Concept

Knowledge Distillation for Reasoning

Knowledge distillation for reasoning is a training method where a compact 'student' LLM learns to emulate the capabilities of a larger 'teacher' LLM. The teacher model produces high-quality reasoning demonstrations, and the student is trained to mimic either the final reasoning outputs or the internal representations of the teacher. The objective is to transfer the sophisticated reasoning abilities of the large model to the smaller, more efficient student model, making advanced reasoning more accessible and less computationally intensive during inference.

0

1

Updated 2026-05-06

Contributors are:

Who are from:

Tags

Ch.5 Inference - Foundations of Large Language Models

Foundations of Large Language Models

Foundations of Large Language Models Course

Computing Sciences