UKP Lab

#UKPLab #NLP #NLProc #InformationRetrieval #DenseRetrieval #MultiHop #FactChecking #QuestionAnswering #RAG @cs-tudarmstadt.bsky.social

👥 Justus-Jonas Erker (@ukplab.bsky.social/@tuda.bsky.social), Nils Reimers (@cohere.com), Iryna Gurevych @igurevych.bsky.social (@ukplab.bsky.social/ @tuda.bsky.social)

See you at #EACL2026 in Rabat 🕌!

GRITHopper: Decomposition-Free Multi-Hop Dense Retrieval State-of-the-art multi-hop dense retrieval without query decomposition. Accepted at EACL 2026. Code, model & paper available.

📄 𝗚𝗥𝗜𝗧𝗛𝗼𝗽𝗽𝗲𝗿: 𝗗𝗲𝗰𝗼𝗺𝗽𝗼𝘀𝗶𝘁𝗶𝗼𝗻-𝗙𝗿𝗲𝗲 𝗠𝘂𝗹𝘁𝗶-𝗛𝗼𝗽 𝗗𝗲𝗻𝘀𝗲 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹
🌐 Project Website: ukplab.github.io/eacl2026-Gri...
🔗 Paper (arXiv): arxiv.org/pdf/2503.07519
🔗 Code: github.com/UKPLab/GritH...
🔗 Model: huggingface.co/UKPLab/GritH...

🔗 𝗨𝗻𝗶𝗳𝗶𝗲𝗱 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴: bridges dense retrieval and generative objectives, showing how post-retrieval generation loss improves retrieval

🧠 𝗦𝗲𝗹𝗳-𝘀𝘁𝗼𝗽𝗽𝗶𝗻𝗴 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴: uses generative capabilities inspired by ReAct to control its own state and determine when to stop

📌 𝗪𝗵𝗮𝘁 𝗺𝗮𝗸𝗲𝘀 𝗚𝗿𝗶𝘁𝗛𝗼𝗽𝗽𝗲𝗿 𝘀𝘁𝗮𝗻𝗱 𝗼𝘂𝘁?
✅ 𝗘𝗻𝗰𝗼𝗱𝗲𝗿-𝗼𝗻𝗹𝘆 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆: each hop requires just a single forward pass
🌍 𝗢𝘂𝘁-𝗼𝗳-𝗱𝗶𝘀𝘁𝗿𝗶𝗯𝘂𝘁𝗶𝗼𝗻 𝗿𝗼𝗯𝘂𝘀𝘁𝗻𝗲𝘀𝘀: stronger generalization than existing decomposition-free baselines

GritHopper’s joint training objective combines contrastive learning for embedding similarity with causal language modeling for next-token prediction.

💡 𝗚𝗿𝗶𝘁𝗛𝗼𝗽𝗽𝗲𝗿 is designed to deliver both efficiency and robustness, while enabling strong multi-hop reasoning without explicit decomposition.

🧩 𝗖𝗼𝗿𝗲 𝗶𝗱𝗲𝗮: unified learning for retrieval and reasoning

⚡ Decomposition-free methods improve efficiency, but often struggle with long reasoning chains and generalization beyond their training data.

So how can we get the best of both worlds?

🚨 Introducing 𝗚𝗿𝗶𝘁𝗛𝗼𝗽𝗽𝗲𝗿, the new State-of-the-Art Multi-Hop Dense Retriever 🦗

Current approaches to multi-hop retrieval face critical trade-offs.

🔎 Decomposition-based methods break complex queries into simpler steps, but they are computationally expensive and difficult to train end-to-end.

👏 Congratulations on this achievement and all the best for Cecilia’s new role as postdoctoral researcher at the @cam.ac.uk!

#NLP #PhDDefense #MultilingualAI #CulturalAI #LanguageModels #UKPLab #TUDarmstadt @cs-tudarmstadt.bsky.social

👥 Jury: @igurevych.bsky.social (TU Darmstadt), Fajri Koto (MBZUAI (Mohamed bin Zayed University of Artificial Intelligence)), Anna Rohrbach (TU Darmstadt), @justusthies.bsky.social (TU Darmstadt)

Cecilia’s thesis marks a strong contribution to 𝗺𝘂𝗹𝘁𝗶𝗹𝗶𝗻𝗴𝘂𝗮𝗹 𝗮𝗻𝗱 𝗰𝘂𝗹𝘁𝘂𝗿𝗮𝗹𝗹𝘆 𝗮𝘄𝗮𝗿𝗲 𝗡𝗟𝗣. It is an important step toward more reliable and context sensitive language technologies.

Her work further demonstrates how situated evaluation and social interaction settings can deepen our understanding of cultural competence in language models and help create more nuanced and context sensitive NLP systems.

... how language models can better generalize across languages and adapt to cultural contexts. Combining parameter efficient fine tuning with insights from cultural learning theory, she shows how targeted training interventions can improve robustness and multilingual performance.

Smiling person standing in a doorway wearing a handmade graduation cap decorated with stickers and photos. They hold the cap with one hand, carry a tote bag over their shoulder, and hold a smartphone in the other hand. A hallway with chairs and large windows is visible in the background.

🎓 𝗗𝗶𝘀𝘀𝗲𝗿𝘁𝗮𝘁𝗶𝗼𝗻 𝗗𝗲𝗳𝗲𝗻𝘀𝗲 – 𝗖𝗲𝗰𝗶𝗹𝗶𝗮
Congratulations to @ccliu.bsky.social on the successful PhD defense at the UKP Lab of @tuda.bsky.social

In her thesis 𝘎𝘦𝘯𝘦𝘳𝘢𝘭𝘪𝘻𝘢𝘵𝘪𝘰𝘯 𝘢𝘯𝘥 𝘈𝘥𝘢𝘱𝘵𝘢𝘵𝘪𝘰𝘯 𝘰𝘧 𝘓𝘢𝘯𝘨𝘶𝘢𝘨𝘦 𝘔𝘰𝘥𝘦𝘭𝘴 𝘵𝘰 𝘓𝘢𝘯𝘨𝘶𝘢𝘨𝘦𝘴 𝘢𝘯𝘥 𝘊𝘶𝘭𝘵𝘶𝘳𝘦𝘴, Cecilia investigates ...

#UKPLab #TUDarmstadt #NLP #NLProc #AIinEducation #AcademicWriting #Feedback #HigherEducation #LLMentor #CSDeptDarmstadt #hessianAI

The team behind the project are Prof. @igurevych.bsky.social, Dr. Thomas Arnold, and Dennis Zyska at @ukplab.bsky.social, as well as Prof. Florian Müller (Mobile Human-Computer Interaction, @cs-tudarmstadt.bsky.social). It has been partially supported by @hessianai.bsky.social

Exposía: Academic Writing Assessment of Exposés and Peer Feedback We present Exposía, the first public dataset that connects writing and feedback assessment in higher education, enabling research on educationally grounded approaches to academic writing evaluation. E...

This work is based on 𝗘𝘅𝗽𝗼𝘀í𝗮, a dataset documenting the full course process from draft essays to final revised versions. It enables systematic research on where AI can reliably support academic writing assessment.

📄 𝗘𝘅𝗽𝗼𝘀í𝗮: arxiv.org/abs/2601.06536
🔗 𝗖𝗔𝗥𝗘: care.ukp.informatik.tu-darmstadt.de

✅ 𝗪𝗵𝗮𝘁 𝗟𝗟𝗠𝗲𝗻𝘁𝗼𝗿 𝗱𝗼𝗲𝘀
suggests preliminary scores for evaluation criteria
provides short justifications
proposes feedback

🙅 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗱𝗼𝗲𝘀 𝗻𝗼𝘁 𝗱𝗼
it does 𝗻𝗼𝘁 grade automatically
it does 𝗻𝗼𝘁 replace lecturers or tutors

All decisions remain with the instructors 👩‍🏫

AI assistance in large lecture courses Learning scientific writing requires precise feedback, which is a particular challenge in very large courses with many participants. In the Department of Computer Science at TU Darmstadt, the AI assis...

📰 @tuda.bsky.social just published a press release about the project, the research and the team behind it:
➡️ www.tu-darmstadt.de/universitaet...

It is integrated into the existing 𝗖𝗔𝗥𝗘 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺 and provides transparent suggestions for assessing research exposes and peer-feedback written by students.

University classroom scene in which students sit at desks while a lecturer presents at the front. One student raises a hand while interacting with a large transparent digital interface showing a document and evaluation tools for academic writing, illustrating AI-assisted feedback and assessment during a seminar.

🎓 𝗔𝗜 𝘁𝗼 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 𝗰𝗼𝘂𝗿𝘀𝗲 𝗶𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗼𝗿𝘀 𝗻𝗼𝘁 𝘁𝗼 𝗿𝗲𝗽𝗹𝗮𝗰𝗲 𝘁𝗵𝗲𝗺!

In 𝘄𝗶𝗻𝘁𝗲𝗿 𝘀𝗲𝗺𝗲𝘀𝘁𝗲𝗿 𝟮𝟬𝟮𝟱/𝟮𝟲, our AI-based system 𝗟𝗟𝗠𝗲𝗻𝘁𝗼𝗿 has supported teaching the course 𝘐𝘯𝘵𝘳𝘰𝘥𝘶𝘤𝘵𝘪𝘰𝘯 𝘵𝘰 𝘚𝘤𝘪𝘦𝘯𝘵𝘪𝘧𝘪𝘤 𝘞𝘰𝘳𝘬 at @cs-tudarmstadt.bsky.social

Thanks a lot to everyone for the support, guidance, mentoring, collaboration, and great moments over the past years! 🙏 Without you, this journey wouldn't have been such a pleasure — and now excited to see what the future brings! 🚀

Problem Solving Through Human–AI Preference-based Cooperation Abstract. While there is a widespread belief that artificial general intelligence—or even superhuman AI—is imminent, complex problems in expert domains are far from being solved. We argue that such pr...

💻 Project page:
ukplab.github.io/arxiv2025-ex...

🔁 Both papers build on HAI-Co2 described in this position paper:
direct.mit.edu/coli/article...

#NLP #NLProc #LLMs #Evaluation #PreferenceLearning #ScientificWriting #HumanAI #HaiCo2 @tuda.bsky.social @cs-tudarmstadt.bsky.social

Expert Preference-based Evaluation of Automated Related Work Generation Expert domain writing, such as scientific writing, typically demands extensive domain knowledge. Although large language models (LLMs) show promising potential in this task, evaluating the quality of ...

📑 𝗘𝘅𝗽𝗲𝗿𝘁 𝗣𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗯𝗮𝘀𝗲𝗱 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 𝗼𝗳 𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗲𝗱 𝗥𝗲𝗹𝗮𝘁𝗲𝗱 𝗪𝗼𝗿𝗸 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻
Furkan Şahinuç, Subhabrata Dutta, Iryna Gurevych

arxiv.org/abs/2508.07955

Reward Modeling for Scientific Writing Evaluation Scientific writing is an expert-domain task that demands deep domain knowledge, task-specific requirements and reasoning capabilities that leverage the domain knowledge to satisfy the task specificati...

🧾 𝗥𝗲𝘄𝗮𝗿𝗱 𝗠𝗼𝗱𝗲𝗹𝗶𝗻𝗴 𝗳𝗼𝗿 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝗳𝗶𝗰 𝗪𝗿𝗶𝘁𝗶𝗻𝗴 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻
Furkan Şahinuç, Subhabrata Dutta, Iryna Gurevych

arxiv.org/abs/2601.11374

Scientific writing is an open-end expert task utilizing user preferences. Evaluating it means capturing both general and expert-specific quality criteria. Our new works push forward preference-aligned evaluation.

Poster titled “Reward Modeling for Scientific Writing Evaluation & Expert Preference Based Evaluation of Automated Related Work Generation,” with authors Furkan Sahinuç, Subhabrata Dutta, and Iryna Gurevych. The graphic illustrates a workflow: cited papers are used by an author to create a gold standard related work section and evaluation criteria, while AI generators produce draft related work sections. An LLM-supported evaluation system ranks outputs based on expert preferences, with feedback loops to improve generation quality.

📣🧪 𝗡𝗲𝘄 𝗿𝗲𝘀𝘂𝗹𝘁𝘀: 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗻𝗴 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝗳𝗶𝗰 𝗪𝗿𝗶𝘁𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗘𝘅𝗽𝗲𝗿𝘁 𝗣𝗿𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝘀
👉 Today we are presenting our first two papers that build on HAI-Co2, our recent position paper on human-ai text co-construction in expert domains:

We thank Andreas for his contributions to the Lab and wish him all the best for his future!

#NLP #PhDDefense #ComputationalArgumentation #Reliability #Interpretability #UKPLab #TUDarmstadt #UniTuebingen #LLMs #NLProc

👥 Jury: Marcus Rohrbach (@tuda.bsky.social), @igurevych.bsky.social (@tuda.bsky.social), Fajri Koto (MBZUAI (Mohamed bin Zayed University of Artificial Intelligence), Simone Schaub-Meyer (@tuda.bsky.social), Yufang Hou (@ituaustria.bsky.social)