HAMV: A Heterogeneous Adaptive Multi-Model Verification Framework for Efficient and Reliable Fact-Checking

ACL ARR 2026 January Submission3250 Authors

Keywords: Hallucination Mitigation, Fact-Checking, Multi-Model Collaboration, Mixture-of-Experts, Efficient LLM Verification

Abstract: Large language models often exhibit hallucinations, while single-model self-verification lacks the knowledge and reasoning diversity required to achieve both reliability and cost-efficiency. We propose HAMV, a framework that reframes multi-model collaboration as a cross-model sparse expert scheduling problem. Inspired by the Mixture-of-Experts (MoE) paradigm, HAMV employs a dynamic routing mechanism to adaptively assign roles (generation, evaluation, verification, and aggregation) based on task and model profiles. It incorporates Dempster–Shafer-based confidence fusion to trigger conditional verification for controlled computational cost. On HalluQA and TruthfulQA, HAMV consistently outperforms representative baselines across varying budgets. Further analysis confirms that dynamic scheduling mitigates position sensitivity, validating the effectiveness of modeling factual verification as a structured scheduling problem.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: fact checking, misinformation detection

Contribution Types: NLP engineering experiment, Approaches low compute settings-efficiency

Languages Studied: Chinese, English

Submission Number: 3250