Self-evolving RL loop on saturated data lifts AIME25 from 22.1 to 32.7 and transfers to Llama-3.1-8B
shangbinfeng · x · 2026-10-22
Follow-up on RL over saturated data: the negative-rollout method still works when saturated and unsaturated data are mixed (+1.8-3.1 pts) and transfers to Llama-3.1-8B. It also enables a self-evolving loop where newly solved prompts join the saturated pool (+443 by step 80) while AIME25 accuracy climbs from 22.1 to 32.7.
More from Research
- Research Progress Graph: modeling authentic research trajectories, 35 agents found incomplete — ZhiyuChen4 · 2026-10-22
- EMNLP 2026 paper uses prediction-powered inference to quantify how much human annotation LLM judges save — shi_weiyan · 2026-10-22
- Ask-to-Distill: synthesizing questions for answer-free on-policy distillation — UIUC-CS · 2026-10-22
- RFD4 team swaps self-conditioning for recurrent updates, echoing looped flows — AllThingsApx · 2026-10-22
- MIT ML technique breaks records for 2-6 week weather forecasts by learning to correct model errors — LesterMackey · 2026-10-22
- Harvard undergrad's ML subseasonal weather forecasting work, started at Microsoft Research, lands in Nature — LesterMackey · 2026-10-22