Entity · paper

Evidence Attribution in Visual Document Understanding without Coordinates or Region Labels

paperactiveprovisionalevidence-attribution-in-visual-document-understanding-without-coordinates-or-region-labels-7cc510aa·1 events·first seen 3d ago

Aliases: Evidence Attribution in Visual Document Understanding without Coordinates or Region Labels

Co-occurring entities

GRPO CiteVQA

More like this (12)

Visual Attribution Distillation Document Visual Question Answering Where Does the Answer Come From? Benchmarking View-Level Visual Evidence Identification in Multi-View MLLMs for Autonomous Driving Native Active Perception as Reasoning for Omni-Modal Understanding How Robust is OCR-Reasoning? Evaluating OCR-Reasoning Robustness of Vision-Language Models under Visual Perturbations Watch, Remember, Reason: Human-View Video Understanding with MLLMs Show Me How You Reason and I'll Tell You Who You Are: Reasoning Graphs for Robust LLM Authorship Attribution Beyond Independent Labels: Schwartz-Geometry Decoding for Human Value Detection Towards Explainable Adjudicative Variance: Quantifying Judicial Discretion via Gated Multi-Task Learning Who Needs Labels? Adapting Vision Foundation Models With the Metadata You Already Have Training-Free Semantic Correction for Autoregressive Visual Models temporal visual event understanding

Recent events (1)

6arXiv · cs.CL·3d ago·source ↗

Quote-and-retrieve pipeline reduces attribution hallucination in visual document understanding without coordinate supervision

Researchers present a study showing that replacing coordinate-based evidence attribution with a language interface—where vision-language models quote evidence verbatim and a multimodal retriever locates the quote—substantially reduces attribution hallucination in visual document understanding. Across six open VLMs on a bilingual CiteVQA subset, evidence recall rises from at most 8 points to 26–47 points and hallucination rate roughly halves. The authors extend this into a GRPO-based training recipe that uses retrieved region crops as reward signal, lifting an 8B model's strict attributed accuracy from 22.4 to 33.8 without any region-level labels. The work offers a practical path to grounded document QA without expensive spatial annotation.

Evaluation and Benchmarking Alignment and RLHF Evidence Attribution in Visual Document Understanding without Coordinates or Region Labels GRPO CiteVQA +1 more