SIS Brings Off-Policy Tokens Back On-Policy
burny_tech · x · 2026-07-12
- This paper introduces SIS to solve the "reusing rollouts after the policy has changed" problem in LLM reinforcement learning.
- Instead of merely clipping the off-policy ratio, it performs per-token detection: if the current model would still sample the token, it's treated as on-policy; otherwise, standard correction applies.
- The author claims this reintegrates tokens previously deemed off-policy back into on-policy training.
- Integrating SIS into GRPO, DAPO, and GSPO yielded improvements in math tasks and agentic search, while stabilizing training under stale rollout and MoE mismatch scenarios.
Related event: SIS Brings Reused Tokens Closer to On-Policy(2 posts)→
More from Research
- Stanford Data-Centric AI Workshop reopens site, extends poster deadline — jure · 2026-10-26
- Hugging Face wraps 6-month Training Agents series with open recordings and code — SergioPaniego · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Ito: a browser that detects browsing intent and auto-organizes tabs into Flows — andrewkuznet · 2026-10-26
- Cornell study says brains process language as linear sequences, challenging Chomsky's hierarchy — anirbanbandyo · 2026-10-26
- Hi3D V1.0 open-sourced with 2M training assets and 90% slicing success rate — kimmonismus · 2026-10-26