Can Large Language Models Detect Errors in Long Chain-of-Thought Reasoning?

TL;DR

This paper introduces DeltaBench, a benchmark for evaluating Large Language Models' ability to detect errors in long Chain-of-Thought reasoning, analyzing different models and critic systems to understand their effectiveness and limitations.

Contribution

The paper presents DeltaBench, a new benchmark dataset for assessing error detection in long CoT reasoning by LLMs, along with comprehensive analysis of model performance.

Findings

Different o1-like models vary in effectiveness for long CoT generation.

Existing critic models have limitations in error detection accuracy.

DeltaBench provides insights for improving LLM reasoning and critique abilities.

Abstract

Recently, o1-like models have drawn significant attention, where these models produce the long Chain-of-Thought (CoT) reasoning steps to improve the reasoning abilities of existing Large Language Models (LLMs). In this paper, to understand the qualities of these long CoTs and measure the critique abilities of existing LLMs on these long CoTs, we introduce the DeltaBench, including the generated long CoTs from different o1-like models (e.g., QwQ, DeepSeek-R1) for different reasoning tasks (e.g., Math, Code, General Reasoning), to measure the ability to detect errors in long CoT reasoning. Based on DeltaBench, we first perform fine-grained analysis of the generated long CoTs to discover the effectiveness and efficiency of different o1-like models. Then, we conduct extensive evaluations of existing process reward models (PRMs) and critic models to detect the errors of each annotated process, which aims to investigate the boundaries and limitations of existing PRMs and critic models. Finally, we hope that DeltaBench could guide developers to better understand the long CoT reasoning abilities of their models.