Martin Gubri @EMNLP (@framart1) header
Martin Gubri @EMNLP (@framart1) profile picture

Assistant professor at École Polytechnique working on Trustworthy AI | he/him Other account: 🦋 mgubri

Palaiseau, France
Joined December 2012
Photo by Martin Gubri @EMNLP (@framart1)Photo by Martin Gubri @EMNLP (@framart1)Photo by Martin Gubri @EMNLP (@framart1)Photo by Martin Gubri @EMNLP (@framart1)Photo by Martin Gubri @EMNLP (@framart1)Photo by Martin Gubri @EMNLP (@framart1)Photo by Martin Gubri @EMNLP (@framart1)Photo by Martin Gubri @EMNLP (@framart1)Photo by Martin Gubri @EMNLP (@framart1)Photo by Martin Gubri @EMNLP (@framart1)
🦹💥 How to detect if my LLM was stolen or leaked? 🤖💥 I am delighted to announce TRAP 🪤, our new #ACL2024 findings paper ☝️ We showcase how to use adversarial prompt as model fingerprint for LLM. A thread 🧵 ⬇️⬇️⬇️
1
4
3
25
4,323
Prefilling the reasoning to detect a way to train on test: simple and clever!
Replying to @JSchaeff3r
In 23% of its attempts, K3.1 cites "instructions" that only the hidden tests check. Its predecessor K3 never does this. Prefilling its reasoning with "let me list all the instructions I was given": K3.1 names hidden requirements in 65% of answers, K3 in 18%.
5
150
Martin Gubri @EMNLP retweeted
We recently released our paper on evaluating compaction in long-horizon benchmarks (read @hrdkbhatnagar's thread below!) Counterintuitively (to me), we found that the compaction method and threshold mostly don’t affect performance; no method beats self-summarization. However, compaction removes many of the initial rules! A simple solution could be to always keep the initial rules in the context window. But what about constraints discovered later, or legitimate changes to these rules? One takeaway for me from this project is that we need other agentic benchmarks for evaluating compaction, where agents learn rules or constraints by interacting with their environment, or where those rules can be overridden or conflict. This is especially important since compaction may drop much of the agent’s nuanced reasoning, including its uncertainty about whether it is violating a rule. Joint work with: @jsalle_ , @hrdkbhatnagar , @MatthiasBethge , and @maksym_andr! Read the paper for me details: arxiv.org/abs/2610.31778
New paper: Benchmarks Are Surprisingly Compaction Friendly Agents commonly go past their context windows, and then "compact" to keep going. Pretty much all frontier agents do this, but no one has evaluated compaction on actual long horizon benchmarks We tested 5 such compaction strategies
4
9
50
1,938
On my way back from #ENMLP2026 by sleeper train. Best carbon footprint and amazing view to wake up :) I had a lovely time. Met plenty of interesting researchers and many interesting conversations. Some constructive feedbacks: 1/
1
18
1,580
- it would be great if the conf organizers could partner with tools like scholar-inbox, so the attendees can have a recommendation of posters grouped by session
1
6
225
- the poster room was too small. More space between posters would be better to properly navigate. The room of the social even would have been perfect Great conference overall. Looking forward to the future @emnlpmeeting !
2
4
222
Interested in multilingualism and watermarking? Come now at poster 194! #EMNLP2026
🌍We've made LLM watermarking equally robust across all languages we studied. Scaling to 100+ languages! Even sota watermarks can be removed by translating to another language, eg. Tamil. This hits hardest in low-resource languages, where moderation tools are already weak. 🧵
3
20
691
Martin Gubri @EMNLP retweeted
🎉 Proud to share our work on multilingual LLM watermarking got accepted to Findings of EMNLP 2026! More details in the post below. Preprint: arxiv.org/abs/2510.18019 (to be updated with the camera-ready version) #EMNLP2026 #LLMWatermarking #TextWatermarking #TrustworthyAI
🌍We've made LLM watermarking equally robust across all languages we studied. Scaling to 100+ languages! Even sota watermarks can be removed by translating to another language, eg. Tamil. This hits hardest in low-resource languages, where moderation tools are already weak. 🧵
1
1
117
🎉 I am extremely excited to announce the biggest news of my career: I will join École Polytechnique as an assistant professor in September!
3
20
887
I am very grateful for the trust placed in me by Sonia Vanier and all the academics, and very excited to work with the current and future members of the ORAILIX lab.
1
1
53
A huge thanks to all my co-authors for their hard work, and to my co-workers, friends and family for their support! I look forward to the years ahead!!
3
53
Martin Gubri @EMNLP retweeted
I'll be at #ACL2026 in San Diego, presenting our recent works spanning AI safety! 📍 Poster (Jul 6) Privacy Collapse: Benign Fine-Tuning Can Break Contextual Privacy in Language Models arxiv.org/abs/2601.15220 (1/2)
4
2
1
20
1,521
Martin Gubri @EMNLP retweeted
Today at #ACL2026, we are presenting out MASEval library for multi-agent system evaluation. @anmgoel is in San Diego to present poster and live demo! 📍Grand Hall | Session 3: Oral/Posters/Demos B 🕑Sunday 2pm-3.30pm #MultiAgentSystem #AI #AIAgents #ACL #Evaluation
1/ Evaluating a single agent harness is hard. Evaluating a multi-agent system? That's a whole different problem. Most eval tools treat the model as the unit of analysis. But in multi-agent systems, the system is what matters. That's why we built MASEval 🧵 #Agents #AI #Eval
1
1
9
740
🏆🪩 DISCO just won the best paper award at ICLR'26-CAO! I am so proud of my co-authors! Huge congrats to @a_rubique, Benjamin, and @coallaoh!
🎉 Our DISCO paper received a Best Paper Award at the ICLR 2026 “Catch, Adapt, and Operate” Workshop! Huge thanks to my coauthors Benjamin, @framart1, and @coallaoh, and to the CAO organizers, led by @sepidshs, and @RBCBorealis, for supporting the workshop!
2
10
380
🪩 If you are at #ICLR today, come the poster of our DISCO paper to discuss LLM efficient benchmarking with @a_rubique!
Come check out our poster to learn a simple (but very effective) trick to speed up your LLM evaluation by 100×! ⏰ 15:15 to 17:45 today! 📍Pavilion 3 P3-#1015
8
207
1/ My contract at @parameterlab ended last week, after 2.5 years (since Sept 2023, with some collaboration before). I had the chance to lead research on trustworthy AI for LLMs alongside an incredible group of people (Neckarfront. All Tübingen researchers have to post it once!)
2
15
1,165
6/ Thank you for the hard work, the patience, and the brilliance! I learned so much from all of you. More to come soon :)
5
88