Q-CueGraph: Query-Conditioned Visual Evidence Graphs

Q-CueGraph asks which visual evidence a multimodal model should actually inspect for a given query. Instead of requiring a frozen multimodal large language model to discover relevant evidence implicitly from an entire high-resolution image, it constructs task-conditioned evidence proposals and turns them into explicit coordinate-level observations.

Authors: Pengcheng Pan and Xinfang Zhang

Task-conditioned observation links human goal-directed gaze with explicit visual evidence selection for multimodal reasoning.

Key idea

Query → evidence selection → explicit observation → frozen multimodal reasoning

For text-rich images, Q-CueGraph builds a reusable OCR/layout graph once per image, activates it for each query, expands through structural relationships, and ranks and composes selected regions into an explicit crop for a frozen reader.

For natural images, query-conditioned visual nodes use the same selection, composition, and budgeting interface. The reported experiments use a frozen Qwen2.5-VL-7B reader.

Selected public results

• V*Bench: accuracy improves from 0.696 with full-image inference to 0.833 while observing approximately 19% of the image area.
• InfographicVQA: approximately 92% of full-image ANLS using about half of the image area.

Paper: arXiv:2608.04452

Keywords: multimodal reasoning, visual evidence acquisition, vision-language models, OCR/layout graphs, query-conditioned observation.