The Confident Liar: Diagnosing Multi-Agent Debate with Log-Probabilities and LLM-as-Judge

A new study of multi-agent debate systems finds that internal confidence signals can detect flawed reasoning before a final answer is given, but the reliability of those signals depends heavily on which agent is speaking [1]. The paper, posted to the arXiv preprint repository on June 9, examines a two-agent debate architecture consisting of a Constructor and an Auditor [1]. An LLM-as-judge scores each agent's reasoning along three dimensions — instruction following, justification quality, and evidence grounding — and also assigns a critical-failure flag when reasoning breaks down [2]. The researchers then compare those rubric scores against token-level log-probability distributions, which serve as a measure of the model's internal confidence [2]. Experiments in the rubric-scoring domain revealed a consistent four-phase confidence trajectory and a substantial role asymmetry [2]. Confidence aligned with externally judged reasoning quality roughly twice as strongly for the Constructor as for the Auditor [2]. The Constructor's confidence-based detection of critical reasoning failures reached an AUROC of 0.804, compared with 0.634 for the Auditor [2]. The work is part of a broader investigation spanning three domains: rubric-based scoring, mathematical reasoning, and factual question answering [2]. Large language models, which underpin these debate systems, are typically based on transformer architectures and are pre-trained to predict the next word before being fine-tuned for instruction-following [8]. Benchmark evaluations for such models attempt to measure reasoning, factual accuracy, alignment, and safety [8]. The paper appears on arXiv, an open-access repository that hosts electronic preprints across mathematics, physics, computer science, and other fields [6]. As of November 2024, the repository receives about 24,000 submissions per month and has surpassed two million total articles [6]. The study's abstract page includes experimental community tools developed through arXivLabs, a framework that allows third-party collaborators to build features such as citation explorers and code finders directly on the site [4][5]. arXivLabs was formalized in 2020 to enable community innovation while requiring partners to adhere to values of openness, community, excellence, and user data privacy [5].