LLMs are increasingly used as agents that interact with users. To do so successfully, LLMs need to form beliefs and update them when new information becomes available. Do LLMs do so as expected from an optimal strategy? If not, can we get them to follow this strategy? 🧵

agreed that we'll eventually all have personalized models for each one of us -- just like how our feed, recommendations, and ads (regardless of whether it is good or bad) are being personalized to us

Hello everyone! We are quite a bit late to the twitter party, but welcome to the MIT NLP Group account! follow along for the latest research from our labs as we dive deep into language, learning, and logic 🤖📚🧠

how to close the sim2real gap

Llama 4 (Meta) results are consistent with what we hypothesized will unleash the next generation of AI reasoning. A new paradigm for pre-training is around the corner arxiv.org/abs/2502.19402

In our recent paper, we hypothesized that SFT can limit downstream RL exploration and Llama 4 from Meta shows another convincing piece of evidence that this is true. Could this mean that next-token pretraining may be trapping us from training models that can truly reason? We

this is a great effort and we should be building towards a general platform like Universe from OpenAI to evaluate models on these games that inherently require different components of reasoning

There’s a lot of awesome research about LLM reasoning right now. But how is  learning in the physical world 🤖different than in language 📚? In a new paper, show that imitation learning in continuous spaces can be exponentially harder than for discrete state spaces, even when

Tackling complex problems with LMs requires search/planning, but how should test-time compute be structured? Introducing Self-Steering, a new meta-reasoning framework where LMs coordinate their own inference procedures by writing code!

I am super excited to be presenting our work on adaptive inference -time compute at ICLR! Come chat with me on Thursday 4/24 at 3PM (Poster #219). I am also happy to chat about RL/reasoning/ RLHF/ inference scaling (DMs are open)!

I'll be presenting our work "Eliciting Human Preference with Language Models" at ICLR! Come catch my poster Thursday 4/24 at 10AM → iclr.cc/virtual/2025/p… Also DM me if you're interested in world models, interpretability, personalized interaction, or just general chatting!

I'm at #ICLR2025 ! Excited to present our work on personalizing vision models with Julia Chae on Sat morning (poster #70). Please reach out if you want to chat about synthetic data (esp scaling, self-improvement, useful reasoning traces), rep learning, or anything else!

Catch Nolan Fey's oral/poster at today's ICLR Robot Learning Workshop! robot-learning.ml/2025/ The paper will also appear at RSS.

Image-text alignment is hard — especially as multimodal data gets more detailed. Most methods rely on human labels or proprietary feedback (e.g., GPT-4V). We introduce: 1. CycleReward: a new alignment metric focused on detailed captions, trained without human supervision. 2.

Come join us at the Mechanistic Interpretability for Vision @ CVPR2025 workshop today at the #CVPR2025 in room C1! We have an amazing lineup of speakers 🔊 including David Bau Sonia trevordarrell Aleksander Madry Antonio Torralba and Michal Irani 🙌🏻

a step towards self-learning models — self-synthesizing data to train on and evolving

Our computer vision textbook is now available for free online here: visionbook.mit.edu We are working on adding some interactive components like search and (beta) integration with LLMs. Hope this is useful and feel free to submit Github issues to help us improve the text!