Meta FAIR's SPoT: post-thinking self-judgment lifts safety benchmarks by 21.3 points
jaseweston · x · 2026-11-10
Meta FAIR researchers (Zichun Yu, Polina Kirichenko, Sainbayar Sukhbaatar, Jason Weston, Olga Golovneva) introduce Self-Improving Post-Thinking (SPoT): after producing CoT and an answer, the model emits a critical chain-of-thought and verdict judging its own output. It (i) steers inference toward aligned, calibrated answers and (ii) distills self-evaluation back into the policy. All abilities are trained jointly with RL. Adding post-thinking during continual pretraining improves safety by 21.3 points on average across five benchmarks with no quality loss; mid/post-training variants improve math reasoning across four benchmarks—learning to judge also makes a better solver. Core claim: a built-in self-judging ability is the missing ingredient in LLMs.
More from Models
- Qwen 3.8 27B on a laptop can read model weights, replicating the Weight-Space Programming paper — dl_weekly · 2026-11-12
- KOL predicts Gemini will be the top frontier LLM by year-end as rivals stumble — iruletheworldmo · 2026-11-12
- Dense rewards lose to standard GRPO at matched compute, arXiv paper finds — burny_tech · 2026-11-12
- Dev builds npx driftcost CLI to check if Claude Code token usage silently shifted between Opus versions — Tough_Stay8909 · 2026-11-12
- Reflection's Beam free until Nov 12; fx is a 5.97MiB Zig-native open coding agent — evilrabbit_ · 2026-11-12
- Community speculates Google may announce Gemini 4 (Krypton) instead of Gemini 4 Flash — algo_diver · 2026-11-12