UC Berkeley Sky

Excited to share that MAP has been selected for ✨ICML Oral✨ We look forward to sharing the insights in the paper with the community And much much appreciations to everyone who participated in our study ❤️ MAP won’t be possible without your contribution to open science

Open-ended coding training data may no longer be the bottleneck: AI can scale open-ended tasks—and even outperform human-expert curation. FrontierCS team is releasing FrontierSmith: a system for synthesizing open-ended coding problems at scale. Starting from closed-ended coding tasks, FrontierSmith mutates, filters, and builds runnable optimization environments for long-horizon coding agents. In our experiments, FrontierSmith data trains stronger models than human-curated open-ended data on FrontierCS and ALE-bench. Blog: frontier-cs.org/blog/frontiers… Paper: arxiv.org/abs/2605.14445 Code: github.com/FrontierCS/Fro… Model: huggingface.co/runyuanhe/qwen…

🚀 Excited to release mKernel: a set of fast multi-node, multi-GPU fused kernels. 💻 Code: github.com/uccl-project/m… 📝 Blog: uccl-project.github.io/posts/mkernel/ mKernel fuses compute + communication into one persistent GPU kernel, covering both intra/inter-node with GPU-initiated communication. Amazing team: @yangzhouy, Chon Lam Lao, Costin Raiciu, Scott Shenker, @istoica05

Learning from rich textual feedback (errors, traces, partial reasoning) beats scalar reward alone for LLM optimization. GEPA demonstrated this for context-space optimization (prompts and agent harnesses), delivering frontier results at a fraction of the cost of RL. But context-only optimization is bounded by the base model's capability ceiling; weight updates can reach further. Very excited about this new line of work on Fast-Slow Training (FST), which interleaves context and model weight optimization! The idea is a clean division of labor between two interleaved loops: 🔹 Fast loop (context): GEPA reads rich rollout feedback updating the context layer. The context becomes a fast-updating scratchpad of what the model needs to know about this task, right now. 🔹 Slow loop (model parameters): RL updates the model's parameters conditioned on the evolving context. Because the prompt already carries task-specific nuances, the model parameters are freed from absorbing them and focus on what actually generalizes across tasks and pushes the frontier. ⦁ 3× more sample-efficient than RL on math, code, and physics reasoning ⦁ ~70% lower KL divergence from base at matched accuracy ⦁ Plasticity preserved: FST checkpoints respond better to additional RL on new tasks than RL-only ones ⦁ Continual learning across changing tasks (HoVer → CodeIO → Physics) where RL stalls the moment the task switches FST is a direction towards: ⦁ Addressing RL's pain points: entropy collapse, sparse rewards, long-horizon exploration ⦁ Providing a clean channel for rich feedback into weight updates ⦁ Demonstrating model-harness co-evolution ⦁ Discovery: Using fast context updates for broad exploration, while leveraging a continually improving model. Check out the full thread below:

1/ Thrilled to introduce T³: a corpus for RAG over reasoning tasks, built from thinking traces. We show that surprisingly RAG can improve reasoning— with the right corpus. Rag with Transformed Thinking Traces T³ gain by up to 43.9% on AIME 2025-2026. 🔗 arxiv.org/abs/2605.03344 🧵

Today, we’re releasing Continual Learning Bench 1.0: the first, realistic benchmark for measuring how AI systems can improve in online settings. Benchmarks today assume models are stateless. Each example is independent, and once a system finishes a task, it moves on as if nothing happened. But deployed AI systems should learn from experience. We tested 10+ frontier systems against novel, expert-validated tasks and find there’s still plenty of headroom for learning. (1/n)

Agent harness is as important as the model for cybersecurity. $300 in compute, 9 OSS-Fuzz projects, 14 security issues and 5 CVEs. The key lesson: you don’t need a secret model to find real security issues. You need an effective, affordable, reliable harness. 5 takeaways 🧵

Excited to announce that FrontierCS has been accepted to ICML 2026! 🚀 We are scaling our open-ended task set to 250 tasks (100 new tasks in 2026 Q1🔥), featuring long-horizon agent settings in Harbor and integration into real-world human contests. More exciting updates to come! Huge thanks to all our collaborators. #ICML2026 #AI #MachineLearning

Excited to share: MAP has been accepted as 🌟 ICML Spotlight 🌟 We hope MAP can provide data-driven insights that help the communities to work on various under-explored research directions around agent systems! Huge thanks & congrats to my amazing co-authors. See you all at Seoul! 🫡

What if one person could run a unicorn company? Today we're open-sourcing OMAR — a TUI that lets a single engineer orchestrate hundreds of AI coding agents in deep, recursive hierarchies. Built at Berkeley. Powered by tmux. github.com/lsk567/omar 🧵

Would you trust an AI agent to negotiate on your country's behalf at the G20? Real coordination is long-horizon, asymmetric, and non-binding; current multi-agent evaluations miss this. We build Cooperate to Compete (C2C): a testbed for LM agents coordinating with rivals. 🤝🔪🎭

Congratulations to Matei Zaharia on being awarded the ACM Prize in Computing! His development of open-source systems helped enable large-scale machine learning, analytics and AI at a global scale. @matei_zaharia @UCBerkeley 🔗 Read more: bit.ly/4vbNujK

🎯 One Year of AI-Driven Research at Berkeley [ADRS Blog #20] For the past year at Berkeley, we have been working on automating discovery with AI. In our blog post this week, we provide an overview of these efforts: the key problems we’re tackling, the frameworks and solutions we’ve built so far, and how these efforts fit into a broader vision for AI-driven scientific discovery. ✍️ Read the blog: ucbskyadrs.github.io/blog/berkeley-… 📖 ADRS Blog Series: ucbskyadrs.github.io

Introducing M²RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling We bring back non-linear recurrence to language modeling and show it's been held back by small state sizes, not by non-linearity itself. 📄 Paper: arxiv.org/abs/2603.14360 💻 Code: github.com/open-lm-engine… 🤗 Models: huggingface.co/collections/op…

Researchers spend hours and hours hand-crafting the strategies behind LLM-driven optimization systems like AlphaEvolve: deciding which ideas to reuse, when to explore vs exploit, and what mutations to try. 🤖But what if AI could evolve its own evolution process? We introduce EvoX, a meta-evolution pipeline that lets AI evolve the strategy guiding the optimization. It achieves high-quality solutions for <$5, while existing open systems and even Claude Code often cost 3-5× more on some tasks. Across ~200 optimization problems, EvoX delivers the strongest overall results: often outperforming AlphaEvolve, OpenEvolve, GEPA, and ShinkaEvolve on math and systems tasks, exceeding human SOTA, and improving median performance by up to 61% on 172 competitive programming problems. 👇

@karpathy Very nice results and great project! Sharing some of our experience with similar agentic frameworks at UC Berkeley: ADRS blog series: ucbskyadrs.github.io/blog/ GEPA: github.com/gepa-ai/gepa KISS: github.com/ksenxx/kiss_ai

AlphaEvolve is closed-source. We release 🌟SkyDiscover🌟, a flexible, modular open-source framework with two new adaptive algorithms that match or exceed AlphaEvolve on many benchmarks and outperform OpenEvolve, GEPA, and ShinkaEvolve across 200+ optimization tasks. Our new algorithms dynamically adapt their search strategy, and can even let the AI optimize its own optimization process on the fly! Results: 📊 +34% median score improvement on 172 Frontier-CS problems. 🧮 Matches/exceeds AlphaEvolve on many math benchmarks ⚙️ Discovers system optimizations beyond human-designed SOTA 🧵👇

We identified an issue with the Mamba-2 🐍 initialization in HuggingFace and FlashLinearAttention repository (dt_bias being incorrectly initialized). This bug is related to 2 main issues: 1. init being incorrect (torch.ones) if Mamba-2 layers are used in isolation without the Mamba2ForCausalLM model class (this has been already fixed: github.com/fla-org/flash-…). 2. Skipping initialization due to meta device init for DTensors with FSDP-2 (github.com/fla-org/flash-… will fix this issue upon merging). The difference is substantial. Mamba-2 seems to be quite sensitive to the initialization. Check out our experiments at the 7B MoE scale: wandb.ai/mayank31398/ma… Special thanks to @kevinyli_, @bharatrunwal2, @HanGuo97, @tri_dao and @_albertgu 🙏 Also thanks to @SonglinYang4 for quickly helping in merging the PR.

Introducing Slingshots // TWO: Research that ships. 14 projects, six institutions – let’s meet the batch 🧵

We are excited to announce that SkyRL now implements the Tinker API. Run Tinker training scripts on your own hardware with zero code changes. Try it out today: novasky-ai.notion.site/skyrl-tinker