Back to AI Research

AI Research

A Theory of Post-hoc Debate Judgement | AI Research

Key Takeaways

  • A Theory of Post-hoc Debate Judgement proposes a formal framework to evaluate how AI agents should resolve debates.
  • Debates have recently emerged as a useful methodology for agentic AI to improve performance as well as to aid explainability and user engagement.
  • For example, LLM-empowered agents may debate internally (with themselves) and/or externally (with other agents).
  • In many settings where debates are used, debates' outcomes and resulting outputs are determined post-hoc by external judges, often LLMs.
  • In this paper we develop and test a novel theory of debate judgement applicable to all settings where agents engage in debates by providing pros and cons for their opinions therein.
Paper AbstractExpand

Debates have recently emerged as a useful methodology for agentic AI to improve performance as well as to aid explainability and user engagement. For example, LLM-empowered agents may debate internally (with themselves) and/or externally (with other agents). In many settings where debates are used, debates' outcomes and resulting outputs are determined post-hoc by external judges, often LLMs. In this paper we develop and test a novel theory of debate judgement applicable to all settings where agents engage in debates by providing pros and cons for their opinions therein. Specifically, we identify a number of formal properties that debate judgement may be required to satisfy in general, as concerns reproducibility, robustness, groundedness and explainability. Then, we explore their satisfaction formally and/or experimentally, for claim verification settings, for two specific alternative debate judgement methods: variants of the LLMs as a judge idea and formal semantics drawn from computational argumentation. We show that the two methods give similar accuracy performances but the former may lack formal guarantees that the latter brings. Overall, our study indicates argumentation semantics as an ideal candidate for principled judges in debate-driven AI.

A Theory of Post-hoc Debate Judgement proposes a formal framework to evaluate how AI agents should resolve debates. When AI agents debate to verify claims, their final outcomes are often determined by an external "judge," typically another LLM. This paper introduces a set of formal properties to assess these judges, ensuring they are reliable, reproducible, and grounded in the arguments provided during the debate.

Formal Properties for Debate Judges

The authors identify seven properties to guide the selection of a debate judge:

  • Determinism: The judge provides the same output for the same input and opinion profile.

  • Permutation Independence: The order in which agents present their opinions does not change the final verdict.

  • Profile Robustness: Small changes in the agents' opinions do not lead to large changes in the judge's stance.

  • Judge Robustness: Minor variations in the judging method itself (such as prompt changes) do not significantly alter the outcome.

  • Non-hallucination: The judge selects a stance already proposed by at least one agent rather than introducing a new one.

  • Judge Unanimity: If all agents agree on a stance, the judge must adopt that consensus.

  • Contestability: The judge responds consistently to changes in the debate; if the evidence becomes more supportive of a claim, the judge’s verdict should reflect that shift.

Comparing Judging Methods

The researchers compared two primary methods for judging debates: "LLMs as a judge" (using GPT-4o) and "semantics as a judge" (using formal computational argumentation).
In the experiments, which focused on claim verification, both methods achieved similar levels of accuracy. However, the authors found that formal argumentation semantics—specifically the DF-QuAD framework—offered better satisfaction of the formal properties defined in the study. While LLM-based judges are common, they often lack the formal guarantees provided by argumentation semantics, which can be critical in high-stakes environments where explainability and consistency are required.

The Role of Argumentation Semantics

The study utilizes Quantitative Bipolar Argumentation Frameworks (QBAFs) to model debates. In this structure, arguments are represented as nodes with attack or support relations. A base score is assigned to each argument, and the final strength of a claim is calculated using aggregation functions. This approach allows the judge to mathematically derive a conclusion based on the dialectical strength of the pros and cons presented by the agents, rather than relying on the probabilistic output of an LLM.

Considerations for Implementation

The authors note that these properties are not necessarily universal requirements for every scenario. For instance, while permutation independence is often desirable to prevent bias, it may be disadvantageous if the judge needs to weigh the expertise of specific agents. Similarly, while non-hallucination prevents the introduction of new stances, it may be problematic if all participating agents are collectively incorrect. The researchers suggest that these properties serve as a toolkit for developers to select the most appropriate judging method based on the specific needs of their AI debate system.

Comments (0)

No comments yet

Be the first to share your thoughts!