STaR (Self-Taught Reasoner) is a method where language models generate their own reasoning traces (chain-of-thought) to solve problems, then fine-tune on their own successful reasoning. This creates a self-improvement loop that enhances reasoning capabilities without external teacher models. STaR is related to on-policy-self-distillation in that both approac...
- Related
- References
STaR (Self-Taught Reasoner)
STaR (Self-Taught Reasoner) is a method where language models generate their own reasoning traces (chain-of-thought) to solve problems, then fine-tune on their own successful reasoning. This creates a self-improvement loop that enhances reasoning capabilities without external teacher models. STaR is related to on-policy-self-distillation in that both approaches leverage the model's own outputs for training, though STaR focuses specifically on reasoning trace generation.
Related
- on-policy-self-distillation — Self-distillation shares the self-improvement paradigm with STaR
- chain-of-thought — STaR generates CoT traces as its core mechanism
- reasoning-distillation — Broader context of transferring reasoning abilities
References
- Zelikman et al. (2022). "STaR: Bootstrapping Reasoning With Reasoning." NeurIPS.