Atomic Contrastive Verification: Fine-Grained Fact-Checking via Claim Decomposition and Knowledge Graph-Grounded Contrastive Reasoning

Large language models (LLMs) frequently produce text that is fluent yet factually inconsistent with source documents. Detecting such inconsistency remains challenging, particularly when errors involve subtle entity substitutions, temporal distortions, or relational misattributions embedded within lengthy outputs. We propose Atomic Contrastive Verification (ACV), a training-free, graph-grounded fact-checking framework that decomposes both generated claims and source documents into atomic claims—minimal, self-contained factual units—and performs structured contrastive reasoning over each unit independently. For each atomic claim, ACV extracts a knowledge graph triple and generates contrastive claim variants through a multi-type perturbation taxonomy covering entity, relation, temporal, and quantitative dimensions. A novel Knowledge-Weighted Contrastive MMR mechanism, integrating graph-structural centrality and NLI-based logical diversity, selects the most discriminative subset of variants. Each selected variant is then pairwise compared against the claim; the resulting comparison responses are summarized to produce a per-claim verdict, and per-claim verdicts are aggregated into a document-level judgment. Experiments on the LLM-AggreFact benchmark (eleven subsets) demonstrate that ACV achieves competitive or superior performance compared to both specialized fine-tuned fact-checkers and large-scale LLMs. Beyond accuracy, ACV provides interpretable, claim-level error localization that existing methods cannot offer.