Bootstrapped Iterative Self-Reasoning Distillation
Compress LLM reasoning chains by distilling concise, correct traces, then using them to guide further distillation for efficiency and accuracy.
When to use
When needing to reduce the length and computational cost of LLM reasoning chains without losing accuracy, especially for deployment on resource-constrained systems.
How to apply
Generate initial reasoning traces from a base LLM using instructions that encourage brevity.
Filter these traces to keep only those that lead to correct answers.
Use the filtered, concise traces to train a smaller or more efficient model through a guided prompting stage.
Apply on-policy self-distillation using the refined traces to further improve the model's reasoning efficiency.
Glossary
- reasoning traces
- A step-by-step breakdown of how an LLM arrives at an answer, showing the intermediate thoughts or logic used.
- correct answers
- The final output or conclusion reached by the LLM that matches the expected or factual outcome for a given problem.
- guided prompting stage
- A phase where an LLM is prompted using examples or specific instructions derived from high-quality, pre-filtered outputs to learn a desired behavior.
- on-policy self-distillation
- A technique where a model learns from its own outputs (on-policy) by distilling knowledge from its own generated reasoning traces, aiming to improve efficiency or accuracy.
- distillation
- A machine learning process where a smaller model learns to mimic the behavior or knowledge of a larger, more complex model.