Q-CueGraph: Query-Conditioned Visual Evidence Graphs for Multimodal Reasoning
A novel AI system named Q-CueGraph, as described in a paper on arXiv, improves multimodal large language models (LLMs) by offering a task-specific policy for image inspection. This system translates a question and an image representation into budgeted, coordinate-level observations for a static reader model. For images rich in text, it employs a reusable OCR/layout graph, while for natural-image searches, it creates query-conditioned visual nodes within the same selection, composition, and budgeting framework. An optional utility refinement identifies which candidate crops the static reader can utilize based on training-answer accuracy, without needing region-box supervision. Utilizing a frozen Qwen2.5-VL-7B reader, Q-CueGraph achieves 0.833 accuracy on V*Bench compared to 0.696 for full-image inference with a 19% image-area budget, and attains 92% of full-image ANLS on InfographicVQA using about half the image area. The paper can be found on arXiv with the identifier 2608.04452.
Key facts
- Q-CueGraph is a query-conditioned visual evidence graph system for multimodal reasoning.
- It provides a task-conditioned policy for deciding where to inspect an image.
- It maps a question and image representation to budgeted, coordinate-level observations.
- It uses a reusable OCR/layout graph for text-rich images.
- It instantiates query-conditioned visual nodes for natural-image search.
- Optional utility refinement learns from training-answer correctness without region-box supervision.
- With a frozen Qwen2.5-VL-7B reader, it achieves 0.833 accuracy on V*Bench vs 0.696 for full-image inference from a 19% image-area budget.
- It reaches 92% of full-image ANLS on InfographicVQA from about half the image area.
Entities
Institutions
- arXiv