UKP Lab

@ukplab.bsky.social

1142 Followers 341 Following

The Ubiquitous Knowledge Processing Lab researches Natural Language Processing (#NLProc) with a strong emphasis on Large Language Models, Conversational AI & Question Answering | @cs-tudarmstadt.bsky.social · @TUDa.bsky.social https://www.ukp.tu-darmstadt

Darya Hryhoryeva holding the SemEval-2026 Best Paper Award certificate for 'UKP_Psycontrol at SemEval-2026 Task 2: Modeling Valence and Arousal Dynamics from Text' at ACL 2026.

Preslav Nakov (left) and Iryna Gurevych (right) standing in front of a projection screen at TU Darmstadt on June 19, 2026, before Preslav Nakov's talk titled 'Human or AI? Interpretable Detection, Multilingual Perception, and the Paradox of Human-Like Text' from Mohamed bin Zayed University of Artificial Intelligence (MBZUAI).

Welcome graphic for Andrea Moleri, Doctoral researcher at UKP Lab. His portrait photo is framed in a red circle, with the UKP Lab and TU Darmstadt logos and the word "Welcome" repeated four times.

Black-and-white illustration of a person wearing a graduation cap, examining a document with a magnifying glass.Black-and-white illustration of a person wearing a graduation cap, examining a document with a magnifying glass. A robotic parrot sits on their shoulder.

Welcome graphic from UKP Lab at TU Darmstadt introducing Phu Gia Hoang as a new Doctoral Researcher, with his portrait next to a 'Welcome' message on a background of scattered letter tiles.

Welcome graphic from the UKP Lab featuring Kurt Micallef, labeled “Postdoctoral researcher.” On the left is a circular portrait of a man standing outdoors at sunset. On the right, the word “WELCOME” appears repeatedly in the background. Logos of the Ubiquitous Knowledge Processing Lab and Technische Universität Darmstadt are included, along with social media icons and the handle @UKPLab.

Promotional graphic announcing a “Major Release” of CARE with the phrase “Opening Dev.” The word “CARE” is split into two blocks: “CA” inside a speech bubble icon and “RE” in a yellow rectangle. The background features faint icons related to research and writing, such as books, charts, links, and documents.

Diagram illustrating a social bias evaluation pipeline for large language models. On the left, an LLM is transformed via a quantizer into a 4-bit quantized LLM. Both models are then evaluated within a “Social Bias Evaluation Framework” that includes categories such as stereotype (SS, RB, WB, BBQ), toxicity (BLD, DTT), sentiment (BLD), and fairness (DE, DEG, DTF). Arrows lead to bar charts on the right labeled “Social Bias Level,” comparing bias levels across models.

Group photo of eight conference participants standing in front of an EACL 2026 backdrop in Rabat, Morocco. They wear name badges and casual to semi-formal attire, posing indoors against a banner that reads “EACL 2026 Rabat, Morocco, March 24–29, 2026” with sponsor logos and decorative patterns.

Promotional graphic for a guest talk featuring Prof. Yova Kementchedjhieva. A portrait of a smiling woman with curly hair, arms crossed, is framed in a red circle against a background of keyboard keys spelling “USE.” Text reads “Prof. Yova Kementchedjhieva” and “MBZUAI Guest Talk.” Logos of the Ubiquitous Knowledge Processing Lab and Technische Universität Darmstadt appear, along with social media icons and the handle @UKPLab.

Young student in a suit stands at a science fair booth, holding a small device with a face-like display and grass on top. Behind him, posters labeled “Regionalsieger” present diagrams and results of the project. On the table in front are hardware components, printed schematics, and a laptop displaying the device interface. Other participants and exhibition stands are visible in the background.

Schematic diagram of analyzing scientific papers. On the left, a document contains color-coded sections representing claim, conclusion, result, and method. In the center, these categories are listed, with dashed lines linking them to corresponding parts of the document. A red cross marks an incorrect linkage between a result and a section in a second document shown on the right, where one passage is highlighted. At the bottom, small robot figures are shown reviewing documents, indicating automated evaluation or peer review processes.

Illustration of the CultureCare framework for culturally aware emotional support in AI systems. The diagram shows a distress message from a 17-year-old in Sudan describing anger issues and stigma around therapy, with cultural signals and emotional distress highlighted. It compares responses: a detailed human response offering supportive suggestions, a generic LLM response acknowledging emotional coping, and an adapted LLM response that incorporates cultural context and suggests speaking with a trusted family member or friend.

Diagram illustrating how AI analyzes peer review discussions. A reviewer asks whether reviewers disagree on the strengths of a paper. The system summarizes that reviewers 1 and 2 both evaluate the method but differ in their assessment, with one seeing it as well-designed and the other noting issues. Another question asks whether other reviewers consider these issues important; reviewer 3 mentions them but treats them as minor concerns. Side labels indicate steps of correlating opinions and weighting arguments.

Three-panel diagram explaining an audit method for LLM unlearning.
Panel A, “LLM Unlearning,” shows a trained language model and a modified “unlearned” model, highlighting that output-based metrics may give a misleading sense of privacy.
Panel B, “Our Audit,” illustrates comparing the two models using a Venn diagram to decompose mutual information into unlearned information and residual information.
Panel C, “Practical Impact,” shows that residual information can correlate with vulnerability to adversarial attacks, and suggests inference-time privacy protection through abstention.

Three-step workflow diagram for evaluating retrieval approaches.
Step 1, “Data generation and validation,” shows synthetic data being aligned with natural data across two domains: peer reviews and news.
Step 2, “Automatic evaluation,” compares multiple approaches (10 retrievers and 3 LLMs) and produces a shortlist of the best-performing models.
Step 3, “Assisted labeling,” illustrates human annotators reviewing candidate outputs and marking them as correct or incorrect.
The final result is the selection of the best approach and the creation of labeled data.

Infographic titled “The Hidden Complexity Behind ‘Not Novel Enough’,” explaining why novelty assessment in peer review is difficult at scale. The graphic shows a researcher surrounded by stacks of papers and question marks, illustrating uncertainty in evaluating novelty. On the left, challenges include fragmented literature, reviewer overload (14+ papers per year), and unclear novelty criteria. On the right, resulting issues include inconsistent decisions, author confusion, and weak scientific feedback. The infographic concludes that novelty assessment requires searching, comparing, and judging prior work under severe time constraints.

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.

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.

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.

Table titled “Taxonomy for evaluation of AI in mental health applications,” organized into columns for quality criteria (validity and reliability) and real-world use (implementation and maintenance). Rows distinguish support types: assessment, intervention, and information synthesis. Each cell lists detailed evaluation questions, such as construct and criterion validity, consistency across populations and time, feasibility, effectiveness, usability, acceptability, safety, and unintended consequences, providing a structured framework for assessing AI systems in mental health contexts.

Research poster for the paper “SPARE: Single-Pass Annotation with Reference-Guided Evaluation for Automatic Process Supervision and Reward Modelling,” featuring logos of UKP Lab, hessian.AI, and Queen’s University. The poster lists the authors and affiliations, and includes a photo of a researcher standing beside the poster at a conference. At the bottom, a diagram explains the SPARE framework, showing inputs such as prompts and reference solutions, and outputs related to evaluation and alignment.

A speaker stands at the front of a lecture room giving a talk, gesturing with both hands. Behind them, a projected slide titled “Instruction vs. Retrieval – A Subtle Difference” explains risks related to prompt-based attacks and retrieval-based manipulation. The slide includes diagrams about LLM agents, public opinion pools, and coordinated ideological input. A laptop and water bottle sit on the podium beside the presenter.

A researcher stands in a lecture room giving a presentation, gesturing toward a projected slide titled “AgentFly: Extensible and Scalable Reinforcement Learning for LM Agents.” The slide is shared via a video conferencing interface visible at the top, showing remote participants. The presenter stands next to a laptop on a small podium, with the projection filling the wall behind them.

Stylized illustration of a researcher wearing a graduation cap and lab coat, holding a magnifying glass and reading a document. A small robotic parrot sits on the researcher’s shoulder, suggesting AI assistance. The image combines themes of academic research, analysis, and human–AI collaboration.

Promotional graphic for a keynote on “Roadmap to Internet Security.” The background shows a glowing digital lock and circuit-board imagery. Large text announces the talk title “How to Make AI-Native Internet Content Secure? Coping with Synthetic and Misleading Data.” On the right is a circular portrait of a woman with shoulder-length curly hair and glasses (Iryna Gurevych), looking toward the camera.

Recruitment graphic for ATHENE, the National Research Center for Applied Cybersecurity. The image shows a digital scene with a human hand and a robotic hand reaching toward a glowing shield icon with a padlock symbol. Overlaid text reads “WE ARE HIRING! HIRING! HIRING!”. Below, a QR code and the text “Postdoc and PhD positions – HAICC: Human-AI Collaboration for Cybersecurity” are displayed, along with the logos of Technische Universität Darmstadt and Goethe University Frankfurt.