🔥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

Embodied chain of thought (ECoT) is a powerful tool to get VLAs to think through problems, but why does it work? In our new work, we analyze various lightweight ECoT-like strategies, including co-training, to see what is the "minimal" amount of reasoning that can boost VLAs 🧵👇

Excited to have two papers accepted to ACL 2025 main! 🎉 1. ChatBench with jake hofman Ashton Anderson - we conduct a large-scale user study converting static benchmark questions into human-AI conversations, showing how benchmarks fail to predict human-AI outcomes.

🔐 Frontier AI is reshaping cybersecurity, raising critical new questions: 🔍 What is its current impact? ⚖️ Who stands to benefit more—attackers or defenders? 🛡️ How can we mitigate the risks? Addressing these challenges requires coordinated efforts across AI & security

Invited talked are announced. icml.cc/virtual/2025/e… Jon Kleinberg Pamela Samuelson Frauke Kreuter Anca Dragan Andreas Krause

It was a pleasure to share my PhD work on measuring pain in patients under anesthesia using wearable devices with UC Joint Computational Precision Health Program! My lab continues this work by studying diseases using novel wearable devices and algorithms. See our website at subramanianlab.com!

Just added a new kinetic energy regularization task to mink, you can get it by upgrading to 0.0.11.

Just did a major revision to our paper on Lifelong Knowledge Editing!🔍 Key takeaway (+ our new title) - "Lifelong Knowledge Editing requires Better Regularization" Fixing this leads to consistent downstream performance! Tom Hartvigsen Ahmed Alaa Gopala Anumanchipalli Berkeley AI Research

How can one efficiently simulate phylodynamics for populations with billions of individuals, as is typical in many applications, e.g., viral evolution and cancer genomics? In this work with Michael Celentano, W. DeWitt, & S. Prillo, we provide a solution. doi.org/10.1073/pnas.2… 1/n

Goal-conditioned RL (GCRL) is great - unsupervised, can use data (in offline mode), flexibility to define tasks at test time. But can we run GCRL on *language data*?? In our new work we show that language GCRL enables sophisticated test-time reasoning for interactive tasks! 🧵👇

🚀 Excited to share the most inspiring work I’ve been part of this year: "Learning to Reason without External Rewards" TL;DR: We show that LLMs can learn complex reasoning without access to ground-truth answers, simply by optimizing their own internal sense of confidence. 1/n

Closed-loop robot policies directly from human interactions. No teleop, no robot data co-training, no RL, and no sim. Just Aria smart glasses. Everyday human data is passively scalable and a massively underutilized resource in robotics...More to come here in the coming weeks.

Happy to announce the first workshop on Pragmatic Reasoning in Language Models — PragLM @ COLM 2025! 🧠🎉 How do LLMs engage in pragmatic reasoning, and what core pragmatic capacities remain beyond their reach? 🌐 sites.google.com/berkeley.edu/p… 📅 Submit by June 23rd

Excited to present FastTD3: a simple, fast, and capable off-policy RL algorithm for humanoid control -- with an open-source code to run your own humanoid RL experiments in no time! Thread below 🧵

Guide your favorite protein generative model with experimental data? Meet ProteinGuide - a method to condition pre-trained models on properties without retraining. We validated it both in silico by guiding ProteinMPNN and ESM3 on 3 tasks and in vitro by engineering base editors.

FastTD3: "Minimum innovation, maximum results" Not the paper we had planned to write, but one of the works I am most proud of. We wanted to make sure our baseline (TD3) was a very solid baseline, so we added a few things that are already known to help in RL (large,

Congratulations to BAIR students and faculty for their Best Paper Awards at the recently held #ICRA2025 in Atlanta. BAIR Researchers from Masayoshi Tomizuka's lab and the Berkeley DeepDrive Consortium won the Best Paper in Automation for their paper "Physics-Aware Robotic

Stare at policy improvement and diffusion guidance, and you may notice a suspicious similarity... We lay out an equivalence between the two, formalizing a simple technique (CFGRL) to improve performance across-the-board when training diffusion policies. arxiv.org/abs/2505.23458

Ever wondered if the way we feed image patches to vision models is the best way? The standard row-by-row scan isn't always optimal! Modern long-sequence transformers can be surprisingly sensitive to patch order. We developed REOrder to find better, task-specific patch sequences.