🚨Reasoning LLMs are e̵f̵f̵e̵c̵t̵i̵v̵e̵ ̵y̵e̵t̵ inefficient! Large language models (LLMs) now solve multi-step problems by emitting extended chains of thought. During the process, they often re-derive the same intermediate steps across problems, inflating token usage and

🚨Super excited to share that our paper “Stable Gradients for Stable Learning at Scale in DRL” was accepted at #NeurIPS as a Spotlight! ✨🌟 Paper: arxiv.org/abs/2506.15544 Huge thanks to my amazing coauthors. I’ve learned so much from each of you. 🫶 Onward! 🚀 #NeurIPS2025

Our work (arxiv.org/abs/2509.13237) can be seen as one instantiation of the paradigm proposed by Andrej Karpathy here. The behavior handbook is a repository of problem solving strategies which we show can be reused to get better and more efficient reasoning in the future. 🧵 for more

🚀We were delighted to anchor ⚓️ at the RoboPapers podcast with Chris Paxton and Michael Cho - Rbt/Acc to talk about SAILOR ⛵️. The episode is now live—tune in!

Your favorite library for training LLMs with RL (most likely) implements the objective incorrectly. I don’t know how this happened, and why this is so prevalent. We have a preprint coming out soon that will set the record straight. And yes, unbiased gradients do matter!

RSA is now on arxiv! 🥳 Check out these beautiful animations made by Moksh Jain which explain the basic idea really well. 📄 arxiv.org/abs/2509.26626

Do people really think he won the Turing award for the bitter lesson essay? What about all of his foundational work on RL?

Thanks for sharing our work Rohan Paul I was honestly surprised to see this simple method works so well across tasks and models. There is definitely a lot more we can get out of models using test time scaling alone.

Exciting to see Qwen adopt RSA (arxiv.org/abs/2509.26626) as their test-time inference method! Note: maintaining a good population size across aggregation is important!

This is really cool

LLMs can also come up with their own skills and reuse them to improve capabilities across various tasks. We explore this in two papers: 📜 - arxiv.org/abs/2405.12205 🧵 - x.com/prfsanjeevaror… 📜 - arxiv.org/abs/2509.13237 🧵 - x.com/Aniket_d98/sta…

[1/9] While pretraining data might be hitting a wall, novel methods for modeling it are just getting started! We introduce future summary prediction (FSP), where the model predicts future sequence embeddings to reduce teacher forcing & shortcut learning. 📌Predict a learned