CriticBench: Benchmarking LLMs for Critique-Correct Reasoning
CriticBench benchmarks 17 LLMs' critique and correction abilities across five reasoning domains, revealing linear relationships and size-dependent knowledge consistency.
Key Findings
Methodology
This work constructs CriticBench, a comprehensive benchmark covering five reasoning domains: mathematical, commonsense, symbolic, coding, and algorithmic. It integrates 15 datasets and responses from 17 models including LLaMA, Vicuna, and GPT series. The evaluation employs a three-phase GQC framework—generation, critique, and correction—using metrics like accuracy and F1. The study analyzes how model size, training strategies (e.g., RLHF, Critique Training), and prompt design influence performance. Inter-model interactions are also examined to understand capabilities and limitations.
Key Results
- GQC capabilities exhibit a linear relationship; larger models like GPT-4 achieve an average accuracy of 55.06%, outperforming smaller models such as LLaMa-2-7b at 51.01%. Critique-focused training significantly boosts performance, especially in logical tasks like coding, where correction effectiveness improves by 4-5%. Knowledge consistency across generation, critique, and correction improves with model size, with GPT-4 showing the highest coherence.
- Task-dependent differences are prominent: models excel in critiquing and correcting logic-based tasks but struggle with detail-oriented algorithmic questions. Mathematical reasoning shows limited correction gains (~3%), indicating difficulty in refining answers in complex numerical tasks. Inter-model critique reveals that stronger models are better at evaluating weaker ones, yet weaker models sometimes outperform stronger ones in self-critique, highlighting the nuanced dynamics of model interactions.
- The experiments demonstrate that model scale, training strategies, and prompt engineering are crucial for optimizing GQC performance. The findings suggest a potential for self-improving AI systems through structured critique and correction processes, with implications for autonomous learning and reliability enhancement.
Significance
This research provides a systematic, quantitative assessment of LLMs' critical reasoning abilities, addressing a key gap in understanding models' self-evaluation and self-improvement potential. By revealing the relationships between model size, training, and task type, it guides future development of more autonomous, reliable AI systems. The insights into inter-model critique dynamics also open avenues for collaborative model improvement and ensemble strategies, ultimately advancing AI's capacity for self-directed learning.
Technical Contribution
The paper introduces the GQC framework, a novel multi-task evaluation paradigm that captures the interconnectedness of generation, critique, and correction. It develops CriticBench, a multi-domain benchmark with diverse datasets, enabling comprehensive performance analysis. The study models the GQC capabilities as a linear relationship, providing a new theoretical perspective on model scaling. It also uncovers inter-model critique patterns, contributing to the understanding of collaborative AI systems and self-improvement mechanisms.
Novelty
This is the first work to systematically evaluate and model the interdependence of generation, critique, and correction in large language models across multiple domains. The introduction of CriticBench as a unified benchmark and the GQC linear relationship framework represent significant innovations, filling a critical gap in the field of AI self-assessment and autonomous learning. The analysis of inter-model critique interactions further distinguishes this work from prior studies focusing solely on single-task performance.
Limitations
- The evaluation relies on static responses and does not incorporate dynamic, interactive feedback mechanisms, which are essential for real-world self-improvement scenarios.
- Dataset coverage, especially for symbolic and algorithmic reasoning, remains limited, potentially affecting the generalizability of findings.
- Variability in training strategies across models introduces confounding factors, necessitating more controlled experiments to isolate specific influences on GQC performance.
Future Work
Future research should explore dynamic, interactive self-evaluation frameworks, integrating reinforcement learning and meta-learning for continuous improvement. Expanding dataset diversity and including real-world tasks will enhance robustness. Investigating multi-model collaboration and ensemble methods could further improve self-correction accuracy, paving the way for autonomous AI systems capable of ongoing self-optimization.
AI Executive Summary
The rapid development of large language models (LLMs) has revolutionized natural language processing, yet their capacity for self-critique and correction remains underexplored. CriticBench emerges as a pioneering benchmark, systematically evaluating 17 models across five reasoning domains—mathematical, commonsense, symbolic, coding, and algorithmic—using a comprehensive set of 15 datasets. This framework assesses models' abilities in three interconnected stages: generation, critique, and correction (GQC). Results reveal a clear linear relationship among these capabilities, with larger models like GPT-4 achieving an average accuracy of 55.06%, significantly outperforming smaller counterparts.
One key insight is that model size correlates positively with knowledge consistency across tasks, indicating that bigger models maintain more coherent reasoning and self-assessment. Task-dependent differences are also prominent: models excel at critiquing and correcting logic-based tasks such as code generation but struggle with detail-oriented algorithmic questions, where correction gains are limited (~3%). Interestingly, the study uncovers dynamic interactions among models—stronger models tend to critique weaker ones more effectively, while weaker models sometimes surpass stronger ones in self-critique—highlighting complex inter-model dynamics.
These findings have profound implications for AI development. They suggest that scaling models and refining training strategies, especially critique-focused methods, can significantly enhance autonomous self-improvement. The research also points toward future directions, including dynamic feedback systems, multi-model collaboration, and expanding task diversity. CriticBench thus provides a vital tool for advancing AI towards more reliable, self-aware systems capable of ongoing self-optimization, ultimately fostering more autonomous, trustworthy AI solutions.
Deep Dive
Abstract
The ability of Large Language Models (LLMs) to critique and refine their reasoning is crucial for their application in evaluation, feedback provision, and self-improvement. This paper introduces CriticBench, a comprehensive benchmark designed to assess LLMs' abilities to critique and rectify their reasoning across a variety of tasks. CriticBench encompasses five reasoning domains: mathematical, commonsense, symbolic, coding, and algorithmic. It compiles 15 datasets and incorporates responses from three LLM families. Utilizing CriticBench, we evaluate and dissect the performance of 17 LLMs in generation, critique, and correction reasoning, i.e., GQC reasoning. Our findings reveal: (1) a linear relationship in GQC capabilities, with critique-focused training markedly enhancing performance; (2) a task-dependent variation in correction effectiveness, with logic-oriented tasks being more amenable to correction; (3) GQC knowledge inconsistencies that decrease as model size increases; and (4) an intriguing inter-model critiquing dynamic, where stronger models are better at critiquing weaker ones, while weaker models can surprisingly surpass stronger ones in their self-critique. We hope these insights into the nuanced critique-correct reasoning of LLMs will foster further research in LLM critique and self-improvement.