\n
\n\n\n\n \n\n \n \n doi\n  \n \n\n \n link\n  \n \n\n bibtex\n \n\n \n\n \n\n \n \n \n \n \n \n \n\n  \n \n \n\n\n\n
\n 
@inproceedings{baum_two_2017,\n\taddress = {Dundee, United Kingdom},\n\ttitle = {Two {Challenges} for {CI} {Trustworthiness} and {How} to {Address} {Them}},\n\tdoi = {10.18653/v1/W17-3701},\n\tbooktitle = {Proceedings of the 1st {Workshop} on {Explainable} {Computational} {Intelligence} ({XCI} 2017)},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Baum, Kevin and Köhl, Maximilian and Schmidt, Eva},\n\teditor = {Pereira-Fariña, M. and Reed, C.},\n\tmonth = sep,\n\tyear = {2017},\n}\n\n\n\n
\n
\n\n\n\n
\n\n\n\n\n\n
\n
\n\n\n\n\n
\n\n\n \n\n \n \n \n \n\n
\n"}; document.write(bibbase_data.data);