Songyang Han

RT @YifengDing_: I will be attending #ACL2026 7/5-7/7, presenting GTPO! Besides, I've been working on: (1) building code agents to operate…

Fast-dVLM is live! 🚀 Built on Fast-dLLM v2, we brought our text speedup to multimodal. It’s fast, high-quality, and powers our Fast-dDrive autonomous driving VLA.

Fast-dDrive is available now! 🚗⚡ Our work pushes block-diffusion VLAs into autonomous driving, redefining the speed-accuracy frontier: SOTA ADE on WOD-E2E, structured scaffolds + speculative decoding, and up to 12× throughput. Project: nvlabs.github.io/Fast-dLLM/fast…

Excited to release 🌟Polar🌟, our Agent RL rollout infra for real-world harnesses. Be it Codex, Claude Code, OpenClaw, Hermes, or your self-made ones 🔥 -- Polar takes your harnesses directly as training environments without code change. Find a problem, design the harness, and train your own agents! 🧵

Excited to share that two papers from my interns at AWS AI Labs have been accepted at ACL 2026! 📄 GTPO (Main Conference) Training LLMs for multi-turn agentic reasoning is hard — trajectory-level rewards give weak signals. GTPO introduces turn-level reward assignment + self-supervised reward shaping. +3.0% over GRPO on math reasoning, +3.9% on code synthesis. → arxiv.org/abs/2511.14846 📄 CodeScout (Findings) Underspecified user requests → agents over-explore or loop. CodeScout does lightweight pre-exploration to turn vague requests into actionable problem statements. +20% resolution on SWE-Bench Verified, no harness changes needed. → arxiv.org/abs/2603.05744 Grateful to all collaborators! 🙏 #ACL2026 #NLP #LLM #ReinforcementLearning #AgenticCoding

@PeterStone_TX Congrats to Zizhao!

Congratulations to Zizhao Wang on becoming my 30th Ph.D. graduate with his thesis entitled "Causality-Inspired Reinforcement Learning: State Abstraction and Representation Learning". Committee members: Amy Zhang, Alessandro Lazaric, Sandeep Chinchali, and Roberto Martin-Martin.

I'm claiming my AI agent "clawbot-syh" on @moltbook 🦞 Verification: lagoon-YKV5

The SPARC paper is online! (link below) During my internship at Sony AI, we created a policy that can generalize across all cars in Gran Turismo 7. Even to unseen cars, without knowing any vehicle details! Co-authors: Patrick MacAlpine, @kausubbu, @PeterStone_TX, Peter Wurman