Research map / Large language models / Retrieval and knowledge
BiDeV: Bilateral Defusing Verification for Complex Claim Fact-Checking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BiDeV, a role-played multi-LLM workflow that simplifies complex claims (perceive-then-rewrite, decompose-then-check) and filters redundant evidence for claim fact-checking on Hover and Feverous-s.
Problem
Complex claim fact-checking struggles with claim vagueness (latent information and complex relations) and evidence redundancy that complicates verification.
Approach
Mimics human-expert fact-checking with role-played LLMs: a Vagueness Defusing module iteratively perceives latent information, queries background knowledge, rewrites the claim, and decomposes it to resolve referential and comparative relations; a Redundancy Defusing module filters evidence paragraphs for relevance.
Key results
- BiDeV reaches 77.59 Macro-F1 on Hover 2-hop gold and 92.39 on Feverous-s gold, the best among compared baselines.
- BiDeV improves Macro-F1 on Hover 4-hop gold to 70.63 versus 66.62 for FactCheck-GPT (best baseline).
- Under close-setting (no evidence), BiDeV reaches 61.12 on Feverous-s versus 60.13 for Direct prompting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BiDeV instead of the strongest reported baseline | Complex claim fact-checking · Hover hop2 gold | Macro-F1: 77.22 -> 77.59 | Helped |
| Use BiDeV instead of the strongest reported baseline | Complex claim fact-checking · Hover hop2 open | Macro-F1: 70.25 -> 73.44 | Helped |
| Use BiDeV instead of the strongest reported baseline | Complex claim fact-checking · Hover hop3 gold | Macro-F1: 66.32 -> 69.91 | Helped |
| Use BiDeV instead of the strongest reported baseline | Complex claim fact-checking · Hover hop3 open | Macro-F1: 61.21 -> 63.62 | Helped |
| Use BiDeV instead of the strongest reported baseline | Complex claim fact-checking · Hover hop4 gold | Macro-F1: 66.62 -> 70.63 | Helped |
| Use BiDeV instead of the strongest reported baseline | Complex claim fact-checking · Hover hop4 open | Macro-F1: 59.25 -> 60.41 | Helped |
| Use BiDeV instead of the strongest reported baseline | Complex claim fact-checking · Feverous-s gold | Macro-F1: 91.98 -> 92.39 | Helped |
| Use BiDeV instead of the strongest reported baseline | Complex claim fact-checking · Feverous-s open | Macro-F1: 67.24 -> 69.01 | Helped |
| Add complex-relation-oriented decomposition rather than direct verification (ablation) | BiDeV · Hover hop2/hop3/hop4 and Feverous-s, gold and open | Macro-F1: Gold: 71.55 -> 77.59; 63.23 -> 69.91; 63.36 -> 70.63; 89.31 -> 92.39. Open: 66.26 -> 73.44; 59.23 -> 63.62; 57.17 -> 60. | Helped |
| Add complex-relation-oriented decomposition rather than naive decomposition (ablation) | BiDeV · Hover hop2/hop3/hop4 and Feverous-s, gold and open | Macro-F1: Gold: 73.61 -> 77.59; 65.34 -> 69.91; 65.57 -> 70.63; 90.15 -> 92.39. Open: 66.85 -> 73.44; 60.05 -> 63.62; 57.65 -> 60. | Helped |
| Add a perceptor that identifies latent information and generates questions (ablation) | BiDeV · Hover and Feverous-s, gold | Macro-F1 | Helped |
| Add a rewriter that substitutes explicit background knowledge into the claim (ablation) | BiDeV · Hover and Feverous-s, gold | Macro-F1 | Helped |
Surprising findings
- A 250M-parameter FLAN-T5 used as querier/checker with BiDeV reportedly surpasses ProgramFC with an 11B sub-task solver.
Limitations
- Relies on gpt-3.5-turbo for several roles, so results depend on a closed API model.
- Evaluated only on two fact-checking benchmarks (Hover, Feverous-s).
Experimental setup
Datasets: Hover Feverous-s
Benchmarks: Hover Feverous-s
Metrics: Macro-F1
Baselines: BERT-FC LisT5 RoBERTa-NLI DeBERTaV3-NLI MULTIVERS Codex FLAN-T5 HiSS FOLK ProgramFC FactCheck-GPT
Builds on: ProgramFC FOLK HiSS
Closest papers by meaning
- DEFAME: Dynamic Evidence-based FAct-checking with Multimodal Experts (ICML 2025 · similarity 0.88)
- Beyond Text: Fine-Grained Multi-Modal Fact Verification with Hypergraph Transformers (AAAI 2025 · similarity 0.85)
- Heterogeneous Graph Reasoning for Fact Checking over Texts and Tables (AAAI 2024 · similarity 0.85)
- MRR-FV: Unlocking Complex Fact Verification with Multi-Hop Retrieval and Reasoning (AAAI 2025 · similarity 0.84)
- Causal Walk: Debiasing Multi-Hop Fact Verification with Front-Door Adjustment (AAAI 2024 · similarity 0.83)
- DECEIVE-AFC: Adversarial Claim Attacks against Search-Enabled LLM-based Fact-Checking Systems (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.