Tao Lin

post training infra @radixark | ex @meta superintelligence labs

[Tao Lin @nblintao]

Had a great time speaking at Training Agents IRL and joining the panel!

Big thanks to @ben_burtenshaw, @SergioPaniego, and everyone who helped make it happen! Really enjoyed the discussions with all the speakers.

Check out the full recording, including my talk on agentic RL with Miles and the panel: https://t.co/vVGvCN99EB

[Ben Burtenshaw @ben_burtenshaw]

We closed out the Training Agents series in person in San Francisco with four talks and a panel on post-training agents with open models in industry.

Enterprise RL (@t2govind Microsoft), environment design (Victor Barres Mercor), scaling agent RL with Miles (@nblintao RadixArk), and OpenEnv Arena (@xdotli BenchFlow).

Full recording, slides and links: https://t.co/fLAuoy4Vmd

[Sergio Paniego @SergioPaniego]

we had a lot of fun last night at @github HQ in San Francisco talking about training agents

🎥 recording coming soon

special thanks to @t2govind, @VBarres86633, @nblintao and @xdotli!

[Tao Lin @nblintao]

Thanks! You'll probably interested in the Fully Asynchronous RL section in our paper https://t.co/VPQ6LmWEWw

[Tao Lin @nblintao]

Excited to be speaking at Training Agents IRL tomorrow!

I’ll share how we’re bringing agentic RL and environments together with Miles, the open-source post-training framework developed by @RadixArk and adopted across the community. I’ll cover integrations with environments like OpenEnv, preventing reward hacking, and lessons learned from real experiments.

Can’t make it in person? Catch the livestream on YouTube (link in the quoted post)!

[Ben Burtenshaw @ben_burtenshaw]

Join us tomorrow on the training agents livestream, and here from experts about how they're scaling RL and evals to train frontier agents.

​We will look beyond single-turn RL to multi-step, environment-driven training: reward design, rollouts, benchmarks, and the practical gap between training gains and real-world agent behavior.

​What to expect Short, 10-minute talks with Q&A, followed by a closing panel on what works in RL for agents, where the bottlenecks are, and what the open-source ecosystem needs next.

​Speakers:
- Ben Burtenshaw (HF)
- Govind Kamtamneni (Microsoft)
- Victor Barres (Mercor)
- Zach Wentz (Reflection)
- Tao Lin (RadixArk)
- Xiangyi Li (Benchflow)

​Who should attend
ML practitioners, agent builders, and RL researchers.

https://t.co/7J7LqXBhfu

[Tao Lin @nblintao]

@EdwardSun0909 DuoBench

[Tao Lin @nblintao]

From inference to RL training and even agent sandboxes, an early look at our work on NVIDIA Vera Rubin. Great to see SGLang and Miles running across the stack!

[SGLang @sgl_project]

We brought SGLang to NVIDIA Vera Rubin and accelerated Kimi K3 inference.
Working closely with @NVIDIA, we optimized attention, MoE, and speculative verification kernels on early-access Rubin hardware.

Highlights:
• Up to 20% faster FP8 MLA at batch 1 / 128K context
• 20% faster KDA verification, with bitwise-identical output
• 5.9% end-to-end inference speedup from MoE tail fusion, removing 276 kernel launches per decode step

SGLang also powers rollouts for Miles' end-to-end RL training on Rubin, including agentic RL with 64 concurrent sandboxes on the Vera CPU.

Full results and engineering details 👉 https://t.co/xGwlR0ax4L

[Tao Lin @nblintao]

@zkwentz Congrats Zach!

[Ben Burtenshaw @ben_burtenshaw]

Training Agent IRL - SF: Next month we'll do our first in-real-life session on training an agent with the open-source stack. Don't worry, we'll live stream it too.

We'll hear 5 practical talks from engineers who've trained. And discuss the best ways to get your projects humming.

- Govind Kamtamneni (@MicrosoftAI)
- Victor Barres (@mercor)
- Zach Wentz (@reflection_ai)
- Tao Lin (@radixark)
- Xiangyi Li (@benchflow_ai)

It's hosted at @github office in San Francisco. It's part of Open Source AI week and adjacent to PyTorchCon, MCPCon, AgentCon, and WhateverCon.

Register to come on down! https://t.co/XGzBAujExC

[Tao Lin @nblintao]

Thanks for the Miles shoutout! Frontier labs have shown how critical scalable environments and sandboxes are for agentic RL. Proud to have Miles building the open training layer and connecting it with tasks, harnesses, and sandboxes, so more people can do agentic RL at scale.

[Sergio Paniego @SergioPaniego]

https://t.co/ibdTpH3xEQ

[Tao Lin @nblintao]

@SergioPaniego Thanks for the Miles shoutout! Frontier labs have shown how critical scalable environments and sandboxes are for agentic RL. Proud to have Miles building the open training layer and connecting it with tasks, harnesses, and sandboxes, so more people can do agentic RL at scale.

[RadixArk @radixark]

The full technical report for Miles is now available.

Miles is built for production-level post-training, with stability, efficiency, and flexibility at its core. The report covers its system design and how it works in practice.

[Tao Lin @nblintao]

@sergiopaniego has a great new piece on how RL environments evolved from the Universe/Gym era to today.

One thing I found especially interesting is how little the Gym reset/step interface itself has changed in almost a decade. With modern agent harnesses, though, that abstraction is starting to change. Sometimes the trainer works with the harness turn by turn. Sometimes the harness runs the full rollout itself, while the trainer captures the model calls for training. In the latter case, step() is no longer the natural unit of interaction.

In Miles, the open-source post-training framework we built, we support both styles through different integrations, including OpenEnv and Harbor. TITO preserves the exact token-in/token-out trajectory in either case, especially when the trainer isn’t driving every turn.

It’s exciting to see OpenEnv moving in this direction too, with support for loop-owning harnesses and the ongoing Harbor integration. It feels like a good time to look back at how we got here and think about what comes next.

[Sergio Paniego @SergioPaniego]

https://t.co/URCixWoMnO

[Tao Lin @nblintao]

@GenAI_is_real 有意思的是，在agent这个词送给硅基生命以前，人们一般把这种品质叫做agency

[Tao Lin @nblintao]

@unseenmars_ @thelocalblender Love the spinwheel idea!

[Tao Lin @nblintao]

@GenAI_is_real 年初的时候我还觉得claude相比于gpt来说写作风格更清新，现在局势反转了。感觉这种可能是最适合它cot的语言方式，现在rl太狠给带出来了

[Tao Lin @nblintao]

@zz30gs day 0 仙人！

[Tao Lin @nblintao]

@zhuzilinallen 于我心有戚戚焉

[Tao Lin @nblintao]

@shao256 @modal @waltergoat @colin_weld @paulgb Beautiful design!

[Tao Lin @nblintao]

@ShiDong14 Many miles to go. More meaningful with this team.

[Tao Lin @nblintao]

@james0zan @radixark Thanks @james0zan! Really enjoyed our conversations around agent environment infra. The https://t.co/XJL3p7ZmWu team has also been fantastic to collaborate with on the AgentENV integration. Excited to keep building together!