🚨 New paper alert 🚨 Using LLMs as data annotators, you can produce any scientific result you want. We call this **LLM Hacking**. Paper: https://t.co/24Fyb4Ik3v

The debate over “LLMs as annotators” feels familiar: excitement, backlash, and anxiety about bad science. My take in a new blogpost is that LLMs don’t break measurement; they expose how fragile it already was. https://t.co/6CweDPv5wG

Cool paper by @ey_985, confirming our LLM hacking findings ( https://t.co/24Fyb4IRT3): ✓ LLMs are brittle data annotators ✓ Downstream conclusions often flip: *LLM hacking risk* is real! ✓ Bias correction methods can help but have tradeoffs ✓ Use human expert whenever possible

We’re delighted to welcome @enfleisig to our @MilaNLProc lab as a visiting PhD student! ✨

📣New paper: Rigorous AI agent evaluation is much harder than it seems. For the last year, we have been working on infrastructure for fair agent evaluations on challenging benchmarks. Today, we release a paper that condenses our insights from 20,000+ agent rollouts on 9

[CL] Large Language Model Hacking: Quantifying the Hidden Risks of Using LLMs for Text Annotation J Baumann, P Röttger, A Urman, A Wendsjö... [Bocconi University & University of Zurich] (2025) https://t.co/ynve6AW3DH

Thank you to the amazing @paul_rottger @AUrman21 @awendsjo @florplaza22 @ Johannes B. Gruber @ Dirk Hovy for this fun collaboration!

Why this matters: LLM hacking affects any field using AI for data analysis–not just computational social science! Please check out our preprint; we'd be happy to receive your feedback! #LLMHacking #Research #Reproducibility #DataAnnotation #nlp #Statistics #LLM

The good news: we present several solutions: ✅ Larger, more capable models are safer (but no guarantee) ✅ Few human annotations beat many AI annotations ✅ Testing several models and configurations on held-out data helps ✅ Pre-registering AI choices can prevent cherry-picking

This also concerns well-intentioned researchers! - Researchers using SOTA LLMs like GPT-4o face a 31-50% chance of false conclusions for plausible hypotheses - Risk peaks near significance thresholds (p=0.05) - Regression correction methods trade off Type I vs. Type II errors

We tested 18 LLMs on 37 social science annotation tasks (13M labels, 1.4M regressions). By trying different models and prompts, you can make 94% of null results appear statistically significant–or flip findings completely 68% of the time.

@MilaNLProc @AUrman21 @RERobertson @ancsaaa3 @tiancheng_hu The @MilaNLProc group is presenting 13 more paper at this year's #ACL2025 , go check them out :) https://t.co/t65ATxE9Rr

@MilaNLProc @AUrman21 @RERobertson @ancsaaa3 Shoutout to @tiancheng_hu for yesterday's stellar presentation of our work benchmarking LLMs' ability to simulate group-level human behavior:

@MilaNLProc I'm at #ACL2025 this week:📍Find me at the FEVER workshop, *Thursday 11am* 📝 presenting: "I Just Can't RAG Enough" - our ongoing work with @AUrman21 @RERobertson @ancsaaa3, showing that RAG does not solve LLM fact-checking limitations!

Breaking my social media silence because this news is too good not to share! 🎉 Just joined @MilaNLProc as a Postdoc, working on large language models and computational social science!

🎉 The @MilaNLProc lab is excited to present 15 papers and 1 tutorial at #ACL2025 & workshops! Grateful to all our amazing collaborators, see everyone in Vienna! 🚀