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

TL;DR

<3-5 sentence high-level summary>

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.

Figures (17)

• Figure 1: (a) Humans infer information by observing and locating supporting evidence in the document. (b) Each element in the reasoning step is linked to a visual attribution via a bounding box during Chain-of-Evidence generation.
• Figure 2: Overview of the proposed LAT framework. Left: A two-stage training pipeline. Stage I generates and filters the CoE data for fine-tuning. Stage II: The model undergoes refinement via RL under the GRPO algorithm. Right: Rule-based reward design. In GRPO training, the model generates CoE reasoning to guide policy updates through the reward signals.
• Figure 3: Comparison of LAT and SFT performance across different settings and ablation study on threshold $\tau$.
• Figure 4: Performance variation with different $\tau$ (Wiki).
• Figure 5: Data examples from Paper-VISA (left) and Wiki-VISA (right). Each image is assigned a unique identifier, with every dataset entry containing a reference image paired with a specific question, ground-truth answer, and answer source localized by a bounding box. In multi-image scenarios, a retriever selects two images and appends their IDs to the reference image, forming a candidate list. For example, the red bounding box (left) indicates the answer source, where pos_idx=0 signifies that the reference image occupies the first position in the candidate list. For entries lacking ground-truth answers (right), the reference image is substituted with an irrelevant image in the candidate list (pos_idx=-1).