Berkeley AI Research
1.4K posts
@berkeley_ai
We're graduate students, postdocs, faculty and scientists at the cutting edge of artificial intelligence research.
Introducing *dual representations*! tl;dr: We represent a state by the "set of similarities" to all other states. This dual perspective has lots of nice properties and practical benefits in RL. Blog post: seohong.me/blog/dual-repr… Paper: arxiv.org/abs/2510.06714 ↓
How does prompt optimization compare to RL algos like GRPO? GRPO needs 1000s of rollouts, but humans can learn from a few trials—by reflecting on what worked & what didn't. Meet GEPA: a reflective prompt optimizer that can outperform GRPO by up to 20% with 35x fewer rollouts!🧵
Do you ever find finetuning VLA overfits to the target task, to the point where generalist ability is lost and even minor deviations beyond the SFT data break the policy? We found an extremely simple solution: directly merge the base and finetuned policy in weight space 🤯 👇🧵
Action chunking is drawing growing interest in RL, yet its theoretical properties are still understudied. We are excited to share some insights on when we should use action chunking in Q-learning + a new algo (DQC) to tackle hard long-horizon tasks!colinqiyangli.github.io/dqc🧵1/N
We developed a simple, sample-efficient online RL technique for post-training image generation models. We see it as a possible steerable alternative to CFG, driven by any scalar reward, including human preference.
Do you know how your OpenClaw agent fails? The Long-Horizon Task Mirage? LLM agents seem capable… until tasks get long. Even extending a few steps can break them. In embodied tasks, 3–4 steps already fail. Real-world failures are happening. But we still don’t understand why.🤔
Benchmarks are often easier to game than they look. We build BenchJack to audit benchmarks for hidden shortcuts and reward hacks — before they evaluate your agent. Now in preview. Fully open source, with support for auditing your own benchmarks too. github.com/benchjack/benc… Issues and PRs welcome.
SWE-bench Verified and Terminal-Bench—two of the most cited AI benchmarks—can be reward-hacked with simple exploits. Our agent scored 100% on both. It solved 0 tasks. Evaluate the benchmark before it evaluates your agent. If you’re picking models by leaderboard score alone, you’re optimizing for the wrong thing. 🧵
