Small Models Struggle to Learn from Strong Reasoners

TL;DR

This paper investigates why small language models struggle with complex reasoning and introduces Mix Distillation, a method that combines reasoning examples of varying complexity to enhance small model performance.

Contribution

The paper identifies the Small Model Learnability Gap and proposes Mix Distillation to improve small model reasoning by balancing reasoning complexity during training.

Findings

Mix Distillation improves small model reasoning accuracy.

Small models perform better with shorter, simpler reasoning chains.

Direct distillation from large models is less effective for small models.

Abstract

Large language models (LLMs) excel in complex reasoning tasks, and distilling their reasoning capabilities into smaller models has shown promise. However, we uncover an interesting phenomenon, which we term the Small Model Learnability Gap: small models ($\leq$3B parameters) do not consistently benefit from long chain-of-thought (CoT) reasoning or distillation from larger models. Instead, they perform better when fine-tuned on shorter, simpler reasoning chains that better align with their intrinsic learning capacity. To address this, we propose Mix Distillation, a simple yet effective strategy that balances reasoning complexity by combining long and short CoT examples or reasoning from both larger and smaller models. Our experiments demonstrate that Mix Distillation significantly improves small model reasoning performance compared to training on either data alone. These findings highlight the limitations of direct strong model distillation and underscore the importance of adapting reasoning complexity for effective reasoning capability transfer.