CHAL framework turns AI debate into structured belief optimization

Current multi-agent LLM debate approaches suffer from structural limitations: belief trajectories act like martingales, most gains come from majority voting, and models show confidence escalation without true calibration. The researchers argue these methods only work for ground-truth tasks, not for defeasible domains where any position can be overturned by better reasoning. They propose CHAL, a dialectic framework that treats argumentation as structured belief optimization.

CHAL equips each agent with a CHAL Belief Schema (CBS) — a graph-structured representation inspired by Bayesian inference. Belief revision uses a gradient-informed dynamic mechanism, leveraging the thesis's strength as a differentiable objective. Meta-cognitive value systems covering epistemology, logic, and ethics become configurable hyperparameters that govern reasoning and adjudication. Ablation experiments confirm systematic effects: the adjudicator's value system determines latent belief trajectories, council diversity refines collective beliefs, and the framework generalizes across fields. CHAL is the first to treat multi-agent debate as structured belief optimization over defeasible domains, producing auditable artifacts for transparent oversight.