RefuteBench 2.0 -- Agentic Benchmark for Dynamic Evaluation of LLM Responses to Refutation Instruction

2025-02-25

AI Summary

Existing evaluations of large language models (LLMs) lack systematic assessment of response quality, memory consistency, and long-term robustness under multi-turn user rebuttals. Method: We introduce the first dynamic benchmark dedicated to multi-turn rebuttal evaluation, proposing an agent-driven paradigm where LLM agents serve as both rebuttal generators and evaluators. Our framework supports both transient and persistent rebuttal instructions and enhances evaluation reliability via attention analysis and meta-evaluation alignment. Results: Experiments reveal that while mainstream models respond promptly to rebuttals, they suffer from long-range information forgetting and progressive performance degradation. Our generated rebuttals exhibit higher human-likeness, and automated evaluation scores correlate strongly with human annotations (Spearman’s ρ > 0.85). This work establishes a novel benchmark and interpretable analytical toolkit for advancing the reliability of interactive LLMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness
• Multiagent Systems: Adversarial Agents

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

In the multi-turn interaction schema, large language models (LLMs) can leverage user feedback to enhance the quality and relevance of their responses. However, evaluating an LLM's ability to incorporate user refutation feedback is crucial yet challenging. In this study, we introduce RefuteBench 2.0, which significantly extends the original RefuteBench by incorporating LLM agents as refuters and evaluators, which allows for flexible and comprehensive assessment. We design both transient and persistent refutation instructions with different validity periods. Meta-evaluation shows that the LLM-based refuter could generate more human-like refutations and the evaluators could assign scores with high correlation with humans. Experimental results of various LLMs show that current models could effectively satisfy the refutation but fail to memorize the refutation information. Interestingly, we also observe that the performance of the initial task decreases as the refutations increase. Analysis of the attention scores further shows a potential weakness of current LLMs: they struggle to retain and correctly use previous information during long context dialogues. https://github.com/ElliottYan/RefuteBench-2.0

Problem

• Evaluate LLM response to refutation
• Assess LLM memory of refutation
• Analyze LLM performance in dialogues

Innovation

• LLM agents as refuters
• Transient and persistent instructions
• Attention score analysis