Adversarial Multi-Agent Evaluation of Large Language Models through Iterative Debate
Keywords: LLM Evals, Adversarial analysis, Mechanism Design
Abstract: We propose a novel framework for evaluating large language model (LLM) outputs using LLMs themselves as interacting agents in an adversarial debate system. Our approach casts LLMs as advocates, judges, and juries within a structured courtroom-inspired setting. Advocate LLMs engage in iterative argumentation to refine and critique responses, while judge and jury LLMs moderate and assess the debate. We introduce a probabilistic model using Beta-Binomial distribution to analyze error reduction dynamics in this iterative process. Comparative studies of ranking versus scoring methods for LLM jurors reveal advantages of fine-grained scoring in capturing nuanced quality assessments. Experiments across diverse language tasks demonstrate our framework's superior performance in agreement with human judgments and provision of interpretable feedback compared to traditional evaluation methods. This work contributes a theoretically grounded, scalable approach to LLM evaluation that addresses limitations of existing techniques and adapts to rapid advancements in language AI technologies.
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12847
Loading