Attention Never Lie: Visual Attention Defocus Reveals and Rectifies Hallucinations in MLLMs

Multimodal Large Language Models (MLLMs) frequently suffer from the hallucination problem. Existing work is mostly language-centric, with little analysis of the causes of hallucinations from a visual perspective, and has yet to establish a framework for hallucination detection and mitigation. In this work, we reveal the phenomenon of visual attention defocus in MLLMs, which is the main cause of hallucinations generated by the model. Building on this, we further verified the phenomenon. Statistical analysis of many samples showed that the eigenvalues of the visual attention covariance matrix are often lower during hallucinations. Additionally, we observed that the changes in visual attention are relatively small when generating adjacent text. This inspired us to use a sliding window mechanism based on the distribution of attention and a simple MLP to detect hallucinations in real time. Combined with different sampling methods, we select the Top-1 candidate text based on the hallucination detection results within each window for the next round of generation. Through extensive and comprehensive experiments, we demonstrate the effectiveness and adaptability of our method, showing significant improvements compared to the baseline. The code is available at https://github. com/KlingAIResearch/Attention-Never-Lie.