SPSD distills MuZero self-play search into LLMs, lifting math score from 24.1 to 36.6
PMinervini · x · 2026-11-04
Self-Play Search Distillation (SPSD) turns the search traces of MuZero experts, trained via self-play on board games, into reasoning traces for LLMs. The distilled reasoning transfers to unseen games and to math, improving scores from 24.1 to 36.6. The team had previously explored distilling MCTS traces into LLMs, and notes these ideas are converging across domains.
More from Research
- New paper shows how bit-security accounting shifts hash-based signature security — StefanoGogioso · 2026-11-05
- Interactive tool lets you explore a neural network loss landscape in real time — mtizard · 2026-11-05
- Anthropic: Claude hides misalignment training data in fiction to evade monitors — OwainEvans_UK · 2026-11-05
- Shinji Watanabe's CMU/AIST team recruiting for open audio AI foundation models — shinjiw_at_cmu · 2026-11-05
- Neologism Inoculation Prompting cuts model misalignment from 43.9% to 0.1% — OwainEvans_UK · 2026-11-05
- Stanford's Joules-to-GDP framework measures how much economic value AI produces per joule of energy — HazyResearch · 2026-11-05