Martin Gubri

Research Lead @parameterlab working on Trustworthy AI | he/him

Other accounts: 🦋 mgubri | 🐘 @[email protected]

[Martin Gubri @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 🧵
⬇️⬇️⬇️

[Martin Gubri @framart1]

@DamienTeney @a_rubique @coallaoh Thanks, Damien ☺️

[Martin Gubri @framart1]

🏆🪩 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!

[Alexander Rubinstein @a_rubique]

🎉 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!

[Martin Gubri @framart1]

@HaritzPuerto @a_rubique @coallaoh Thanks a lot!!

[Martin Gubri @framart1]

🪩 If you are at #ICLR today, come the poster of our DISCO paper to discuss LLM efficient benchmarking with @a_rubique!

[Alexander Rubinstein @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

[Martin Gubri @framart1]

6/ Thank you for the hard work, the patience, and the brilliance! I learned so much from all of you.

More to come soon :)

[Martin Gubri @framart1]

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!)

[Martin Gubri @framart1]

5/ And to all my brilliant co-authors since 2023:
@dnnslmr, @oodgnas, @hwaran_lee, Siwon, @HaritzPuerto, @tommasogreen, @anmgoel,
@CorEmde, Ahmed, @esruzzetti, Salman, @amohamed264, @a_rubique, @adamdaviesnlp,
@_elinguyen, @michael_simeone, Erik.

[Martin Gubri @framart1]

@CFGeek There is work on that: you can check model fingerprinting and model watermarking (black & white box methods). I recommend this survey & benchmark paper: https://t.co/xmxcG6zngs
Which includes our TRAP paper: https://t.co/IO1rIGrBX5
Nonetheless, I wish we have such model lineage

[Martin Gubri @framart1]

🎉Our privacy collapse paper has been accepted at #ACL 2026 (main)!

Contextual privacy is fragile: fine-tune an LLM on benign data, and it can overshare personal information.

This is silent: suites don't usually measure contextual privacy, which is a big issue with LLM agents.

[Anmol Goel @anmgoel]

🚨 Fine-tuning your model to be more helpful or empathetic might be making it less private, without you noticing.

In our latest work, we show that benign fine-tuning can silently break contextual privacy in language models while safety & general capabilities appear intact.

⬇️

[Martin Gubri @framart1]

@tom_doerr Our Neurips paper proposes a benchmark for these GEO methods and shows that none (that we tested) work: https://t.co/8n18q3NuHZ

[Martin Gubri @framart1]

Would love to hear thoughts from people working on watermarking :)
cc @ni_jovanovic @sahar_abdelnabi @jonasgeiping

[Martin Gubri @framart1]

🌍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.
🧵

[Martin Gubri @framart1]

Kudos to Asim Mohamed for his first research paper!

Paper: https://t.co/slv2Kxo3YC
Code: https://t.co/V3Qo0UeaXB