UKP Lab

Quantum AI and NLP Conference 2025 Quantum AI and NLP Conference 2025 Website. The conference will be held from the 6th to the 8th of August 2025 in Bloomington, Indiana at Indiana University.

The conference brought together talented researchers at the intersection of quantum computing and #NLP, sparking ideas and discussions that push the field forward.

➡️ Check out the event: qnlp.ai

(3/3)

#NLProc #UKPLab #TUDarmstadt #QuantumComputing

Quantum AI and NLP Conference 2025 Quantum AI and NLP Conference 2025 Website. The conference will be held from the 6th to the 8th of August 2025 in Bloomington, Indiana at Indiana University.

This work was made possible through a great collaboration with

👩‍🔬 Anna Schroeder & Mariami Gachechiladze (Quantum Computing Group, @tuda.bsky.social)

👨‍🔬 Yue Zhang (Westlake University)

👩‍💻 @igurevych.bsky.social (@ukplab.bsky.social, @tuda.bsky.social)

(2/🧵)

Federico Tiblias from the UKP Lab gives a presentation at the Quantum AI and NLP 2025 conference. He stands at a podium in front of a projection screen displaying a slide titled Hamiltonian Classifier with formulas and diagrams. An inset portrait of Federico Tiblias appears in the upper right corner. The UKP Lab logo and text “UKP Lab @ Quantum AI and NLP 2025” are shown at the bottom. Audience members are seated in the foreground.

✨ Exciting news from the #QNLP Conference 2025 in Bloomington, Indiana!

@akatief.bsky.social gave a talk on “An Efficient Quantum Classifier Based on Hamiltonian Representations” — exploring how quantum computing can open new frontiers for natural language processing.

(1/🧵)

Consider following the authors @jitaishik.bsky.social, Prottay Kumar Adhikary, Hiba Arnaout, @igurevych.bsky.social, and @tanmoy-chak.bsky.social

(3/3)

#NLProc #NLP4MentalHealth #NLPsych

We expose key gaps and chart how richer, more inclusive datasets can unlock the full potential of Mental Health AI.

📃Paper: arxiv.org/pdf/2508.09809

(2/🧵)

Flowchart illustrating aspects of clinical mental health datasets. Starting from a dataset icon, branches include: Disorders: depression, anxiety, PTSD, bipolar disorder, schizophrenia. Accessibility: public, controlled, private. Tasks: classification, questionnaire scores, response generation, analysis. Synthetic Data: artificial data creation. Modalities: single modality (text, audio, video, brain signals) and multiple modality combinations. Languages/Cultures: diversity of linguistic and cultural contexts.

👉 Interested in working on AI for mental health? This is a 𝗺𝘂𝘀𝘁-𝗿𝗲𝗮𝗱 𝘀𝗶𝗻𝗰𝗲 𝘁𝗵𝗲 𝗱𝗮𝘁𝗮 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗶𝘀 𝗮 𝗺𝗮𝗷𝗼𝗿 𝗯𝗼𝘁𝘁𝗹𝗲𝗻𝗲𝗰𝗸!

Our survey maps the mental health dataset landscape, covering a wide range of disorders, tasks, modalities, accessibility, synthetic data, and cultural reach.

(1/🧵)

📚 Discovery Science 2025 is part of the AI for Science Conference in #Ljubljana

🔗 ds2025.ijs.si/speakers/iry...

#DiscoveryScience2025 #AIforScience #AI #UKPLab

Her talk “𝘗𝘭𝘦𝘢𝘴𝘦 𝘮𝘦𝘦𝘵 𝘈𝘐, 𝘰𝘶𝘳 𝘥𝘦𝘢𝘳 𝘯𝘦𝘸 𝘤𝘰𝘭𝘭𝘦𝘢𝘨𝘶𝘦. 𝘐𝘯 𝘰𝘵𝘩𝘦𝘳 𝘸𝘰𝘳𝘥𝘴: 𝘤𝘢𝘯 𝘴𝘤𝘪𝘦𝘯𝘵𝘪𝘴𝘵𝘴 𝘢𝘯𝘥 𝘮𝘢𝘤𝘩𝘪𝘯𝘦𝘴 𝘵𝘳𝘶𝘭𝘺 𝘤𝘰𝘰𝘱𝘦𝘳𝘢𝘵𝘦?” addresses how AI and LLMs can support scientists throughout different stages of the research process – from writing and revising manuscripts to peer review and scientific discussions.

Promotional graphic for the Discovery Science 2025 International Conference keynote. The text reads: “Please meet AI, our dear new colleague. In other words: can scientists and machines truly cooperate?” Keynote by Prof. Iryna Gurevych. On the right is a photo of Prof. Iryna Gurevych standing in front of bookshelves. The logos of UKP Lab and Discovery Science are included at the bottom.

🚀 𝗞𝗲𝘆𝗻𝗼𝘁𝗲 𝗯𝘆 𝗣𝗿𝗼𝗳. 𝗜𝗿𝘆𝗻𝗮 𝗚𝘂𝗿𝗲𝘃𝘆𝗰𝗵 𝗶𝗻 𝗟𝗷𝘂𝗯𝗹𝗷𝗮𝗻𝗮
Prof. @igurevych.bsky.social, Head of the UKP Lab, is giving a keynote today at the 28th International Conference on Discovery Science 2025 in Ljubljana.

Stay tuned as Yang reflects on his research path, current projects, and the lasting impact of his time at UKP Lab 💻💡.

(5/5)

#UKPLab #UKPAlumni #LanguageModels #RLHF #ReinforcementLearning #DeepMind #Gemini #NLProc #AI #TextGeneration #ComputationalLinguistics

Before joining Google DeepMind in 2021, Yang held positions at Royal Holloway, University of London and at the Chinese Academy of Sciences, following a PhD at @imperialcollegeldn.bsky.social & and postdoc at @ukplab.bsky.social.

(4/🧵)

During his time at UKP, Yang and his collaborators were among the first to explore 𝗥𝗟 𝘄𝗶𝘁𝗵 𝗛𝘂𝗺𝗮𝗻 𝗙𝗲𝗲𝗱𝗯𝗮𝗰𝗸 #RLHF for text generation, pioneering techniques that are foundational to today’s large language models. His work highlighted the importance of 𝗿𝗲𝘄𝗮𝗿𝗱 𝗺𝗼𝗱𝗲𝗹𝗶𝗻𝗴 in RL-based NLP systems.

(3/🧵)

Yang joined the UKP Lab as a postdoctoral researcher in 2017, bringing with him a background in 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 #RL. His time at UKP marked a turning point: it was here that he began integrating RL with #NLP — an approach that would go on to shape much of his later work.

(2/🧵)

Graphic for the UKP Lab Alumni Spotlight series. It features a circular portrait photo of Yang Gao, smiling and wearing glasses and a jacket, with a blurred building in the background. Text above the image reads: “Spotlight on UKP Lab Alumni.” Below the portrait: “Yang Gao, Research Scientist at Google DeepMind.” The bottom right corner shows the Ubiquitous Knowledge Processing (UKP) Lab logo.

🚀 We’re excited to continue our 𝘚𝘱𝘰𝘵𝘭𝘪𝘨𝘩𝘵 𝘰𝘯 𝘜𝘒𝘗 𝘓𝘢𝘣 𝘈𝘭𝘶𝘮𝘯𝘪 series with Yang Gao, now a Research Scientist at #Google #DeepMind, where he works on the Google #Gemini language models.

(1/🧵)

🔁 Prof. Iryna Gurevych, Director of the UKP Lab at TU Darmstadt, is featured in this ATHENE video portrait.

🎥 The video (in German) introduces her research on disinformation and cognitive security, where methods from #NLP intersect with questions of #cybersecurity.

⬇️ Watch the video portrait here

And consider following the authors Furkan Şahinuç, Subhabrata Dutta, @igurevych.bsky.social for more information or an exchange of ideas. (8/8)

For further details, check our project page, paper, dataset, and code!

🌐Project webpage: ukplab.github.io/arxiv2025-ex...
📄Paper: arxiv.org/abs/2508.07955
💻Code: github.com/UKPLab/arxiv...
📁 Data: tudatalib.ulb.tu-darmstadt.de/handle/tudat...

(7/🧵)

𝗪𝗵𝘆 𝗗𝗼𝗲𝘀 𝗜𝘁 𝗠𝗮𝘁𝘁𝗲𝗿?
GREP bridges a gap in evaluating AI-generated scientific text—moving from superficial scoring to expert-aligned evaluation. Whether you’re interested in RLHF, domain-specific generation, or enhancing scientific writing pipelines, this work offers insight and robust tools.
(6/🧵)

𝗥𝗲𝘀𝘂𝗹𝘁 𝘁𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀: Strong LLMs often fail to satisfy hard constraints and struggle most with maintaining citation coherence, highlighting the need for iterative, feedback-driven human-AI collaboration to refine outputs. (5/🧵)

GREP distinguishes between hard constraints (non-negotiable requirements, such as correct citations) and soft constraints (stylistic preferences and readability). This layered evaluation better reflects how experts truly assess quality in scientific writing. (4/🧵)

GREP goes beyond standard automatic metrics and LLM-as-a-judge systems—tailored for general NLP tasks—failing to account for expert preferences and domain-specific quality standards. (3/🧵)

We propose 𝗚𝗥𝗘𝗣—a novel, preference-aligned, multi-turn evaluation framework aimed at elevating how we assess and refine automatically generated Related Work sections in scientific writing. (2/🧵)

Diagram illustrating a workflow for generating and evaluating a “Related Work” section using large language models (LLMs). Inputs: Cited papers and an author drafting a related work section, along with criteria or preferences and a gold-standard related work section. Generation step: Multiple LLM-based generators (depicted as small robots in different colors) produce draft sections. Evaluation step: An LLM-supported evaluation system (robot with judge’s gavel) compares the drafts against criteria and the gold section. Ranking step: The generated drafts are ranked on a podium (1st, 2nd, 3rd place). Feedback loop: Feedback is sent back to the generators to improve performance. The figure shows an iterative system combining generation, evaluation, and ranking for producing high-quality related work sections.

Fail to utilize LLM-as-a-judge systems for evaluating expert writing tasks?🤔

They often don’t know what to judge (missing expert criteria) and don’t know how to judge (lacking structured evaluation). In our new paper, we integrate task-specific criteria with expert preferences. 🚀
(1/🧵)

👏 Congratulations to all authors and collaborators for their excellent work! We are looking forward to presenting these results at EMNLP 2025 in #Suzhou this November. Stay tuned for more details!

#NLProc #MachineLearning #UKPLab #Research #EMNLP2025

🔟 𝘐𝘥𝘦𝘯𝘵𝘪𝘧𝘺𝘪𝘯𝘨 𝘈𝘴𝘱𝘦𝘤𝘵𝘴 𝘪𝘯 𝘗𝘦𝘦𝘳 𝘙𝘦𝘷𝘪𝘦𝘸𝘴
Sheng Lu, @ikuznetsov.bsky.social , @igurevych.bsky.social

9️⃣ 𝘓𝘦𝘢𝘬𝘺 𝘛𝘩𝘰𝘶𝘨𝘩𝘵𝘴: 𝘓𝘢𝘳𝘨𝘦 𝘙𝘦𝘢𝘴𝘰𝘯𝘪𝘯𝘨 𝘔𝘰𝘥𝘦𝘭𝘴 𝘈𝘳𝘦 𝘕𝘰𝘵 𝘗𝘳𝘪𝘷𝘢𝘵𝘦 𝘛𝘩𝘪𝘯𝘬𝘦𝘳𝘴
Tommaso Green, @mgubri.bsky.social , Haritz Puerto, Sangdoo Yun, @coallaoh.bsky.social

8️⃣ 𝘛𝘶𝘳𝘯𝘪𝘯𝘨 𝘓𝘰𝘨𝘪𝘤 𝘈𝘨𝘢𝘪𝘯𝘴𝘵 𝘐𝘵𝘴𝘦𝘭𝘧: 𝘗𝘳𝘰𝘣𝘪𝘯𝘨 𝘔𝘰𝘥𝘦𝘭 𝘋𝘦𝘧𝘦𝘯𝘴𝘦𝘴 𝘛𝘩𝘳𝘰𝘶𝘨𝘩 𝘊𝘰𝘯𝘵𝘳𝘢𝘴𝘵𝘪𝘷𝘦 𝘘𝘶𝘦𝘴𝘵𝘪𝘰𝘯𝘴
Rachneet Sachdeva, Rima Hazra, @igurevych.bsky.social

7️⃣ $\𝘵𝘦𝘹𝘵𝘵𝘵{𝘋𝘳𝘰𝘪𝘥}$: 𝘈 𝘙𝘦𝘴𝘰𝘶𝘳𝘤𝘦 𝘚𝘶𝘪𝘵𝘦 𝘧𝘰𝘳 𝘈𝘐-𝘎𝘦𝘯𝘦𝘳𝘢𝘵𝘦𝘥 𝘊𝘰𝘥𝘦 𝘋𝘦𝘵𝘦𝘤𝘵𝘪𝘰𝘯
Daniil Orel, @ineil77.bsky.social, @igurevych.bsky.social, Preslav Nakov

6️⃣ 𝘛𝘩𝘦 𝘎𝘰𝘰𝘥, 𝘵𝘩𝘦 𝘉𝘢𝘥 𝘢𝘯𝘥 𝘵𝘩𝘦 𝘊𝘰𝘯𝘴𝘵𝘳𝘶𝘤𝘵𝘪𝘷𝘦: 𝘈𝘶𝘵𝘰𝘮𝘢𝘵𝘪𝘤𝘢𝘭𝘭𝘺 𝘔𝘦𝘢𝘴𝘶𝘳𝘪𝘯𝘨 𝘗𝘦𝘦𝘳 𝘙𝘦𝘷𝘪𝘦𝘸'𝘴 𝘜𝘵𝘪𝘭𝘪𝘵𝘺 𝘧𝘰𝘳 𝘈𝘶𝘵𝘩𝘰𝘳𝘴
Abdelrahman "Boda" Sadallah, @timbmg.bsky.social, @igurevych.bsky.social, Ted Briscoe

5️⃣ 𝘍𝘳𝘰𝘮 𝘚𝘶𝘳𝘷𝘦𝘺𝘴 𝘵𝘰 𝘕𝘢𝘳𝘳𝘢𝘵𝘪𝘷𝘦𝘴: 𝘙𝘦𝘵𝘩𝘪𝘯𝘬𝘪𝘯𝘨 𝘊𝘶𝘭𝘵𝘶𝘳𝘢𝘭 𝘝𝘢𝘭𝘶𝘦 𝘈𝘥𝘢𝘱𝘵𝘢𝘵𝘪𝘰𝘯 𝘪𝘯 𝘓𝘓𝘔𝘴
Muhammad Farid Adilazuarda, @ccliu.bsky.social, @igurevych.bsky.social, Alham Fikri Aji