What Happened in LLMs Layers when Trained for Fast vs. Slow Thinking: A Gradient Perspective

TL;DR

This paper analyzes how fast and slow thinking training methods affect layer-wise gradients in large language models, revealing differences in stability and the ability to reflect response correctness, with implications for building more reliable AI systems.

Contribution

It provides the first gradient-based analysis of fast versus slow thinking training in LLMs, highlighting their impact on learning stability and response correctness reflection.

Findings

Fast thinking leads to larger gradients and instability.

Slow thinking improves learning stability and correctness detection.

Gradient patterns differ significantly between reasoning and non-reasoning tasks.

Abstract

What makes a difference in the post-training of LLMs? We investigate the training patterns of different layers in large language models (LLMs) through the lens of the gradient. We are specifically interested in how fast vs. slow thinking affects the layer-wise gradients, given the recent popularity of training LLMs on reasoning paths such as chain-of-thoughts (CoT) and process rewards. In our study, fast thinking without CoT leads to larger gradients and larger differences of gradients across layers than slow thinking (Detailed CoT), indicating the learning stability brought by the latter. Additionally, we study whether the gradient patterns can reflect the correctness of responses when training different LLMs using slow vs. fast thinking paths. The results show that the gradients of slow thinking can distinguish correct and irrelevant reasoning paths. As a comparison, we conduct similar gradient analyses on non-reasoning knowledge learning tasks, on which, however, trivially increasing the response length does not lead to similar behaviors of slow thinking. Our study strengthens fundamental understandings of LLM training and sheds novel insights on its efficiency and stability, which pave the way towards building a generalizable System-2 agent. Our code, data, and gradient statistics can be found in: https://github.com/MingLiiii/Layer_Gradient.