H²SD mixes hindsight and self-distillation to improve RLVR reasoning training
Qiye Cai · hf · 2026-07-22
H²SD is a hybrid hindsight self-distillation method for RLVR that tries to fix sparse trajectory-level supervision.
The paper compares RLVR with OPD, OPSD, and RLSD, then combines their strengths: for successful trajectories, the teacher is used to modulate update magnitudes without changing the reward-driven direction; for failed trajectories, the teacher is conditioned on a hint plus verified answer and optimized with reverse KL. Experiments on several reasoning benchmarks show consistent gains over representative baselines while keeping optimization stable and generation efficient.
More from Research
- Stanford Data-Centric AI Workshop reopens site, extends poster deadline — jure · 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
- Ortet evaluates 34 open-weight models on 50 health tasks, finds general leaderboards insufficient — elmanmansimov · 2026-10-26