Look as You Think: Unifying Reasoning and Visual Evidence Attribution for Verifiable Document RAG via Reinforcement Learning

TL;DR

This paper introduces a reinforcement learning framework called Look As You Think (LAT) that trains vision-language models to generate verifiable, evidence-grounded reasoning paths for visual document question answering, improving accuracy and traceability.

Contribution

The paper proposes the Chain-of-Evidence paradigm and LAT framework to unify reasoning and visual evidence attribution with process-level self-verification in VD-RAG.

Findings

LAT improves model performance by 8.23% in soft EM and 47.0% in IoU@0.5.

LAT outperforms supervised fine-tuning baseline in accuracy and generalization.

Experiments demonstrate LAT's effectiveness on Paper- and Wiki-VISA benchmarks.

Abstract

Aiming to identify precise evidence sources from visual documents, visual evidence attribution for visual document retrieval-augmented generation (VD-RAG) ensures reliable and verifiable predictions from vision-language models (VLMs) in multimodal question answering. Most existing methods adopt end-to-end training to facilitate intuitive answer verification. However, they lack fine-grained supervision and progressive traceability throughout the reasoning process. In this paper, we introduce the Chain-of-Evidence (CoE) paradigm for VD-RAG. CoE unifies Chain-of-Thought (CoT) reasoning and visual evidence attribution by grounding reference elements in reasoning steps to specific regions with bounding boxes and page indexes. To enable VLMs to generate such evidence-grounded reasoning, we propose Look As You Think (LAT), a reinforcement learning framework that trains models to produce verifiable reasoning paths with consistent attribution. During training, LAT evaluates the attribution consistency of each evidence region and provides rewards only when the CoE trajectory yields correct answers, encouraging process-level self-verification. Experiments on vanilla Qwen2.5-VL-7B-Instruct with Paper- and Wiki-VISA benchmarks show that LAT consistently improves the vanilla model in both single- and multi-image settings, yielding average gains of 8.23% in soft exact match (EM) and 47.0% in IoU@0.5. Meanwhile, LAT not only outperforms the supervised fine-tuning baseline, which is trained to directly produce answers with attribution, but also exhibits stronger generalization across domains.