MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation

TL;DR

This work investigates why multimodal LLMs hallucinate, showing that before final decoding the model can recognize objects yet prior-guided biases suppress these facts. It introduces DeCo, a training-free dynamic decoding framework that selects an anchor preceding layer and corrects the final logits with anchor-layer knowledge and dynamic modulation. DeCo is model-agnostic and integrates with common decoding strategies, improving hallucination suppression across image captioning and VQA benchmarks while incurring modest latency overhead. The results indicate that leveraging preceding-layer information is an effective and practical pathway to reduce hallucinations in MLLMs, with strong empirical support and analysis of robustness and hyperparameters.

Abstract

Multimodal Large Language Models (MLLMs) frequently exhibit hallucination phenomena, but the underlying reasons remain poorly understood. In this paper, we present an empirical analysis and find that, although MLLMs incorrectly generate the objects in the final output, they are actually able to recognize visual objects in the preceding layers. We speculate that this may be due to the strong knowledge priors of the language model suppressing the visual information, leading to hallucinations. Motivated by this, we propose a novel dynamic correction decoding method for MLLMs DeCo, which adaptively selects the appropriate preceding layers and proportionally integrates knowledge into the final layer to adjust the output logits. Note that DeCo is model agnostic and can be seamlessly incorporated with various classic decoding strategies and applied to different MLLMs. We evaluate DeCo on widely-used benchmarks, demonstrating that it can reduce hallucination rates by a large margin compared to baselines, highlighting its potential to mitigate hallucinations. Code is available at https://github.com/zjunlp/DeCo.

Figures (13)

• Figure 1: Overall results of the probing experiment with MLLMs, indicating that they possess a certain level of awareness regarding the presence of visual objects (Figure \ref{['fig:arc1']}), with prediction accuracy being higher in the preceding layers (Figure \ref{['fig:arc2']}) but gradually decline afterward.
• Figure 2: Illustration of token probabilities across transformer layers, which reveals distinct trends for target hallucinated (orange) and non-hallucinated (green) tokens. In the preceding layers, non-hallucinated tokens exhibit a higher probability. In the final layers, hallucinated tokens demonstrate increased probabilities, while the probability of non-hallucinated tokens drops sharply.
• Figure 3: Distribution of activated ground-truth tokens across layers.
• Figure 4: Framework of DeCo. DeCo first dynamically selects an appropriate anchor layer from the preceding layers and then correct the knowledge in the final layer with dynamic coefficient.
• Figure 5: DeCo generally improves the MLLM's performance.