RealCritic: Towards Effectiveness-Driven Evaluation of Language Model Critiques

TL;DR

This paper introduces RealCritic, a benchmark for evaluating the critique capabilities of LLMs through a closed-loop approach involving self-critique, cross-critique, and iterative critique across reasoning tasks.

Contribution

It presents a novel benchmark with a closed-loop evaluation method for assessing LLM critique abilities, highlighting differences between classical and advanced models.

Findings

Classical LLMs lag behind advanced models in critique tasks.

Classical LLMs may underperform in self-critique and iterative critique.

The benchmark distinguishes reasoning capabilities of different LLMs.

Abstract

Critiques are important for enhancing the performance of Large Language Models (LLMs), enabling both self-improvement and constructive feedback for others by identifying flaws and suggesting improvements. However, evaluating the critique capabilities of LLMs presents a significant challenge due to the open-ended nature of the task. In this work, we introduce a new benchmark designed to assess the critique capabilities of LLMs. Unlike existing benchmarks, which typically function in an open-loop fashion, our approach employs a closed-loop methodology that evaluates the quality of corrections generated from critiques. Moreover, the benchmark incorporates features such as self-critique, cross-critique, and iterative critique, which are crucial for distinguishing the abilities of advanced reasoning models from more classical ones. We implement this benchmark using eight challenging reasoning tasks. We have several interesting findings. First, despite demonstrating comparable performance in direct chain-of-thought generation, classical LLMs significantly lag behind the advanced reasoning-based model o1-mini across all critique scenarios. Second, in self-critique and iterative critique settings, classical LLMs may even underperform relative to their baseline capabilities. We hope that this benchmark will serve as a valuable resource to guide future advancements. The code and data are available at \url{https://github.com/tangzhy/RealCritic}.