STaR: Bootstrapping Reasoning With Reasoning
What this work does
Iteratively generates rationales and fine-tunes on correct answers; rationalization conditions on known answers to recover additional training examples.
Connection to discovery
A reasoning-training comparison for success-filtered self-training, especially when contrasting known-answer tasks with discovery lacking answer labels.
Evidence
Inspected the bootstrapping setup and arithmetic/commonsense evaluation, including discussion of judging rationale quality.
Scope and limitations
Known training answers are required for filtering/rationalization. Correct final answers do not guarantee faithful reasoning or valid experiment choice.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.