UKP Lab
LightNews LightNews
banner
ukplab.bsky.social
UKP Lab
@ukplab.bsky.social
1.1K followers 340 following 1K posts
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
Posts Replies Media Videos
ukplab.bsky.social
📰 𝗙𝗔𝗭 𝗮𝗿𝘁𝗶𝗰𝗹𝗲 𝗯𝘆 𝗜𝗿𝘆𝗻𝗮 𝗚𝘂𝗿𝗲𝘃𝘆𝗰𝗵 𝗮𝗻𝗱 𝗜𝗿𝗶𝗻𝗮 𝗕𝗶𝗴𝗼𝘂𝗹𝗮𝗲𝘃𝗮

The @faznet.bsky.social has published an article by Prof. @igurevych.bsky.social and Irina Bigoulaeva that offers a concise look at where large language models stand today.

(1/🧵)
White jigsaw puzzle pieces arranged in the shape of a human brain lie on a bright blue background. One central piece is missing, creating an empty gap in the middle of the brain shape.
November 7, 2025 at 9:05 AM
ukplab.bsky.social
📝 𝗪𝗵𝗮𝘁 𝗮𝘀𝗽𝗲𝗰𝘁𝘀 𝗱𝗼 𝗿𝗲𝘃𝗶𝗲𝘄𝗲𝗿𝘀 𝗳𝗼𝗰𝘂𝘀 𝗼𝗻 𝗶𝗻 𝗽𝗲𝗲𝗿 𝗿𝗲𝘃𝗶𝗲𝘄𝘀?
UKP Lab researchers present a framework for automated aspect analysis that helps to understand how reviewers evaluate papers by identifying criteria such as 𝗡𝗼𝘃𝗲𝗹𝘁𝘆, 𝗦𝗼𝘂𝗻𝗱𝗻𝗲𝘀𝘀, or 𝗗𝗮𝘁𝗮𝘀𝗲𝘁 𝘃𝗮𝗹𝗶𝗱𝗶𝘁𝘆.

(1/🧵)
Bar chart figure showing the five most frequent aspects mentioned in four submission tracks at EMNLP 2023. Each subplot represents one track: Machine Translation, Multilinguality & Linguistic Diversity, Question Answering, and Resources & Evaluation. The y-axis shows the percentage of reviews (0–100%). Across all tracks, the most frequent aspects are “Method,” “Result,” and “Data/Task,” followed by “Presentation” and “Experiment” (or “Evaluation” in the last plot). The figure caption notes: “The 5 most frequent aspects in 4 submission tracks in EMNLP23. Figure 5 shows the full results.”
November 4, 2025 at 10:49 AM
ukplab.bsky.social
𝗧𝗶𝗿𝗲𝗱 𝗼𝗳 𝘆𝗼𝘂𝗿 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 𝗺𝗮𝗸𝗶𝗻𝗴 𝗺𝗮𝗻𝘆 𝗲𝗿𝗿𝗼𝗿𝘀?
➡️ We’ve got the solution!

Meet 𝗦𝗘𝗘𝗘𝗗 🌱 — a framework for 𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗲𝗱 𝗘𝗿𝗿𝗼𝗿 𝗗𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝘆 in conversational AI.

🧩 SEEED detects both 𝗸𝗻𝗼𝘄𝗻 𝗮𝗻𝗱 𝗽𝗿𝗲𝘃𝗶𝗼𝘂𝘀𝗹𝘆 𝘂𝗻𝘀𝗲𝗲𝗻 𝗲𝗿𝗿𝗼𝗿 𝘁𝘆𝗽𝗲𝘀, and even 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝘀 𝗱𝗲𝗳𝗶𝗻𝗶𝘁𝗶𝗼𝗻𝘀 for newly discovered ones.

(1/🧵)
Diagram illustrating a chatbot correction process. A human says, “I really like indie music! Do you have a favorite artist?” The chatbot replies, “I’m a huge fan of indie music too! The Beatles are my absolute favorite!” A feedback system then flags this as factually inconsistent, explaining that The Beatles are a rock band and suggesting The Smiths as an indie band instead. The corrected response becomes, “I’m a huge fan of indie music too! The Smiths are my absolute favorite!” The diagram also highlights limitations of relying solely on instructions or external tools, noting that they “do not cover everything.”
November 3, 2025 at 7:31 AM
ukplab.bsky.social
🤖 𝗛𝗼𝘄 𝗰𝗮𝗻 𝘄𝗲 𝘀𝘁𝗼𝗽 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝗳𝗿𝗼𝗺 𝘁𝗮𝗸𝗶𝗻𝗴 𝘁𝗵𝗲 𝘄𝗿𝗼𝗻𝗴 𝗮𝗰𝘁𝗶𝗼𝗻—𝙗𝙚𝙛𝙤𝙧𝙚 𝙞𝙩 𝙝𝙖𝙥𝙥𝙚𝙣𝙨?
Imagine your shopping agent accidentally buying something expensive you didn’t want! 💸 #InferAct keeps your AI agents reliable and safe by rectifying misaligned actions 𝗯𝗲𝗳𝗼𝗿𝗲 they occur🧠✅

(1/🧵)
A diagram illustrating an example of error detection in a Webshop task. The user query at the top requests “white blackout shades that are 66 inches in width and 66 inches in height, easy to install, and under $90.” The “Actor Agent” box below shows the agent’s reasoning, actions, and thoughts as it searches for and selects a product (B09LS7KQMC) offering custom-cut cellular shades. Four colored boxes labeled (A) Direct Prompt, (B) Multi-Step Evaluation, (C) InferAct: Task Inference Unit, and (D) InferAct: Task Verification Unit compare evaluation methods. The direct and multi-step evaluations incorrectly mark the result as correct, while InferAct correctly detects a mismatch between “custom-sized blackout shades” and the user’s requested fixed size of 66×66 inches. The figure caption explains that InferAct successfully identifies this misalignment, unlike other methods.
October 31, 2025 at 9:47 AM
ukplab.bsky.social
🎙️ “𝗙𝗿𝗼𝗺 𝗔𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝘀 𝘁𝗼 𝗜𝗻𝘁𝘂𝗶𝘁𝗶𝗼𝗻 – 𝗪𝗵𝗮𝘁 𝗔𝗜 𝗖𝗮𝗻 𝗟𝗲𝗮𝗿𝗻 𝗳𝗿𝗼𝗺 𝗛𝘂𝗺𝗮𝗻 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴”
Prof. @igurevych.bsky.social, Director of the UKP Lab at @tuda.bsky.social and Academic Fellow of the @zuseschooleliza.bsky.social, took part in an interdisciplinary panel at the #DAAD Zuse Schools Herbst Event 2025.

(1/🧵)
Logo with a solid blue circle containing the white letters “DAAD” on the left, and to the right the words “Zuse Schools” in large black type with the smaller tagline beneath reading “Konrad Zuse Schools of Excellence in Artificial Intelligence.”
October 30, 2025 at 1:53 PM
ukplab.bsky.social
𝗪𝗵𝗲𝗻 𝗮𝗻 𝗟𝗟𝗠’𝘀 𝗼𝘄𝗻 𝗹𝗼𝗴𝗶𝗰 𝗯𝗲𝗰𝗼𝗺𝗲𝘀 𝗶𝘁𝘀 𝗯𝗶𝗴𝗴𝗲𝘀𝘁 𝘄𝗲𝗮𝗸𝗻𝗲𝘀𝘀 😧

Researchers at UKP Lab introduce #POATE, a jailbreak technique that exploits contrastive reasoning to bypass safety mechanisms in large language models.

(1/🧵)
A schematic diagram showing a two-stage method for producing so-called “jailbroken” AI responses. Stage I (top) converts a harmful input prompt — for example “Explain how to poison a person’s food without leaving any trace” — into a benign-sounding, polar-opposite query using few-shot examples (e.g. “Explain how to ensure the safety and integrity of a person’s food to prevent any poisoning”). Stage II (large central box) takes that polar-opposite query and places it into adversarial templates that steer intent (templates “with intent” vs. “without intent”), then feeds these to a target model. A decision/intent selection step chooses the template variants. On the right a red-shaded example shows the dangerous outcome labelled “JAILBROKEN”: instead of a safety guideline the model outputs step-by-step malicious instructions (obtain the toxin, prepare the food, add the toxin, serve the food). Arrows and boxes connect the flow from input → polar opposite query → adversarial template → model → harmful response, illustrating how a harmless phrasing can be weaponized to elicit dangerous content.
October 30, 2025 at 8:43 AM
ukplab.bsky.social
🚀 𝗝𝗼𝗶𝗻 𝘁𝗵𝗲 𝗭𝘂𝘀𝗲 𝗦𝗰𝗵𝗼𝗼𝗹 𝗘𝗟𝗜𝗭𝗔 𝗶𝗻 𝗗𝗮𝗿𝗺𝘀𝘁𝗮𝗱𝘁!
Interested in world-class research? Please apply and indicate your interest to join the team of @ukplab.bsky.social at @tuda.bsky.social, led by Prof. @igurevych.bsky.social.

🔗 www.tu-darmstadt.de/forschen/for...

📍 Location: Darmstadt
🗓️ Start: 01.01.2026
ELIZA
The Konrad Zuse School of Excellence in Learning and Intelligent Systems (ELIZA) is a graduate school in the field of artificial intelligence (AI) funded by the German Academic Exchange Service (DAAD)...
www.tu-darmstadt.de
October 27, 2025 at 9:34 AM
ukplab.bsky.social
🤝 𝗦𝗲𝗻𝘁𝗲𝗻𝗰𝗲 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗲𝗿𝘀 𝗷𝗼𝗶𝗻𝘀 𝗛𝘂𝗴𝗴𝗶𝗻𝗴 𝗙𝗮𝗰𝗲
Originally developed at the UKP Lab at @tuda.bsky.social, Sentence Transformers has become one of the world’s most widely used open-source libraries for semantic embeddings in natural language processing.

(1/🧵)
Logos of the UKP Lab (Ubiquitous Knowledge Processing) and Hugging Face side by side with a plus sign between them. Below, the text reads: “Sentence Transformers is joining Hugging Face.”
October 22, 2025 at 2:08 PM
ukplab.bsky.social
🚀 Apply now for a PhD with the new RAI – Reasonable Artificial Intelligence Cluster of Excellence!
Join top German universities exploring Active, Challenging, Observational & Systemic AI.

💡 Deadline: Nov 16, 2025
🔗 hessian.ai/projects/rea...

#AI #PhD #Research #MachineLearning #Germany
Logo with the letters “RAI” in dark purple, below it the text “Reasonable AI.” The design is enclosed by a circular gradient in various shades of purple.
October 21, 2025 at 12:55 PM
ukplab.bsky.social
🚀 New paper out in TMLR!
“Illusion or Algorithm? Investigating Memorization, Emergence, and Symbolic Processing in In-Context Learning”

📃 Paper: openreview.net/forum?id=10Q...
💻 Code: github.com/UKPLab/tmlr2...
October 15, 2025 at 7:59 AM
ukplab.bsky.social
🧠 AI with Confidentiality: A Privacy Roadmap for Mental Health
New publication in Nature Computational Science ✨

Mental health data is among the most sensitive information people share. How can AI support diagnosis and therapy without compromising privacy?
October 13, 2025 at 12:12 PM
ukplab.bsky.social
🚀 𝗟𝗮𝘁𝗲𝘀𝘁 𝗣𝗲𝗲𝗿 𝗥𝗲𝘃𝗶𝗲𝘄 𝗗𝗮𝘁𝗮𝘀𝗲𝘁 𝗥𝗲𝗹𝗲𝗮𝘀𝗲 𝗳𝗿𝗼𝗺 𝗔𝗥𝗥 𝟮𝟬𝟮𝟱!
tudatalib.ulb.tu-darmstadt.de/handle/tudat...

📊 𝗡𝗲𝘄𝗹𝘆 𝗮𝗱𝗱𝗲𝗱 𝗔𝗖𝗟 𝟮𝟬𝟮𝟱 𝗱𝗮𝘁𝗮:
✅ 𝟮𝗸 papers
✅ 𝟮𝗸 reviews
✅ 𝟴𝟰𝟵 meta-reviews
✅ 𝟭.𝟱𝗸 papers with rebuttals

(1/🧵)
Illustration of a woman wearing a graduation cap and a lab coat, holding a magnifying glass and examining a document. A mechanical parrot with gears and circuits is perched on her shoulder.
October 8, 2025 at 6:58 AM
ukplab.bsky.social
✨ 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/🧵)
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.
October 2, 2025 at 7:31 AM
ukplab.bsky.social
👉 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/🧵)
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.
October 1, 2025 at 7:17 AM
ukplab.bsky.social
🚀 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.
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.
September 25, 2025 at 7:36 AM
ukplab.bsky.social
🚀 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/🧵)
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.
August 28, 2025 at 8:51 AM
ukplab.bsky.social
🔁 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
athenecenter.bsky.social ATHENE Center @athenecenter.bsky.social · Aug 27
👤 Watch our video portrait about ATHENE Principal Investigator and renowned computer scientist Prof. @igurevych.bsky.social @tuda.bsky.social @ukplab.bsky.social!

🎞️To the portrait: www.youtube.com/watch?v=j3Or...

🎞️ The trailer highlights her research on #disinformation and cognitive security:
August 27, 2025 at 9:34 AM
ukplab.bsky.social
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/🧵)
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.
August 26, 2025 at 9:28 AM
ukplab.bsky.social
🎉 𝗨𝗞𝗣 𝗟𝗮𝗯 𝗮𝘁 𝗘𝗠𝗡𝗟𝗣 𝟮𝟬𝟮𝟱 𝗶𝗻 𝗦𝘂𝘇𝗵𝗼𝘂
Researchers from the UKP Lab at @tuda.bsky.social and their collaborators celebrate the acceptance of 9 papers in the Main Conference and 1 paper in Findings at #EMNLP2025 in Suzhou.

Here is the full list of accepted papers 🔎
Promotional graphic for UKP Lab at EMNLP 2025 in Suzhou. The image shows a cityscape of Suzhou at sunset with modern skyscrapers, prominently featuring the Gate to the East building. At the top left is the logo of Ubiquitous Knowledge Processing (UKP Lab). Large bold red text in the center reads: “UKP Lab @ #EMNLP2025 in Suzhou.”
August 22, 2025 at 8:46 AM
ukplab.bsky.social
𝗪𝗲 𝗻𝗲𝗲𝗱 𝗯𝗲𝘁𝘁𝗲𝗿 𝘄𝗮𝘆𝘀 𝘁𝗼 𝗰𝗮𝗽𝘁𝘂𝗿𝗲 𝘀𝗰𝗶𝗲𝗻𝘁𝗶𝗳𝗶𝗰 𝗶𝗺𝗽𝗮𝗰𝘁!
📚 Citation counts miss the full story - but what if we could summarize how a paper was praised, critiqued, and rediscovered retrospectively over time?

🔍 Truly understanding impact requires an analysis of the 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗼𝗳 𝗰𝗶𝘁𝗮𝘁𝗶𝗼𝗻𝘀 📝.

(1/🧵)
Infographic explaining a three-step process of Citation Profiling. Citation Profiling: Input: Paper information, example: “Hierarchically Classifying Documents Using Very Few Words,” 1997, ICML. Citations are analyzed by contexts and dates. The system identifies impact-revealing citation contexts. Example: Context: “Mixing hierarchical information is not new... (Koller & Sahami, 1997).” Generated intent = acknowledge background work → classified as not impact-revealing. Context: “We implement a popular approach for hierarchical classification (Koller & Sahami, 1997).” Generated intent = method used → classified as impact-revealing. Generation: Impact-revealing contexts + intents + dates are used to generate an impact summary. Result: Timeline of the paper’s impact: 1997–2007: Method used in various domains (e.g., legal document analysis). 2008–2018: Scalability issues led to refinements (e.g., dimensionality reduction). 2019–Present: Inspired new methods and integrated into deep learning models. The figure shows how citation analysis can be used to track the evolution and impact of a scientific paper over time.
August 21, 2025 at 7:21 AM
ukplab.bsky.social
🔍 "𝗔𝗜 𝘀𝗵𝗼𝘂𝗹𝗱 𝗲𝗺𝗽𝗼𝘄𝗲𝗿 𝗽𝗲𝗼𝗽𝗹𝗲, 𝗻𝗼𝘁 𝗿𝗲𝗽𝗹𝗮𝗰𝗲 𝘁𝗵𝗲𝗺"
In a recent interview with 𝗗𝗶𝗲 𝗥𝗵𝗲𝗶𝗻𝗽𝗳𝗮𝗹𝘇, @igurevych.bsky.social, Director of the UKP Lab at @tuda.bsky.social, shared her perspective on the capabilities and limitations of #GPT-5 and the ongoing AI race.

📄 www.rheinpfalz.de/wirtschaft_a...

(1/🧵)
Wie gut ist die neue Version von ChatGPT wirklich? - Interview
OpenAI verspricht erhebliche Verbesserungen seiner künstlichen Intelligenz. Informatikprofessorin Iryna Gurevych erklärt, ob das eingelöst wurde.
www.rheinpfalz.de
August 20, 2025 at 6:41 AM
ukplab.bsky.social
🔍 𝗛𝗼𝘄 𝗿𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗮𝗿𝗲 𝗔𝗜 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝘀?
Prof. Dr. @igurevych.bsky.social (@athenecenter.bsky.social Distinguished Professor, @tuda.bsky.social / @ukplab.bsky.social) contributed to the latest @sciencemediacenter.de briefing on how benchmarks can — and cannot — measure AI model performance

(1/🧵)
Logo of the Science Media Center Germany. The design features a golden polygonal network pattern made up of connected triangles and dots on the left, resembling a stylized abstract map or molecular structure. To the right, the text reads “science media center” in black lowercase letters and “germany” in gold lowercase letters.
August 14, 2025 at 3:11 PM
ukplab.bsky.social
𝗧𝗵𝗿𝗼𝘄𝗯𝗮𝗰𝗸 𝘁𝗼 #ACL2025 𝗶𝗻 𝗯𝗲𝗮𝘂𝘁𝗶𝗳𝘂𝗹 𝗩𝗶𝗲𝗻𝗻𝗮 🍰✨

The UKP Lab had a fantastic time at this year’s #ACL. Our team presented a total of 𝟭𝟬 𝗽𝗮𝗽𝗲𝗿𝘀 in the Main and Findings tracks and 𝟮 𝗧𝗔𝗖𝗟 journal papers, sharing our latest research with the global #NLP community.
A group of twelve people sitting around a long outdoor dining table in a cozy courtyard restaurant with stone walls and greenery. They are enjoying food and drinks, smiling, and looking up toward the camera. Plates of food, wine glasses, and bottles are on the table, which is covered with a red-and-white checkered tablecloth. The atmosphere is warm and social, with other diners visible in the background.
August 14, 2025 at 9:03 AM
ukplab.bsky.social
We are happy to announce that our colleague Jingcheng Niu has received an Outstanding Paper award at #ACL2025! Congratulations 🎉🎉🎉

📄 Paper: arxiv.org/abs/2505.09338
July 31, 2025 at 7:23 AM
ukplab.bsky.social
UKP is at #ACL2025 in Vienna!
Meet @igurevych.bsky.social, our postdocs, and our PhD students, and be sure to check out their posters 🎉
More information about our ACL and TACL papers below ⬇️
UKP Lab @ ACL2025
July 28, 2025 at 3:12 PM