Meta researchers unveil SPoT post-thinking: +17.7pp safety, +4.7pp math accuracy over RL baseline
polkirichenko · x · 2026-11-12
Meta AI researchers (incl. Jason Weston, Olga Golovneva) released Self-Improving Post-Thinking (SPoT): models generate a critical reasoning trace and verdict after answering, self-assessing safety, correctness, and confidence.
- Jointly trains answer generation and self-judgment via SFT distillation or RL, and can be introduced at continued pre-training, mid-training, or post-training stages
- Self-judgments let models both select the best rollout and learn to generate better answers
- Reported numbers: +17.7pp generation safety across 5 benchmarks over the base model (+3.6pp more from self-selection); +4.7pp accuracy across 4 math benchmarks over standard RL post-training (+6.3pp more from self-selection)
Paper: arXiv 2601.21343, "Self-Improving Training & Thinking".
Related event: Meta's SPoT: self-improving post-thinking training(4 posts)→
More from Research
- As models get stronger, the next trainable tasks look increasingly like alignment itself — herbiebradley · 2026-11-12
- RoboCap-10k: 10,000 hours of open egocentric robot data across 21 industries released — micoolcho · 2026-11-12
- Talent vs. luck model reproduced: equal funding is ~9x more efficient than funding winners — ChenhaoTan · 2026-11-12
- Stanford HAI dashboards show employment in AI-exposed jobs down over 4% this year — StanfordHAI · 2026-11-12
- Shun-ichi Amari on how 'do as you like' freedom birthed the math behind AI's 60 years — FrnkNlsn · 2026-11-12
- Formal verification proves a 3,000-step Sokoban optimum, with a bounty on beating it — banteg · 2026-11-12