Scaling Reasoning Hop Exposes Weaknesses: Demystifying and Improving Hop Generalization in Large Language Models

TL;DR

This paper investigates why large language models struggle with reasoning steps beyond their training distribution, identifying internal attention head conflicts as a key factor, and proposes a method to improve reasoning generalization.

Contribution

It uncovers the role of erroneous attention heads in reasoning failures and introduces a test-time correction technique to deactivate these heads, enhancing reasoning hop generalization.

Findings

Removing erroneous processing heads improves reasoning accuracy.

Test-time correction consistently enhances reasoning hop generalization.

Internal competition among attention heads causes reasoning errors.

Abstract

Chain-of-thought (CoT) reasoning has become the standard paradigm for enabling Large Language Models (LLMs) to solve complex problems. However, recent studies reveal a sharp performance drop in reasoning hop generalization scenarios, where the required number of reasoning steps exceeds training distributions while the underlying algorithm remains unchanged. The internal mechanisms driving this failure remain poorly understood. In this work, we conduct a systematic study on tasks from multiple domains, and find that errors concentrate at token positions of a few critical error types, rather than being uniformly distributed. Closer inspection reveals that these token-level erroneous predictions stem from internal competition mechanisms: certain attention heads, termed erroneous processing heads (ep heads), tip the balance by amplifying incorrect reasoning trajectories while suppressing correct ones. Notably, removing individual ep heads during inference can often restore the correct predictions. Motivated by these insights, we propose test-time correction of reasoning, a lightweight intervention method that dynamically identifies and deactivates ep heads in the reasoning process. Extensive experiments across different tasks and LLMs show that it consistently improves reasoning hop generalization, highlighting both its effectiveness and potential.