Extracting Visual Facts from Intermediate Layers for Mitigating Hallucinations in Multimodal Large Language Models (2507.15652v1)

Abstract: Multimodal LLMs (MLLMs) have made significant strides by combining visual recognition and language understanding to generate content that is both coherent and contextually accurate. However, MLLMs continue to struggle with object hallucinations, where models produce seemingly plausible but factually incorrect outputs, including objects that do not exist in the image. Recent work has revealed that the prior knowledge in MLLMs significantly suppresses visual information in deep layers, causing hallucinatory outputs. However, how these priors suppress visual information at the intermediate layer stage in MLLMs remains unclear. We observe that visual factual knowledge and the differences between intermediate-layer prior/original probability distributions show similar evolutionary trends in intermediate layers. Motivated by this, we introduce Decoding by Extracting Visual Facts (EVA), a simple, training-free method that dynamically selects intermediate layers with the most significant visual factual information. By contrasting the output distributions of the selected layer derived from the original input and pure-text input, EVA extracts visual factual knowledge and proportionally incorporates it into the final layer to correct the output logits. Importantly, EVA is model-agnostic, seamlessly integrates with various classic decoding strategies, and is applicable across different MLLMs. We validate EVA on widely-used benchmarks, and the results show that it significantly reduces hallucination rates compared to baseline methods, underscoring its effectiveness in mitigating hallucinations.