Snowflake Doubles Down on Agentic RL: Overfit First, Then Scale to 65.7 on SWE-bench
StasBekman · x · 2026-10-28
Snowflake AI Research shares three Agentic RL milestones, arguing long-horizon agent training demands the right recipe, eliminating reward hacking, and co-designing systems with modeling — not just compute. 1) Arctic-SWE: overfit first — cleanly fit and hold a 160-task sandbox (96% reward held) where public recipes Prime-RL and SkyRL crash or collapse. 2) HybridResearcher: the SWE recipe transfers to deep research, bridging SQL tables and web search in one reasoning chain. 3) Curated data lifts Qwen3.5-4B from 47.4 to 65.0 and Qwen3.5-9B from 56.2 to 65.7 on held-out SWE-bench Verified.
More from coding & agent
- Agents hate latency: Claude Opus drops APIs past 7.2s, and 'agent marketing' is born — danshipper · 2026-10-28
- Freelancer lands a Xiaohongshu gig in 5 minutes using Skillry's Motion skill — yihui_indie · 2026-10-28
- LaunchMotion skills let Claude Code render 4K product launch videos in 5 minutes — yihui_indie · 2026-10-28
- Dev rebuilds app with Rork and checks its iPhone Duo listing at Apple's SoHo store — rudrank · 2026-10-28
- Frameworks should be built for AI agents, not humans: the case for rigor over taste — samgoodwin89 · 2026-10-28
- Nous Research ships ElevenLabs SFX plugin for Hermes Agent with looping and variation tools — NousResearch · 2026-10-28