Meta FAIR's SPoT teaches models to self-critique after answering, boosting safety and calibration

Meta AI's research team (authors include Jason Weston, Olga Golovneva, Ellen Xiaoqing Tan, etc.) released Self-Improving Post-Thinking (SPoT): a method that teaches models to generate a critical reasoning trajectory (post-thinking) and a verdict after producing their answer, addressing the problem of LLMs giving unsafe, incorrect, or overconfident outputs without realizing it. The paper (arXiv 2601.21343, with Chapter 3 covering post-thinking), official blog post, and Jason Weston's tweet thread are all live.

Confirmed

• Safety: across 5 benchmarks, generation safety improved 17.7pp over the base model; using self-evaluation for self-selection added another 3.6pp
• Reasoning: 4.7pp improvement across 4 math benchmarks
• Calibration: via RL with Brier score as the reward, the model was trained to predict answer correctness probability; confidence-estimation AUROC rose from 0.61 to 0.84, selective accuracy improved from 79%, and the model learned to abstain when uncertain
• Training: answer generation and self-judgment are trained jointly via SFT distillation or RL
• Generalization: trained only on answerable math problems, SPoT also improved the model's ability to abstain on unanswerable questions and generalized beyond the math domain (clear AUROC gains on out-of-domain benchmarks like BBQ, MoralChoice, GPQA-D, MediQ)

Why It Matters

SPoT internalizes "self-critique and confidence estimation" as a trainable capability rather than a bolt-on patch, improving safety, reasoning, and calibration simultaneously; its ability to generalize from math-domain training to out-of-domain unanswerable questions suggests the method could boost model reliability and self-awareness in open-ended settings.

2026-11-12 ~ 2026-11-12 · 7 related posts

• [source] Meta FAIR's SPoT: post-thinking self-judgment lifts safety benchmarks by 21.3 points — jaseweston · 2026-11-10
• Meta researchers unveil SPoT post-thinking: +17.7pp safety, +4.7pp math accuracy over RL baseline — polkirichenko · 2026-11-12
• How SPoT jointly trains answer generation and self-judgment across training stages — polkirichenko · 2026-11-12
• [source] SPoT results: +17.7pp safety across 5 benchmarks, +4.7pp math accuracy over RL — polkirichenko · 2026-11-12
• SPoT boosts calibration: AUROC 0.61 to 0.84 while learning to abstain — polkirichenko · 2026-11-12
• SPoT generalizes out-of-distribution: models learn to flag unanswerable questions — polkirichenko · 2026-11-12
• [source] SPoT paper "Self-Improving Training & Thinking" now on arXiv with blog and thread — polkirichenko · 2026-11-12