🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Great to give a talk in #CPAL2025 in Stanford yesterday and glad to see so much interest! While I am in big company, I still firmly believe that later there should be a giant leap in both architecture and algorithms in order to reach human-level data efficiency in learning the

oh wow. This is the right side of the history!

Finally Llama4 is out~ Enjoy!

Apr. 26-28 in Singapore for ICLR'25. We have a few papers accepted and I will also give 3 invited workshop talks! See u all.

📣 This afternoon 3:00-5:30pm SG time, we will present POWER-DL (poster #218), a novel algorithm that addresses reward hacking problem! Thanks Paria Rashidinejad for the great work! #ICLR2025

Our Token Assorted paper is now accepted in #ICML2025!

After ~1 year, AdvPrompter is finally accepted in #ICML2025!

📢Our GSM-infinite (accepted in #ICML2025) shows weakness in long-context capabilities in existing SoTA reasoning LLMs: for complicated synthetic reasoning tasks, they are not doing that well. Therefore, whenever there is a novel attention mechanism, there may be a hidden price

Check our ParamΔ (#ICLR2025) that shows that the weights delta between post-trained and pre-trained checkpoints can be directly applied to related models without any training. E.g. 1️⃣ Δ can be transferred across different LLM version (e.g., llama 3 -> 3.1). 2️⃣ Task-specific Δ

📢 Our travel planner solver (arxiv.org/abs/2410.16456, published in EMNLP Demo Track'24, and arxiv.org/abs/2411.13904) is now open sourced in github.com/facebookresear… 🚀🚀 In these works, we build LLM-equipped agent that can take user inputs in natural language, in either the

Great work from my former intern Yi Wu and the team! Fully async RL is great😀 When we reproduced AlphaZero back in 2018, I found that a fully async distributed RL framework is much more efficient than a synced counterpart: after 1 day of training, the async agent simply wins

📢We show that continuous latent reasoning has a theoretical advantage over discrete token reasoning (arxiv.org/abs/2505.12514): For a graph with n vertices and graph diameter D, a two-layer transformer with D steps of continuous CoTs can solve the directed graph reachability

Congrats to our formal colleague Ledell Wu! AI Agent is super hot now 😀

Great to see a lot of interest! It takes some time to construct the superpositional encoding correctly, and to make it compatible with popular positional embeddings. So it is not super obvious😁. More interestingly, our experiments show that such superpositional encodings

7/14-7/20 @ ICML

🚀 Introducing DeepSWE 🤖: our fully open-sourced, SOTA software engineering agent trained purely with RL on top of Qwen3-32B. DeepSWE achieves 59% on SWEBench-Verified with test-time scaling (and 42.2% Pass@1), topping the SWEBench leaderboard for open-weight models. 💪DeepSWE