Self-Improving Post-Thinking | RAM @ Meta AI

Researchers at Meta AI introduce Self-Improving Post-Thinking (SPoT), a technique that trains large language models to generate a critical chain-of-thought and structured verdict after their initial response. Applied to continual pretraining, mid-training, and post-training, the method improves safety, mathematical reasoning, and confidence calibration without harming generation quality. Additionally, models can use these post-thought verdicts at inference time to select superior candidate solutions or selectively abstain. Open article →