Google just dropped "Attention is all you need (V2)" This paper could solve AI's biggest problem: Catastrophic forgetting. When AI models learn something new, they tend to forget what they previously learned. Humans don't work this way, and now Google Research has a solution.

Releasing a new "Agentic Reviewer" for research papers. I started coding this as a weekend project, and Yixing Jiang made it much better. I was inspired by a student who had a paper rejected 6 times over 3 years. Their feedback loop -- waiting ~6 months for feedback each time -- was

Love to see more fully open post-training recipes (this one multimodal reasoning). It's surprising how rare post-training data is because the opportunity for impact is huge. Lots of people will try it and simple data methods still can improve on SOTA.

The original PPO-based RLHF pipeline had 4 model copies: 1. Policy 2. Reference 3. Critic 4. Reward Model Recent GRPO-based RLVR pipelines have eliminated all of these models except for the policy. - The critic is no longer needed because values are estimated from group

A few updates * Had an amazing summer at MSR NE -- big thanks to David Alvarez Melis * Defending my PhD this semester and, assuming all goes well (😂), will join Brenden Lake's group at Princeton as a postdoc next year! * I’ll be at NeurIPS 12/1~12/7 -- happy to catch up or meet new folks!

🚀Introducing TMLR Beyond PDF! 🎬This is a new, HTML-based submission format for TMLR, that supports interactive figures and videos, along with the usual LaTeX and images. 🎉Thanks to TMLR Editors in Chief Hugo Larochelle Gautam Kamath ✈️ NeurIPS 2025 Naila Murray Nihar B. Shah Laurent Charlin!

LLM as a judge has become a dominant way to evaluate how good a model is at solving a task, since it works without a test set and handles cases where answers are not unique. But despite how widely this is used, almost all reported results are highly biased. Excited to share our

Using On-policy distillation leads to great performance boost even for small models😃Great work from Zechun Liu Ernie Chang and Changsheng (Vikas Chandra's group)!

I will be at #NeurIPS2025 in San Diego next week! Please stop by my poster if you are interested in the underlying mechanisms of #KnowledgeDistillation in generative models! - Paper: Why Knowledge Distillation Works in Generative Models: A Minimal Working Explanation - Wed 3

This paper shocked me 🤯 Everyone on X keeps bragging about “LLM-as-a-judge” like it’s some magical truth oracle. But this paper shows something insane: Most LLM evaluations you’ve seen are biased by design not because models are bad, but because the judge itself quietly

The paper shows that models that judge answers well usually also give good answers themselves. Chatbots are tested by humans or a strong model choosing better answers for prompts, which costs money. This work studies how closely 2 skills move together, writing good answers and

Got burned by an Apple ICLR paper — it was withdrawn after my Public Comment. So here’s what happened. Earlier this month, a colleague shared an Apple paper on arXiv with me — it was also under review for ICLR 2026. The benchmark they proposed was perfectly aligned with a

How to Properly do LLM-as-a-Judge Raw LLM-as-a-Judge scores are inherently biased due to how LLMs would often make mistakes This paper proposes a simple statistical method to correct the scores and calculate valid confidence intervals via a human-verified calibration set

AI agents can't learn from experience. Until now. Earlier this year, one of our customers asked us: "If your web agent uses my website once, will it be easier the next time it visits?" The question highlighted a fundamental gap between human intelligence and AI agents: Humans

I just arrived in San Diego and received my badge for NeurIPS 2025! I will be attending the conference on 12/1 - 12/7. If you'd like to discuss continual learning, machine unlearning, knowledge distillation, and related topics with me, please contact me! :)

Why can humans learn continually, but LLMs can’t? In our oral presentation at the NeurIPS CCFM workshop, through the lens of a human-curriculum, we explore how continual learning impacts LLMs. We also release our 23.4B-token dataset, models, and data + training pipeline. 🧵

Nice shout-out for Continual & Meta-Learning research by Rich Sutton!

We embedded all 5000+ NeurIPS papers! exa.ai/neurips Cool queries: - "new retrieval techniques" - "the paper that elon would love most" - "intersection of coding agents and biology, poster session 5" It uses our in-house model trained for precise semantic retrieval 😌

Check out our new paper Context Tuning for test-time adaptation!