Research map / Large language models / Retrieval and knowledge
MRR-FV: Unlocking Complex Fact Verification with Multi-Hop Retrieval and Reasoning
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 MRR-FV, a generative multi-hop evidence retriever (BART-based, with an LLM-distilled query compressor and trie-constrained decoding) coupled with a hierarchical reasoner (recurrent memory transformer plus graph reasoning) for fact verification on FEVER and HOVER.
Problem
Fact verification on complex claims needing multiple pieces of evidence; one-hop retrievers miss interrelated evidence, and iterative multi-hop retrieval suffers from intention shift as queries grow with accumulated evidence.
Approach
Generate next-hop evidence autoregressively with a constrained (trie-based) BART retriever over a compressed query, where the compressor is pre-trained by distilling from an LLM and aligned with InfoNCE; verification is done with a hierarchical reasoner combining recurrent memory transformer (intra-sentence) and latent-edge graph attention (inter-sentence), jointly trained with retrieval.
Key results
- On FEVER test, MRR-FV reaches 80.83 LA and 78.25 FEVER score, above the best listed baseline Proofver (79.47 / 76.82).
- On HOVER dev, MRR-FV scores 71.66 / 65.09 / 62.72 Macro-F1 for 2/3/4-hop subsets, outperforming ProgramFC (70.30 / 63.43 / 57.74) on each.
- On FEVER dev evidence retrieval, MRR-FV reaches 71.54 F1 versus 37.69 for ESIM and 67.40 for GERE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the query compressor (ablation) | MRR-FV · FEVER test | LA: 80.06 -> 80.83 | Helped |
| add pre-training for the query compressor (ablation) | MRR-FV · FEVER test | LA: 80.21 -> 80.83 | Helped |
| use multi-hop retrieval instead of one-step retrieval (ablation) | MRR-FV · FEVER test | LA: 79.69 -> 80.83 | Helped |
| use the generative retriever instead of dense retrieval (ablation) | MRR-FV · FEVER test | LA: 80.27 -> 80.83 | Helped |
| add recurrent-memory intra-sentence reasoning (ablation) | MRR-FV · FEVER test | LA: 80.43 -> 80.83 | Helped |
| add inter-sentence reasoning (ablation) | MRR-FV · FEVER test | LA: 79.83 -> 80.83 | Helped |
| use MRR-FV rather than the strongest listed baseline | FEVER · FEVER test | FEVER score: 76.82 -> 78.25 | Helped |
| use MRR-FV rather than the strongest listed baseline | HOVER · HOVER 4-hop | Macro-F1: 60.35 -> 62.72 | Helped |
| use a pre-trained query compressor rather than concatenating claim and prior evidence | MRR-FV retrieval · intention-shift similarity, 3-hop | cosine similarity: 0.71 -> 0.79 | Helped |
Limitations
- Hyperparameters such as compression length and hop count are dataset-dependent (sensitivity analysis shows an optimum).
- HOVER evaluated only on the dev set since test sets are not public.
- Compression mitigates but does not eliminate intention shift (cosine similarity drops from 0.85 to 0.79 at 3 hops).
Experimental setup
Datasets: FEVER HOVER
Benchmarks: FEVER HOVER
Metrics: Label Accuracy FEVER score Macro-F1 Precision Recall F1
Baselines: BERT Concat GAT GEAR HESM KGAT DREAM TARSA Proofver EvidenceNet TwoWingOS KGAT+FER CD
Builds on: GERE BART Recurrent Memory Transformer KGAT ProgramFC
Closest papers by meaning
- Enhancing Multi-Hop Fact Verification with Structured Knowledge-Augmented Large Language Models (AAAI 2025 · similarity 0.89)
- Causal Walk: Debiasing Multi-Hop Fact Verification with Front-Door Adjustment (AAAI 2024 · similarity 0.87)
- Beyond Text: Fine-Grained Multi-Modal Fact Verification with Hypergraph Transformers (AAAI 2025 · similarity 0.87)
- Grounding Multi-Hop Reasoning in Structural Causal Models via Group Relative Policy Optimization (ICML 2026 · similarity 0.86)
- CofCA: A STEP-WISE Counterfactual Multi-hop QA benchmark (ICLR 2025 · similarity 0.86)
- Heterogeneous Graph Reasoning for Fact Checking over Texts and Tables (AAAI 2024 · similarity 0.86)
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.