The paper, which has been accepted for presentation at EMNLP 2026, tackles a key challenge faced by long-context visual language models: their potential inaccuracy in pinpointing relevant evidence when dealing with lengthy materials that combine illustrations and text (image-text). To address this, the paper introduces the innovative concept of Multimodal Retrieval Heads (MMRetHeads). Drawing inspiration from QRHead, this concept outlines a method for detecting such heads by closely examining their question-evidence attention patterns.

The study encompasses an analysis of six models under various multi-task and multi-context length conditions. It uncovers that both text and image retrieval tasks share certain attention heads. Through a series of masking experiments, the researchers demonstrate that these retrieval heads play a causal role in the model's capacity to access evidence. Specifically, when these heads are masked, there is a notable decline in performance across retrieval tasks, long document question answering, and multimodal reasoning tasks.

Furthermore, the research team capitalized on the evidence signals emanating from these heads to enhance multimodal document retrieval, all without the need for additional training of a retrieval module. On the MMDocIR benchmark, the Qwen3-VL-8B model exhibited significant improvements, achieving 7.7 and 6.3 percentage point increases in page-level and layout-level Recall@1, respectively, when compared to the most robust baseline.