Empowering individuals and organisations to safely use foundational AI models.
Tübingen, Germany
Joined March 2023
- Tweets140
- Following116
- Followers281
- Likes134
‼️ New paper from Parameter Lab!
⛓️💥 We identify privacy collapse, a silent failure mode of LLMs: LLMs fine-tuned on seemingly benign data can lose their ability to respect contextual privacy norms.
Done by @anmgoel during his internship!
Check-out 👇
ALT Privacy collapse
Parameter Lab retweeted
🚨 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.
⬇️
Parameter Lab retweeted
For more insights:
🌐 Project page: parameterlab.github.io/priva…
📄 Paper: arxiv.org/abs/2601.15220
Work done with the amazing team at @parameterlab: @CorEmde, @oodgnas, @coallaoh and @framart1 with support from @NAVER_AI_Lab.
#NLProc #AISafety #Privacy #LLMs
👏 Proud to share that the paper that Ahmed Heakl authored during his internship at Parameter Lab was accepted at #ICLR2026!
See how 🩺Dr.LLM increases accuracy and decreases inference computations of frozen LLMs: lnkd.in/dqzByRkT
Parameter Lab retweeted
🎉Delighted to announce that our 🫗Leaky Thoughts paper about contextual privacy with reasoning models is accepted to #EMNLP main!
Huge congrats to the amazing team @tommasogreen @HaritzPuerto @coallaoh @oodgnas
Delighted by this great thread from @omarsar0 presenting our new Leaky Thoughts paper!
We show that reasoning models pose serious privacy risks when used as personal agents. Reasoning traces are a new attack vector.
Work led by @tommasogreen during his internship @parameterlab!
ALT Teaser figure. Prior studies on contextual privacy focused on LLM output. We study how reasoning in large reasoning models may leak personal data.
🧪 Our latest research: Does SEO boost the visibility of content in LLM-based conversational search?
We present C-SEO Bench, a benchmark to evaluate conversational SEO strategies.
Key takeaway: SEO methods that target LLM do not work. But surprisingly, traditional SEO is not dead: it still matters, as LLMs tend to favour content already ranked higher in their input.
ALT Paper thumbnail
🔎 Does Conversational SEO (C-SEO) actually work? Our new benchmark has an answer.
Excited to announce C-SEO Bench: Does Conversational SEO Work?
🌐 RTAI: researchtrend.ai/papers/2506…
📄 Paper: arxiv.org/abs/2506.11097
💻 Code: github.com/parameterlab/c-se…
📊 Data: huggingface.co/datasets/para…
Authored by @tommasogreen @framart1 @HaritzPuerto @oodgnas @coallaoh
Huge thanks to this amazing team for the great work!
🎉 Very excited to see our new Leaky Thoughts 🫗 paper featured among last week's top AI papers by both @dair_ai and @TheAITimeline!
- teapawt.org/dair_ai/status/1936825…
- teapawt.org/TheAITimeline/status/1…
➡️ Learn more about the paper in this great thread by @omarsar0: teapawt.org/omarsar0/status/193571…
➡️ ArXiv link: arxiv.org/abs/2506.15674
ALT Main figure.
Parameter Lab retweeted
We see news like this from time to time and that’s why it’s vital to keep researching on tools to prove these cases! Our #NAACL2025 paper shows that with over 10k tokens, we can reliably detect whether a text was part of an LLM’s training data aclanthology.org/2025.findin…
Disney, Universal sue image creator Midjourney for copyright infringement reut.rs/3SMqtlu reut.rs/3SMqtlu
Parameter Lab retweeted
Proud to have co-led more than 200 of those 999 during my time at Google.
Check out the 999 open models that Google has released on @huggingface:
huggingface.co/google
(Comparative numbers: 387 for Microsoft, 33 for OpenAI, 0 for Anthropic).
Parameter Lab retweeted
Do you want to prove that your copyrighted document/corpus was trained by an LLM? Come to poster 46 #NAACL2025
Parameter Lab retweeted
#NAACL2025 has started! I’ll be presenting my work at @parameterlab about detecting pretraining data on Friday
🗓️ May 2, 11:00 AM - May 2, 12:30 PM
🗺️ Poster Session 8 - APP: NLP Applications
Location: Hall 3
Work with @framart1 @oodgnas @coallaoh
🧵 It is assumed that Membership Inference Attacks (MIA) do not work on LLMs, but our new paper shows it can work at the right scale! MIA is effective if the number of input tokens is large enough, such as in long documents and collections of them. 📃arxiv.org/abs/2411.00154
Parameter Lab retweeted
I will be in person at #NAACL2025 🌵🇺🇸 to present Scaling Up Membership Inference: When and How Attacks Succeed on LLMs. Come and say hi 👋 if you want to know how to proof if an LLM was trained on a data point!
🧵 It is assumed that Membership Inference Attacks (MIA) do not work on LLMs, but our new paper shows it can work at the right scale! MIA is effective if the number of input tokens is large enough, such as in long documents and collections of them. 📃arxiv.org/abs/2411.00154
🔎 Wonder how to prove an LLM was trained on a specific text? The camera ready of our Findings of #NAACL 2025 paper is available!
📌 TLDR: longs texts are needed to gather enough evidence to determine whether specific data points were included in training of LLMs: arxiv.org/abs/2411.00154
🧵 It is assumed that Membership Inference Attacks (MIA) do not work on LLMs, but our new paper shows it can work at the right scale! MIA is effective if the number of input tokens is large enough, such as in long documents and collections of them. 📃arxiv.org/abs/2411.00154
Parameter Lab retweeted
We just wanted to say: Membership inference is unlikely to succeed on n-grams or even paragraphs. Language models require **multiple documents** to gather enough evidence to determine whether specific data points were included in training. Accepted to #NAACL2025 Findings.
I'm excited to announce that my internship paper at @parameterlab was accepted to Findings of #NAACL2025 🎉 Huge thanks to @framart1 @coallaoh and @oodgnas! Amazing team!!
Parameter Lab retweeted
I'm excited to announce that my internship paper at @parameterlab was accepted to Findings of #NAACL2025 🎉 Huge thanks to @framart1 @coallaoh and @oodgnas! Amazing team!!
🧵 It is assumed that Membership Inference Attacks (MIA) do not work on LLMs, but our new paper shows it can work at the right scale! MIA is effective if the number of input tokens is large enough, such as in long documents and collections of them. 📃arxiv.org/abs/2411.00154
Parameter Lab retweeted
techcrunch.com/2025/01/09/ma… From time to time we hear news like this. However, proving that an LLM was trained on a specific document is very challenging 🥴 This motivated my latest work, where we show that current methods can be effective if we use enough data 🧐
Parameter Lab retweeted
🔥 Curious about the latest trends in AI?
ResearchTrend.AI gives you an intuitive overview of diverse topics in AI. Dive into our dashboard featuring 55 communities ready for you to explore!
👉 Visit: researchtrend.ai/communities
Browse and follow your favourite communities. You’ll receive a daily digest directly on your landing page after logging in.
Got an idea for a new community? Let us know, and we’ll add it for you!
👉 Suggest a community: researchtrend.ai/communities…
🚀 Start exploring today!
#ResearchTrendAI
🚨📄 Exciting new research! Discover when and at what scale we can detect if specific data was used in training LLMs — a method known as Membership Inference (MIA)! Our findings open new doors for using MIA as potential legal evidence in AI. More info in the 🧵below.
🧵 It is assumed that Membership Inference Attacks (MIA) do not work on LLMs, but our new paper shows it can work at the right scale! MIA is effective if the number of input tokens is large enough, such as in long documents and collections of them. 📃arxiv.org/abs/2411.00154