SEVD: Mitigating Object Hallucinations in Multimodal Large Language Models via Semantic Evidence Verification Decoding

Multimodal large language models (MLLMs) have demonstrated strong capabilities in visual understanding and open-ended generation, yet they remain prone to hallucination, where generated content is inconsistent with or unsupported by the visual input. Existing training-free decoding methods mainly intervene through token-level generation scores or output distributions, while the semantic content represented by candidate tokens is not explicitly verified. We propose Semantic Evidence Verification Decoding (SEVD), a training-free decoding method that elevates hallucination mitigation from token-level intervention to semantic-level verification. SEVD first constructs explicit semantic units from candidate tokens, performs semantic evidence verification by comparing how the model's support for the same semantic unit changes when visual evidence is preserved, weakened, or removed, and finally maps the resulting outcomes back to decoding decisions through selective intervention. Extensive experiments across different MLLM backbones and benchmarks demonstrate substantial hallucination mitigation while preserving general multimodal capability, supporting the effectiveness and transferability of SEVD. Code repository: https://github.com/sielan479-wq/SEVD