Learning from Partial Chain-of-Thought via Truncated-Reasoning Self-Distillation

Researchers introduce Truncated-Reasoning Self-Distillation (TRSD), a post-training method that enables AI language models to maintain accuracy while using shorter reasoning traces. The technique reduces computational costs by training models to produce correct answers from partial reasoning, achieving significant inference-time efficiency gains without sacrificing performance.

• →TRSD is a lightweight post-training procedure that improves model efficiency by teaching them to reason with partial information.
• →The method uses a teacher-student architecture where the student learns to match full reasoning performance using only truncated traces.
• →Models trained with TRSD naturally generate shorter reasoning traces even without forced truncation, reducing inference costs.
• →The technique demonstrates improved robustness across multiple reasoning benchmarks and various token budgets.
• →TRSD addresses the computational overhead problem of chain-of-thought reasoning without significant accuracy tradeoffs.