Institute of Formal and Applied Linguistics

Dny AI jsou tady! Do programu národního festivalu @dny.ai se zapojují i vědci z ÚFAL MFF UK: 🔹 T. Sourada: AI: Hrozba nebo dar? 🔹 R. Rosa: AI for UNI 🔹 M. Popel: Jak hodnotit výkony generativní AI v dějepise 📍 Program a odkaz na akce: dny.ai #DnyAI #UFAL #Matfyz #AI #NLP

mff.cuni.czNejlépe hodnocení pedagogové v akademickém roce 2025/2026V úterý 6. října byly vyhlášeny výsledky studentské ankety o nejlépe hodnocené pedagogy loňského akademického roku.

Milan Straka is one of the top-rated teachers at MFF in 2025/26 www.mff.cuni.cz/cs/verejnost.... Congratulations! Your remarkable track record over the past nine years clearly demonstrates your excellent work. Very well deserved.

Our NLG group led by @tuetschek.bsky.social was at #INLG2026 in Utrecht (Oct 17–21) with 2 long papers, 1 short paper, 1 system demo and 4 poster presentations. And a best short paper award! 🏆 A thread 🧵👇

Demo: CAP (Contrastive Activation Patching) is an open-source toolkit for finding, comparing and steering the neurons behind behaviours like hallucination, across languages. It’s joint work with @UniBonn, Fraunhofer IAIS & Lamarr Institute (Semenikhin, Frey, Dušek, Flek, Schmidtová).

Our posters covered: 📊 A retrospective on NLG evaluation (Schmidtová, Mahamood, Dušek) 💬 Why reasoning gets harder for LLMs inside a dialogue (Kartáč, Lango, Dušek) 🐾 VetScore: risk-weighted fact-checking for veterinary QA (Kartáč, Tovarys, Lango, Dušek) 🌍 Modular multilingual LMs (Kumar, Dušek)

“Can LLM Coding Agents Reason About Time Series?” (Filip Rechtorík, @tuetschek.bsky.social , @zdenekkasner.cz ) compares LLMs reading raw numbers, LLMs writing Python code to analyse the data, and a hybrid of both. 2026.inlgmeeting.org/static/proce...

2026.inlgmeeting.org2026.inlgmeeting.org

“Strategies for Span Labeling with Large Language Models” (Danil Semin, @tuetschek.bsky.social , @zdenekkasner.cz ) looks at how LLMs can mark parts of an input text, e.g. named entities or errors. Three families of approaches: tagging, indexing and matching. 2026.inlgmeeting.org/static/proce...

🏆 Best Short Paper Award: “Do LLMs Make More Mistakes If They Do Not Believe the Input Data?” by Peter Kochelka, Aleš Manuel Papáček, Vojtěch Dvořák & Ondřej Dušek. Congratulations! 🎉 2026.inlgmeeting.org/static/proce...

The paper asks whether LLMs describe data less faithfully when the data contradicts what they “know”. The team built factual, counterfactual and fictional data from local Czech and Slovak knowledge, and had 9 open models write text in English, Czech, Slovak and Upper Sorbian.

#EMNLP2026 is about to start, and folks from @ufal.mff.cuni.cz will be there presenting their work. Here is where you can find us at the main conference: @jlibovicky.bsky.social presents on Monday afternoon; Adnan on Tuesday morning 👇

Institute of Formal and Applied Linguistics reposted

Jindřich Libovický

@jlibovicky.bsky.social

I’m heading to #EMNLP2026 in Budapest 🇭🇺! Here’s where you can find us 🧵👇

✨ ÚFAL at #DHECC2026 in Denmark! 🏆 Huge congrats to Christopher Brückner & team (N. S. Andersen, A. Revenko, U. Yorsh, K. Tauchmanová, A. Turčan, P. Pecina) for winning the Best Paper Award for their NLP work on Holocaust testimonies in the MEMORISE project! 🧵👇 #DigitalHumanities #NLP

🇨🇿 Big news to wrap up: DHECC 2027 will be hosted by ÚFAL right here in Prague in September 2027! We look forward to hosting the digital heritage community next year! 🏛️✨ 📸 Photo: Martin Pedersen #DHECC2027 #UFAL #Prague

🎓 Filip Makara presented research from his BSc thesis on Dense Passage Retrieval & custom UI for the Visual History Archive. The award-winning paper "Semantic Enrichment of Historical Archives" will appear in the International Journal of Digital Humanities! 📖 #OralHistory #AIinResearch

From data highways to oral history: ÚFAL researchers presented at the CLARIN Conference in Barcelona. Ondřej Košarko & Pavel Straňák compared EOSC-CZ & CLARIN infrastructure architectures, while Christopher Brückner showcased MalachNER, turning oral history into searchable data. #NLP #CLARIN2026

Institute of Formal and Applied Linguistics reposted

Jindřich Libovický

@jlibovicky.bsky.social

21 Sept

I really enjoyed giving a talk at the Machine Translation Marathon 2026 about our research on tokenization evaluation. Thank you for having me! For those interested, the slides from my presentation are available here: docs.google.com/presentation...

arxiv.orgPredicting Multilingual Classification and Translation Performance of LLMs with Cross-Lingual Alignment -- Is English Enough?Multilingual large language models (LLMs) have been shown to perform better on non-English classification tasks when the representations of the given language are more aligned to English within the mo...

2 Sept

Congrats, Adnan! 🎉 If you're at #EMNLP2026 in Budapest 🇭🇺, don't miss this one.

Jindřich Libovický@jlibovicky.bsky.social

Very happy to share that Adnan Al Ali's first big PhD project has been accepted to #EMNLP2026! 🎉 Joint work with @kathaem.bsky.social, me & Alex Fraser. Paper: arxiv.org/abs/2608.03446 What predicts multilingual LLM performance in a language? 🧵

Institute of Formal and Applied Linguistics reposted

Jana Straková

@janastrakova.bsky.social

13 July

We're releasing a new multilingual model for NameTag 3: nametag3-multilingual-260521. It achieves state-of-the-art performance on 33 test datasets across 23 languages. Try the demo: lindat.mff.cuni.cz/services/nam... Documentation: ufal.mff.cuni.cz/nametag/3

lindat.mff.cuni.czlindat.mff.cuni.cz

#ACL2026 continues with the main conference (main + findings + demos + industry + SRW), and @ufal.mff.cuni.cz folks will present 8️⃣ papers. Stop by and check with our colleagues. All times in PDT.

9 July

Nalin Kumar & @tuetschek.bsky.social presented Modular Monolingual Adaptation using Pretrained Language Models aclanthology.org/2026.acl-ind... with their tricks for saving parameters and improving performance when fine-tuning pretrained LMs for low-resource languages.

8 July

🎉 Congratulations to our colleague Dan Zeman, who together with Marie-Catherine de Marneffe, @chrmanning.bsky.social & Joakim Nivre has won the ACL 2026 Computational Linguistics High Impact Paper Award for "Universal Dependencies"! 👏 aclanthology.org/2021.cl-2.11

universaldependencies.orgUniversal Dependencies

8 July

What are Universal Dependencies? 🌍 UD is a framework for annotating grammar consistently across languages, now covering 100+ languages. It's used across multilingual NLP: parsing, linguistic typology studies, and model interpretability studies. Explore: universaldependencies.org

8 July

Reasoning Gets Harder for LLMs Inside A Dialogue aclanthology.org/2026.acl-lon... by @ivankartac.bsky.social, M. Lango & @tuetschek.bsky.social Same reasoning tasks, isolated vs. inside a dialogue: 9 LLMs 📉 consistently do worse once conversation, roles & tool-use requirements enter the picture.

Semantic-pragmatic Annotations in the Prague Dependency Treebank aclanthology.org/2026.finding... @mariemikulova.bsky.social , E. Hajicova, J. Mirovsky, A. Nedoluzhko, M. Novak, P. Synkova, J. Stepanek, B. Stepankova, @hajicjan.bsky.social A 3M+ token corpus, fully annotated & freely available.

Thesis Proposal by Kristýna Onderková: Intentional Inference for Insight Generation aclanthology.org/2026.acl-srw... Why do LLM insights feel shallow? Often, it is unstated assumptions that fill gaps in the task. This thesis makes models surface them & push toward deeper, more trustworthy reasoning.

Thesis proposal by Honza Bronec (supervised by @jindrahelcl.bsky.social) 👉 Targeted and Unified Cross-Lingual Unlearning from Multilingual Language Models aclanthology.org/2026.acl-srw... Aims to unify cross-lingual unlearning benchmarks and edit only the layers that store the knowledge.

LLM as a Meta-Judge: Synthetic Data for NLP Evaluation Metric Validation aclanthology.org/2026.acl-srw... by Lukáš Eigler, @jlibovicky.bsky.social & David Hurych Rankings from synthetic LLM-generated data almost perfectly match real human judgments. 🤖⚖️ at Student Research Workshop

One-step Nonautoregressive Natural Language Generation with Shortcut Flow Matching Models by Jędrzej Warczyński, Mateusz Lango & @tuetschek.bsky.social aclanthology.org/2026.acl-sho... One-step generation that actually works: BLEU more than doubles vs classic flow matching.

6 July

Dušan Variš presents joint work w/ @abyste.bsky.social & @jlibovicky.bsky.social: TokCollate: A Comprehensive Tool for Tokenizer Evaluation and Visualization across Languages aclanthology.org/2026.acl-dem... A dashboard for your tokenizer 👀 See which languages you left behind. Open-source, MIT.

#ACL2026 starts with two days of tutorials and workshops. @ufal.mff.cuni.cz folks will be there and present their work, here's our schedule, so don't miss our presentations 👇

UFAL-CUNI at SemEval-2026 Task 11: An Efficient Modular Neuro-symbolic Method for Syllogistic Reasoning by @ivankartac.bsky.social, Honza Bronec, Kristýna Onderková, @zdenekkasner.cz, Mateusz Lango & @tuetschek.bsky.social aclanthology.org/2026.semeval... 💪 4B model + prover beats zero-shot LLMs

#ACL2026 starts with two days of tutorials and workshops. @ufal.mff.cuni.cz folks will be there and present their work, here's our schedule, so don't miss our presentations 👇

Lucie Poláková presents: Challenges in Machine Translation of Interactive Multimodal Exercises aclanthology.org/2026.bea-1.1... MT isn't as solved as it seems: terminological, multimodal, XML-encoded exercises bring new challenges, incl. LLMs 'fixing' counterfactuals.

#ACL2026 starts with two days of tutorials and workshops. @ufal.mff.cuni.cz folks will be there and present their work, here's our schedule, so don't miss our presentations 👇

4 July

A Pocket Offline Model for Simultaneous Speech Translation as CUNI Submission to IWSLT2026 aclanthology.org/2026.iwslt-1.22 AlignAtt4LLM: Fast AlignAtt for Decoder-Only LLMs @ IWSLT2026 Simultaneous Speech Translation Task aclanthology.org/2026.iwslt-1.32 presented simultaneously by Dominik Macháček

#ACL2026 starts with two days of tutorials and workshops. @ufal.mff.cuni.cz folks will be there and present their work, here's our schedule, so don't miss our presentations 👇

4 July

@straka-milan.bsky.social winning the Unconstrained track in this shared task with his system CorPipe aclanthology.org/2026.codi-1.27 CorPipe 26 tops the unconstrained track by 9.5 points and the LLM track by 2.8, with ablations on model size, empty node prediction & zero-shot transfer.

#ACL2026 starts with two days of tutorials and workshops. @ufal.mff.cuni.cz folks will be there and present their work, here's our schedule, so don't miss our presentations 👇

4 July

Michal Novák presented "Findings of the Fifth Shared Task on Multilingual Coreference Resolution: Expanding Datasets for Long-Range Entities" on the CODI-CRAC workshop, the CRAC26 shared task overview. CorefUD 1.4: 27 datasets, 19 languages. #NLProc aclanthology.org/2026.codi-1....

2 July

Congrats! Stop by to chat with @ivankartac.bsky.social at #ACL2026 in San Diego 🇺🇸 He's presenting 2 papers: A modular neuro-symbolic system pairing small LLMs with a theorem prover for syllogistic reasoning, and BOULDER, a new benchmark showing LLM reasoning drops sharply inside multi-turn dialogue.

Ivan Kartáč@ivankartac.bsky.social

Heading to San Diego for #ACL2026, where I’ll be presenting two papers (see 🧵). Stop by to chat about evaluating reasoning embedded in task-oriented dialogue, or how to use small LLMs in modular neuro-symbolic approaches to syllogistic reasoning!

aclanthology.orgaclanthology.org

If you are at #ACL2026, don't miss @mariemikulova.bsky.social et al. presenting a semantic-pragmatic annotation layer for the 3M+ token Prague Dependency Treebank aclanthology.org/2026.finding...

Marie Mikulová@mariemikulova.bsky.social

Semantic-pragmatic annotation in the Prague Dependency Treebank — three times at #ACL2026! 📍 Jul 3, 13:30–14:15 LAW La Jolla, talk 📍 Jul 3, 14:20–15:30 CODI-CRAC Regatta, poster 📍 Jul 6, 11–12:30 Findings Poster, Grand Hall See you in San Diego! @ufal.mff.cuni.cz aclanthology.org/2026.finding...

Institute of Formal and Applied Linguistics reposted

Open Euro LLM

@openeurollm.bsky.social

15 June

Input, more input 🤖⚡ Just like Jonny 5 in Short Circuit, our baby model is reading every single token from its pretraining dataset. So far: 10 trillion tokens, 36 languages + code & math as their own "languages" 📚🌍💻 We’re tracking progress & sharing it openly 👇 (1/2)

Institute of Formal and Applied Linguistics reposted

Tokenization Workshop (TokShop) @COLM2026

@tokshop.bsky.social

23 June

📣 TokShop 2026 deadline extended! 🗓️ New submission deadline: Friday, June 26, 2026 (AoE) Research papers (up to 9 pages) and extended abstracts (up to 2 pages) are welcome. Submit: openreview.net/group?id=col... More info: tokenization-workshop.github.io

Institute of Formal and Applied Linguistics reposted

Jindřich Libovický

@jlibovicky.bsky.social

9 June

Lukáš Eigler defended his thesis (co-supervised with David Hurych, @valeoai.bsky.social) 🎉 Congrats! #NLP metric validation needs 🐌💰 human judgment data. Our fix: generate synthetic data for metric validation. ✅ Tested on MT, QA, summarization. To appear #ACL2026 SRW: arxiv.org/abs/2603.09403

arxiv.orgLLM as a Meta-Judge: Synthetic Data for NLP Evaluation Metric ValidationValidating evaluation metrics for NLG typically relies on expensive and time-consuming human annotations, which predominantly exist only for English datasets. We propose LLM as a Meta-Judge, a scalabl...

arxiv.orgCHALIS: A Challenge Dataset for Language Identification in Difficult ScenariosWe present CHALIS (Challenging Language Identification Samples), a new benchmark dataset explicitly designed to address difficult cases in language identification: cousin languages and orthographic no...

23 June

One of a dozen master's theses defended at our institute this season. Thesis season is always a reminder of the talented students passing through here. 🎓

Jindřich Libovický@jlibovicky.bsky.social

Is "Je to obrovský problém, tvrdí české ministerstvo obrany." Czech or Slovak? Both. Language ID systems must pick one: and that's just one way they break. Our new benchmark CHALIS exposes how fragile they really are. 📄 arxiv.org/abs/2606.06088 🤗 huggingface.co/datasets/michal-tichy/CHALIS

Institute of Formal and Applied Linguistics reposted

Tokenization Workshop (TokShop) @COLM2026

@tokshop.bsky.social

22 May

Announcing First Call for Papers: Second Tokenization Workshop 🔡 📣 ▶️ Non-archival submissions of two types: Research papers (up to 9 pages) ▶️ Extended abstracts (up to 2 pages) Submission deadline June 23, 2026 (AoE) Acceptance notification on July 24, 2026 (AoE) tokenization-workshop.github.io

Institute of Formal and Applied Linguistics reposted

Marie Mikulová

@mariemikulova.bsky.social

Great discussions, inspiring talks, and lots of interest around the 🌲Prague Dependency Treebank🌲 at #LREC2026. Hope we helped give PDT some well-deserved visibility there! @ufal.mff.cuni.cz ➡️ lrec.elra.info/lrec2026-mai... ➡️ lrec.elra.info/lrec2026-mai... ➡️ lrec.elra.info/lrec2026-mai...

#LREC2026 is over. It had huge @ufal.mff.cuni.cz presence 👯🙆‍♂️🧑‍🤝‍🧑🙋🧓🧒 of 2️⃣6️⃣ people and 1️⃣4️⃣ main conference papers. If you missed our presentations, you can still checkout our #NLProc and #CL papers 👇

SEEM-CZ: Annotation and Classification of Epistemic Markers in Czech by Bára Štpánková, Michal Novák, Tomáš Musil, Lucie Poláková lrec.elra.info/lrec2026-mai... Czech epistemic markers dataset (~4,000 uses) with annotations and XLM-RoBERTa classifiers for NLP.

Automatic Suggestions Help Extending Eventive Ontology: A Case Study on SynSemClass by @janastrakova.bsky.social, Eva Fučíková, Zdenka Urešová & @hajicjan.bsky.social lrec.elra.info/lrec2026-mai... Auto-suggestions improve semantic ontology annotation agreement in SynSemClass case study.

CzechDocs: A Multiway Parallel Dataset of Formatted Documents for Minority Languages in Czechia 🇨🇿 by Pepa Jon and Ondřej Bojar 📝 lrec.elra.info/lrec2026-mai... 📂 github.com/cepin19/Czec...