Quiet-STaR introduces a novel approach for language models to teach themselves reasoning by inferring underlying thought processes behind texts. Discover more in the full text.

• The model generates rationales at each token to justify future text segments, enhancing model prediction accuracy.
• A new sampling algorithm and learnable tokens are introduced to track the beginning and ending of thoughts.
• Generated rationales significantly aid the model in addressing challenging tokens and incrementally improve performance on difficult questions.
• Quiet-STaR’s zero-shot learning shows gains on benchmarks such as GSM8K and CommonsenseQA without task-specific fine-tuning.

Quiet-STaR’s self-teaching capabilities mark an advancement in teaching LMs to reason, suggesting a future where AI could autonomously learn and apply rational thought processes.