Post-Hoc Judgement Theory for Argumentative AI: Toward a Principled Referee
This paper addresses the prevalent post-hoc judgement problem in agent debate and proposes and evaluates a general theory of debate evaluation. As LLM-driven agents debate internally or against one another, their outcomes and outputs are often decided afterward by an external referee (typically an LLM). The authors first derive a set of formal properties that debate evaluation should satisfy—covering reproducibility, robustness, groundedness, and interpretability—and then, in the setting of claim verification, combine formalization with experiments to test two alternative approaches: a variant of "LLM-as-a-judge" and the formal semantics originating from computational argumentation. Experiments show that both approaches achieve similar accuracy, but the former lacks the formal guarantees offered by the latter. The study ultimately argues that the formal semantics of computational argumentation is the ideal candidate for a principled referee in argumentative AI.
Background and Context
Large language model-driven agents increasingly settle questions by debating—either within a single agent or across competing agents—as a means to boost accuracy, explainability, and user engagement. But the debate itself rarely decides its own outcome. In many real deployments the final verdict and the resulting output are assigned afterward by an external referee, most often another LLM. This post-hoc judgement step is pervasive yet under-examined, and it is precisely what the arXiv paper targets as its core research problem.
Rather than merely benchmarking which referee performs better, the authors develop a general theory of debate evaluation applicable across settings. The theory applies whenever agents argue the merits of their claims, elevating "how should a debate be judged" from engineering intuition to an analysable, arguable scientific question. The work therefore reframes a routine engineering decision as a formal problem with testable requirements.
Deep Analysis
The authors establish standards before testing methods, deriving four formal properties that debate evaluation should satisfy. Reproducibility demands that identical debates under identical referees yield stable, consistent results. Robustness requires the referee to remain steady against perturbations or rephrasings in the argument. Groundedness insists that verdicts rest on evidence actually present in the debate rather than external information the referee introduces on its own. Interpretability requires the referee to produce convincing reasons for its decision.
These criteria are then examined in the concrete setting of claim verification, where two alternative approaches are compared. The first is a family of variants on "LLM-as-a-judge," in which a large model simply reads both sides and renders a verdict. The second draws on computational argumentation, borrowing formal argumentation semantics to adjudicate. Its defining feature is that decisions follow from a strict set of formal semantic rules, granting provable formal properties.
For each method the authors combine formal analysis with empirical validation, checking whether the properties hold. The experiments show both approaches reach similar accuracy, meaning the formal-semantics referee is not outperformed on raw correctness. The decisive gap lies in assurance: the LLM-as-a-judge variant, though practical and flexible, lacks formal guarantees for groundedness, reproducibility, and interpretability, whereas the computational-argumentation semantics supplies them.
Industry Impact
The study offers a set of design principles for judging argumentative AI as agent systems move from single models toward multi-agent collaboration and self-debate. Deciding debate outcomes fairly, robustly, and explainably becomes central to system reliability. The authors argue that the formal semantics of computational argumentation is the ideal candidate for a principled referee, a judgment with implications for both open-source communities and industrial deployment.
Because formal semantics yield provable properties, the judging process becomes easier to audit, debug, and keep compliant—particularly valuable in high-stakes settings where trust matters. The general property framework also provides a standard coordinate system for evaluating different referee methods, helping the field shift from empirical trial-and-error toward theory-guided design.
Outlook
The central tension the paper surfaces—comparable empirical performance but unequal formal guarantees—suggests future referee design will weigh assurance as heavily as accuracy. Where auditability and defensibility matter, formal semantics from computational argumentation present a stronger foundation than ad hoc LLM verdicts.
The authors ultimately position these formal semantics as the ideal candidate for a principled referee in argumentative AI, framing the work as a response to the open question of how to find a trustworthy judge for AI debate. For researchers focused on LLM interpretability, agent debate, and AI governance, the paper provides an answer combining theoretical depth with practical relevance.
The authors note that the abstract reports no specific numerical metrics, so precise accuracy figures are not asserted here; the core finding remains that the two approaches perform similarly on accuracy while differing sharply on formal guarantees.