We’ve all wondered “is this text already in my LLM training data? 🤔” Want a fast and compact way to check data membership? Introducing Data Portraits 🖼️ Work w/ Benjamin Van Durme Demo: dataportraits.org

I’m hearing chatter of PhD students not knowing what to work on. My take: as LLMs are deployed IRL, the importance of studying how to use them will increase. Some good directions IMO (no training): 1. prompting 2. evals 3. LM interfaces 4. safety 5. understanding LMs 6. emergence

Happy to share the news of this Best Paper Award for the COMPS paper (arxiv.org/abs/2210.01963). Congratulations Kanishka!

Hong, Aswathy, Greg, Eamon, Kyle and Ian's paper "The Diminishing Returns of Masked Language Models to Science”, has been accepted to the Findings of ACL 2023 (doi.org/10.48550/arXiv…)!📜

How does natural language shape bilateral bargaining? To study this, we contribute a dataset of spoken natural language (NL) negotiations & a control without NL. We find that NL promotes cooperation. Being patient and reactive is an effective strategy. 📄arxiv.org/abs/2306.07117

Our paper on analyzing language modeling acceptability judgments with systematically manipulated contexts was recognized as an outstanding paper—thanks so much to the reviewers and the best paper award committee!

In the past weeks I received many questions (from undergrads especially) about AI research, so I'm putting together a "Ask Me Anything" doc. Add any questions to the doc, I'll answer all of them: docs.google.com/document/d/1QR… Yes, I'll actually answer them all, because writing answers

Excited to be attending #EMNLP2023 in 🇸🇬! Will be giving an oral presenting my paper about localizing and rectifying multi-hop reasoning failures in LMs at BlackboxNLP on Dec 7th at (2:00-3:30 pm). Please stop by and say 👋/DM if you want to connect! msakarvadia.github.io/memory_injecti…

🤫Jailbreak prompts make aligned LMs produce harmful responses.🤔But is that info linearly decodable? ↗️We show many refused concepts are linearly represented, sometimes persist through instruction-tuning, and may also shape downstream behavior❗ arxiv.org/abs/2507.00239 🧵1/

Prompting is our most successful tool for exploring LLMs, but the term evokes eye-rolls and grimaces from scientists. Why? Because prompting as scientific inquiry has become conflated with prompt engineering. This is holding us back. 🧵and new paper: arxiv.org/abs/2507.00163

HR Simulator™: a game where you gaslight, deflect, and “let’s circle back” your way to victory. Every email a boss fight, every “per my last message” a critical hit… or maybe you just overplayed your hand 🫠 Can you earn Enlightened Bureaucrat status? (link below)

For those who missed it, we just releaaed a little LLM-backed game called HR Simulator™ You play an intern ghostwriting emails for your boss. It’s like you’re stuck in corporate email hell…and you’re the devil 😈 link and an initial answer to “WHY WOULD YOU DO THIS?” below

An interesting collaboration and mentoring experience -- this time with AIs🤖. Check out the paper written by our AI scientists: openreview.net/forum?id=sVaRg…

❓ Does an LLM know thyself? 🪞 Humans pass the mirror test at ~18 months 👶 But what about LLMs? Can they recognize their own writing — or even admit authorship at all? In our new paper, we put 10 state-of-the-art models to the test. Read on 👇 1/n 🧵

Redesigning an LLM course that goes from perceptrons 👁️ to the geometry of the residual stream 📐! Heavily weighting in-class assessments, due to AI tools. (2024 was project-focused) Would be grateful for any exams/etc. folks can share, this is new territory for me! Please♻️!!

We're launching a weekly competition where the community decides which research ideas get implemented. Every week, we'll take the top 3 ideas from IdeaHub, run experiments with AI agents, and share everything: code, findings, all the successes and failures. It's completely free