Amazon and Duke: keeping gradient from just 1 token per rollout improves LLM reasoning
jiqizhixin · x · 2026-10-11
In "Extremely Sparse Supervision Incentivizes Reasoning Ability," Amazon and Duke University randomly keep the gradient signal from only one token per multi-thousand-token math rollout and mask the rest — training should collapse, but reasoning improves instead. The result challenges the assumption that dense token-level supervision is essential to On-Policy Distillation and sheds light on its inner mechanics.
More from Research
- MathEconLib open-sourced: Axiom Math and top economists bring machine-checked proofs to economics in Lean 4 — soumitrashukla9 · 2026-10-23
- Ken Ono invites the community to build MathEcon on the new Lean 4 economics library — soumitrashukla9 · 2026-10-23
- IFM open-sources RL360: 8,192 concurrent sandboxes and ~1.3M rollouts for agentic RL — rupspace · 2026-10-23
- Science lifts expression of concern on 2024 AI debunking-conspiracies paper; conclusions unchanged — DG_Rand · 2026-10-23
- Embodied First Exam benchmarks coding agents on 1,090 tasks across 44 robots — ManlingLi_ · 2026-10-23
- Anima Anandkumar on neural operators: AI weather models beat supercomputers — AnimaAnandkumar · 2026-10-23