Epistemic Filtering and Collective Hallucination: A Jury Theorem for Confidence-Calibrated Agents

Researchers propose a new framework for collective decision-making where AI agents can abstain from voting when uncertain, extending the Condorcet Jury Theorem to confidence-gated settings. The study shows this selective participation approach can improve group accuracy and potentially reduce hallucinations in large language model systems.

• →Framework allows heterogeneous agents to learn their own reliability and selectively abstain from voting when uncertain.
• →Research extends classical Condorcet Jury Theorem to sequential, confidence-gated decision-making scenarios.
• →Non-asymptotic lower bounds derived for group success probability with selective participation.
• →Monte Carlo simulations validate the theoretical bounds for confidence-calibrated voting systems.
• →Potential applications include mitigating hallucinations in collective LLM decision-making for AI safety.