1/8 Excited to announce the release of SkyRL v0.5.0!
Some highlights:
- Support for Beam-1.0, DeepSeek-V4.1-Flash, Qwen3.8-Flash-Next
- Score centering for stable RL with 4-bit rollouts
- 1M-token context training for GLM-5.3-Flash + optimizer offload to disk
- A native On-Policy Distillation (OPD) trainer
- Better VLM training: sequence packing, context parallelism and images in agentic RL
Blog: blog.skyrl.ai/blog/skyrl-v0-…
Details in 🧵
Nice empirical results! We need a solid study of SC + async RL. Lots still open:
• SC alone ≠ on-policy gradient in expectation (can even flip sign)
• Full trajectory-level IS is unbiased, but variance explodes with length
• Truncation → SC+TIS keeps residual bias under lag
Need more principled, scientific work on async RL for LLMs, esp. long-horizon tasks.
🤔 Is score centering (SC) all we need for stable async RL?
People often stack truncated importance sampling (TIS) or KL penalties on top of SC. In a fun side project, we tested how much they help.
We found that SC alone often matched on-policy. But when it fell short, adding more corrections doesn't necessarily help. 🧵
Now that we've been asking GPT, Claude and Gemini and let them search for us, what make them click, read, and present?
It turns out they are highly susceptible, and we are able to learn a simple content presentation strategy that makes 𝐆𝐏𝐓-𝟓-𝐦𝐢𝐧𝐢 𝐜𝐥𝐢𝐜𝐤 𝟗𝟖.𝟓% 𝐨𝐟 𝐭𝐡𝐞 𝐭𝐢𝐦𝐞 (as opposed to 17.1%), and it even transfers to the frontier 𝐆𝐏𝐓 𝟓.𝟓, 𝐦𝐚𝐤𝐢𝐧𝐠 𝐢𝐭 𝐜𝐥𝐢𝐜𝐤 𝟗𝟑.𝟑% of the time, 𝐎𝐩𝐮𝐬 𝟒.𝟖 𝐚𝐭 𝟗𝟒.𝟏% of the time, and 𝐆𝐞𝐦𝐢𝐧𝐢 𝟑 𝐅𝐥𝐚𝐬𝐡 𝐚𝐭 𝟗𝟖.𝟗% of the time!
Introducing our latest study AgentBait: agentbait.github.io
Do you know that past experience can substantially reduce the token cost of your coding agent?
Introducing our recent work TokenEconomizer: 18.4–30.9% lower cache-aware cost on SWE-Bench, with comparable task success. We deployed it as a plugin for mainstream coding agents - feel free to try it out and let us know your feedback!
Plugin: github.com/microsoft/TokenEc…
Paper: arxiv.org/pdf/2610.22732
Project page: tokeneconomizer.github.io
(1/3)
Introducing ✨Token-Gated Language Models✨ for up to 11x faster prefill by learning to skim context, conditionally forwarding only tokens that matter.
Akin to the shift from dense models to MoEs, the next frontier of efficient LLMs is conditional computation over a new scaling axis: context tokens🧵👇
Releasing FLAT: Flash Linear Attention on TPUs 🚀
Pallas training kernels for Linear Attention, Mamba-2 and GLA. On one TPU v6e, forward + backward is up to 3x faster than FLA on an H100 for GLA, and 1.5x faster for Mamba-2.
🧵
Excited to announce the release of SkyRL v0.4.0!
Some highlights:
- Support for GLM 5.3 Flash, GLM 5.3, and Kimi K2.7
- Top-P sampler replay and optimized R3 data transfer
- Full FP8/MXFP8 RL
- Sharded P2P weight transfer
Blog: blog.skyrl.ai/blog/skyrl-v0-…
Details in 🧵
What happens if you run @gepa_ai style evolutionary search on @NVIDIAAI's Graph-as-Policies? Do the robot policies actually get any better?
My latest blogpost explores this: kael.tech.blog/2026/09/26/ro…
Does your Claude model really need Claude Code…? 🤔
We evaluate 7 models on Claude Code, Codex, and Pi. Three surprising findings emerge:
1️⃣Harness choice has little effect on task success rate, but can significantly affect the cost
2️⃣A simple harness can be competitive
3️⃣The native harness isn’t always the best.
Millions of people are using coding agents, but the impact of harness choice remains unclear.
(1/n) More details in the thread. 🧵
Agents let us build systems for different workloads and requirements. But… can we trust what they build?
We release 🌟SkySynth🌟: an engine for synthesizing high-performance, just-in-time (JIT) systems we can trust, by co-evolving formal proofs and tests alongside the code.
Results:
💿 KV stores up to 2.3× faster than Redis and FASTER + formally verified stores with 2.9× Claude Code's pass rate
🚏 Model routers up to 48% cheaper than a general router
⚡ Specialized inference engine with 2.2× the throughput of vLLM/SGLang
🧵👇
Today, SkyPilot is out of stealth.
Building custom intelligence is now existential. We help frontier AI teams build intelligence faster by removing their biggest bottleneck: AI compute fragmentation.
Frontier teams like @appliedcompute, @AbridgeHQ, @hippocraticai, @hcompany_ai, and @nubank already run on SkyPilot, with 10x faster time-to-intelligence and double-digit increase in GPU utilization.
AI teams today get compute anywhere they can. They then firefight compute fragmentation across providers. Researchers burn time on workload setup. Infra gets paged when GPUs go down. Frontier teams build slowly even on the fastest compute.
@skypilot_org turns your fragmented compute into one AI supercomputer, so you run frontier workloads faster. Many users manage 10,000+ GPUs across providers with SkyPilot. GPU hours consumption has grown 6x in the last 6 months.
1/ We're launching SkyPilot Platform — the AI compute platform for frontier AI teams to manage large GPU fleets and accelerate building custom intelligence.
Optimized for fleet management, team governance, and frontier workloads — pretraining, post-training, multi-cluster serving, and sandboxes. SkyPilot open source users can switch to the platform with a server URL change.
2/ We've raised over $20M led by @Lux_Capital (@breeves08), with participation from @AmplifyPartners (@dauber, @lennypruss), @coatuemgmt, @FoundationCap (@ashugarg , @JayaGup10), @RaceCapital, @thehousefund, and top operators like @alighodsi (CEO, Databricks), @JeffDean (Chief Scientist, Google), @rauchg (CEO, Vercel), @amasad (CEO, Replit), @ClemDelangue (CEO, @huggingface) and more.
We're hiring across Engineering and GTM to deliver the platform for the next decade of AI.
Above all, I'm excited to be building with the incredible team we've assembled, along with my cofounders Zhanghao @Michaelvll1, Romil @bromil101, Scott, and Ion @istoica05.
If you firefight AI compute, let's build.
How can we train small agentic models that are highly capable of terminal use and coding?
Announcing OpenThoughts-Agent + OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
What if a robot policy weren't a neural net or a test-time chat loop, but a multi-file code repo selected from a Pareto frontier of genetically evolved candidates? RHO moves all its LLM exploration to training time, then runs that repo on scenes it was never trained on. 🧵👇🏽
👀Humans compare images by looking back and forth. Many open-weight VLMs encode each image independently, and defer comparison to the LM.
We introduce SVE: Stateful Visual Encoders for Vision-Language Models, where the visual encoder itself becomes change-aware.
🌐Project: statefulvisualencoders.githu…
📰Paper: arxiv.org/abs/2606.04433
💻Code: github.com/StatefulVisualEnc…
1/n
Static benchmarks are dying — they tend to get saturated quickly.
Evaluation and training data should co-evolve with frontier models.
We released BenchEvolver — a framework that automatically evolves saturated problems into harder, verified tasks for evaluating frontier models, which can also serve as useful self-improvement signals for RL.
New work from UC Berkeley @berkeley_ai@BerkeleyRDI@BerkeleySky
Project Page: benchevolver.github.io
Paper: arxiv.org/abs/2606.01286
Agents are finding more vulnerabilities than ever. But it turns out there are gaps in existing vulnerability discovery. Over the past 90 days vs. a year ago, web vulnerabilities (XSS/SQLi/CSRF) are down 66% and memory safety exploitability is down 3.5x.
We built the Agentic Vulnerability Coverage Map to track it all, updated daily. Introducing the Berkeley Vulnerability Initiative: vuln.cs.berkeley.edu. ⤵️