DIVER: Enhancing Complex Fact Verification via Dynamic Evidence Retrieval and Iterative Reasoning

ACL ARR 2025 July Submission1450 Authors

Abstract:

Fact‑verification tasks involving sequences of claims remain challenging due to high claim density, low accuracy in open-domain evidence retrieval, and multi-hop reasoning requirements, which are difficult to address using traditional methods. In this paper, we propose DIVER (Dynamic and Iterative fact VERification), a fact verification framework that decomposes paragraphs into context-independent sentences and applies a dynamic and iterative claim extraction and evidence retrieval strategy. Unlike prior one‑shot or list‑style approaches, DIVER introduces a fine-grained iterative claim extraction mechanism, allowing the system to better capture verifiable atomic claims, and incorporates a novel evidence-filtering and query recommendation module to robustly handle insufficient or ambiguous evidence, significantly enhancing multi-hop reasoning capabilities.
Additionally, we propose a heuristic-driven revision step to detect long-distance contextual errors overlooked by previous approaches. These mechanisms collectively improve the model's calibration, ensuring the verifier fires only when supported by sufficient evidence—an essential property for dependable fact checking. Experimental results on three widely-used challenging fact-checking benchmarks (FEVEROUS, LIAR, and AVeriTeC) demonstrate that DIVER substantially outperforms existing LLM-based approaches and pipelines.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: fact checking

Languages Studied: English

Submission Number: 1450