Claim level graph reasoning for training free LLM hallucination mitigation

Abstract

Large language models (LLMs) often generate fluent yet unsupported or contradicted claims, and such hallucinations become particularly harmful when expressed with high confidence. Existing mitigation methods typically verify responses at the answer level or rely on support-oriented retrieval, making it difficult to distinguish heterogeneous factual risks across atomic claims and to reject superficially relevant evidence. We propose FreeGuard, a training-free hallucination mitigation framework that formulates response governance as claim-level graph reasoning. FreeGuard decomposes an initial response into atomic claims, propagates hallucination risk over a typed claim graph initialized by semantic-verbal uncertainty mismatch, routes high-risk claims to bidirectional evidence verification, and applies counterfactual evidence-chain testing to identify fragile support. Experiments on long-form factuality, retrieval-augmented hallucination, real-world verification, and scientific claim verification benchmarks show consistent improvements in hallucination detection and mitigation. Averaged across five benchmarks, FreeGuard improves AUROC from 83.1 to 85.5 over the strongest recent detection baseline, reduces overall hallucination rate from 32.0% to 16.2%, reduces confident hallucination rate from 25.9% to 9.8%, and improves correctness rate from 63.2% to 70.2%. These results support claim-level risk propagation, bidirectional evidence reasoning, and counterfactual evidence discrimination as complementary components of post-hoc hallucination governance.