New Theory Proposes Principled Debate Judgement for Agentic AI

Debates have emerged as a valuable methodology for agentic AI, enhancing performance, explainability, and user engagement. AI agents, particularly those powered by Large Language Models (LLMs), can engage in internal self-debates or external debates with other agents. In many such scenarios, the outcomes and resulting outputs of these debates are determined post-hoc by external judges, often themselves LLMs.This research introduces a novel theory of debate judgment applicable to any setting where agents present pros and cons for their opinions. The theory identifies several formal properties crucial for effective debate judgment, including reproducibility, robustness, groundedness, and explainability. The study then formally and experimentally explores how two alternative debate judgment methods satisfy these properties in claim verification settings: variants of the "LLMs as a judge" paradigm and formal semantics drawn from computational argumentation.The findings indicate that while both methods achieve similar accuracy, the "LLMs as a judge" approach may lack the formal guarantees that computational argumentation semantics inherently provide. Overall, the study positions argumentation semantics as an ideal candidate for developing principled and reliable judges in debate-driven AI systems, offering a more robust foundation for determining debate outcomes.