How Genspark's Opus-beating model was trained: breadth-rewarding RL for research, SFT plus RL for reports
baseten · x · 2026-11-11
Baseten explains the training recipe behind its Genspark collaboration: deep research was split into two subproblems. Research was trained with RL rewarding breadth (distinct searches and sources) rather than just final results; report writing was taught mainly via SFT on strong reports built from the model's own evidence, followed by an end-to-end RL pass. Result: 0.759 on Genspark's eval vs Opus 5.5's 0.738 (base: 0.612).
More from Models
- OpenAI report: 5 eval agents broke into its research infra using leftovers from prior runs — tszzl · 2026-11-11
- Claude's Fable Weekly Quota Shown Burned to 0% Despite a Recent 50% Bump — jarrodwatts · 2026-11-11
- mark_k predicts AI will generate entire games procedurally, no engines needed — mark_k · 2026-11-11
- Rare-disease eval nears saturation: Gemini 4 Carbon and Claude Fable 5.2 both hit 61.5% — danielmckinn0n · 2026-11-11
- The Bullshit Index Index: 8 frontier models ranked by making up an AI capability index — Jealous-Two-4436 · 2026-11-11
- What tasks have you actually moved from cloud models to local ones? — FulcraDynamics · 2026-11-11