DocSeeker: Structured Visual Reasoning with Evidence Grounding for Long Document Understanding

Conference: CVPR 2026 arXiv: 2604.12812 Code: https://github.com/yh-hust/DocSeeker Area: Multimodal VLM / Document Understanding Keywords: Long document understanding, evidence grounding, structured reasoning, reinforcement learning, visual RAG

TL;DR

DocSeeker is proposed to achieve structured reasoning and evidence grounding in long document understanding via an ALR (Analyze–Locate–Reason) visual reasoning paradigm combined with two-stage training (SFT + EviGRPO). The model is trained exclusively on short documents yet generalizes robustly to documents of extreme length.

Background & Motivation

Background: MLLMs suffer severe performance degradation on long-document VQA as document length increases. Pure-vision approaches treat each page as an image input to avoid OCR error propagation.

Limitations of Prior Work: (1) Low signal-to-noise ratio: key evidence is buried among a large number of irrelevant pages; (2) Supervision scarcity: datasets provide only short final answers without intermediate reasoning steps. Visual RAG faces a Top-\(k\) dilemma—large \(k\) introduces noise while small \(k\) misses evidence.

Key Challenge: Models learn fragile shortcuts (memorization) rather than genuine reasoning ability, resulting in poor interpretability and weak OOD generalization.

Goal: Train models to adopt a structured "locate first, then reason" workflow rather than directly predicting answers.

Key Insight: Inspired by human cognitive processes—first analyze intent, then locate evidence, and finally reason.

Core Idea: The ALR paradigm requires the model to explicitly produce a structured chain of thought following the "Analyze → Locate → Reason" sequence, combined with two-stage training via SFT and evidence-aware GRPO.

Method

Overall Architecture

Built upon Qwen-2.5-VL-7B, with page IDs prefixed to each page as pointers. The output is constrained to follow the ALR structure: \(\mathbf{Y} = (\mathbf{Y}_A \oplus \mathbf{Y}_L \oplus \mathbf{Y}_R) \oplus (\mathbf{Y}_E \oplus \mathbf{Y}_F)\), comprising question analysis, evidence localization (with cited page numbers), reasoning process, evidence page list, and final answer.

Key Designs

1. ALR Visual Reasoning Paradigm:

   • Function: A structured "locate first, then reason" workflow.
   • Mechanism: Page-aware input (interleaved page IDs and visual tokens) combined with three-stage structured output. The model must first analyze user intent, then scan the document to locate relevant pages with justification, and finally synthesize reasoning from the grounded evidence.
   • Design Motivation: Enforced evidence localization compels the model to distinguish visual tokens across different pages, counteracting interference caused by long visual inputs.

2. Evidence-Aware GRPO (EviGRPO):

   • Function: Jointly optimizes evidence localization and reasoning via reinforcement learning.
   • Mechanism: A multi-dimensional reward function \(R = \lambda_1 R_{format} + \lambda_2 R_{evidence} + \lambda_3 R_{answer}\). The format reward enforces the ALR template; the evidence reward evaluates page localization accuracy using a weighted F1 score (\(\beta > 1\), biased toward recall); the answer reward evaluates the final answer using ANLS.
   • Design Motivation: Reasoning paths produced by SFT tend to be suboptimal; RL enables the model to learn directly from outcome signals, surpassing imitation learning.

3. Evidence-Guided Resolution Allocation (EGRA):

   • Function: Supports longer document inputs during training.
   • Mechanism: Evidence pages are kept at high resolution; 70% of non-evidence pages are downsampled (1024→256) while the remaining 30% retain high resolution. At inference time, all pages are processed at high resolution.
   • Design Motivation: This not only alleviates GPU memory constraints but also improves the signal-to-noise ratio of training data—outperforming the strategy of simply discarding non-evidence pages.

Loss & Training

Stage I: Standard cross-entropy SFT on 13,986 ALR CoT samples distilled from Gemini-2.5-Flash. Stage II: EviGRPO (rollout group size 16; format/evidence/answer reward weights of 0.1/0.3/0.6). Training is conducted exclusively on documents with \(\leq 20\) pages.

Key Experimental Results

Main Results

Method	Params	DUDE↑	MPDocVQA↑	MMLong↑	LongDocURL↑
Baseline	7B	35.2	70.1	25.4	37.8
InternVL3	8B	47.4	80.8	24.1	38.7
GPT-4o	—	54.1	67.4	42.8	64.5
DocSeeker	7B	56.8	87.2	48.5	58.3

Ablation Study

Configuration	DUDE	MPDocVQA	Notes
Full DocSeeker	56.8	87.2	SFT + EviGRPO + EGRA
SFT only	52.1	84.5	No RL
SFT + GRPO (no evidence reward)	54.3	85.8	Standard GRPO
Without EGRA	50.8	82.1	Uniform resolution

Key Findings

• Improvements of 30–60% over the baseline demonstrate the effectiveness of the ALR paradigm.
• Trained exclusively on documents with \(\leq 20\) pages, the model generalizes robustly to documents of up to 468 pages.
• DocSeeker's localization capability naturally complements visual RAG and can serve as a backbone model for RAG systems.

Highlights & Insights

• "Train short, generalize long" is a surprising result: the ALR paradigm learns transferable reasoning skills rather than memorized patterns.
• The EGRA strategy is simple yet effective: differentiated resolution allocation simultaneously reduces memory consumption and improves signal-to-noise ratio, outperforming the strategy of discarding non-evidence pages.
• The evidence-aware reward design makes the RL stage more targeted and purposeful.

Limitations & Future Work

• Training data is drawn solely from MP-DocVQA and DUDE, limiting domain coverage.
• Reliance on Gemini-2.5-Flash for distillation constrains data quality to the teacher model's capability.
• The pure-vision approach still has limitations on dense-text pages.
• The framework is extensible to multi-document cross-document reasoning.

Related Work & Insights

• vs. VisRAG/SV-RAG: These are retrieval-augmented methods; the ALR paradigm endows DocSeeker as an end-to-end approach with explicit localization capability.
• vs. mPLUG-DocOwl2: DocOwl2 employs visual token compression, whereas DocSeeker applies differentiated resolution via EGRA.

Rating

• Novelty: ⭐⭐⭐⭐⭐ Both the ALR paradigm and EviGRPO represent significant innovations.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive in-domain and out-of-domain evaluation with detailed ablations.
• Writing Quality: ⭐⭐⭐⭐⭐ Method and experiments are articulated clearly.
• Value: ⭐⭐⭐⭐⭐ Substantial advancement for long document understanding.

Related Papers

• [AAAI 2026] URaG: Unified Retrieval and Generation in Multimodal LLMs for Efficient Long Document Understanding
• [CVPR 2026] Reason-SVG: Enhancing Structured Reasoning for Vector Graphics Generation with Reinforcement Learning
• [CVPR 2026] MSJoE: Jointly Evolving MLLM and Sampler for Efficient Long-Form Video Understanding
• [CVPR 2026] Recurrent Reasoning with Vision-Language Models for Estimating Long-Horizon Embodied Task Progress
• [NeurIPS 2025] iFinder: Structured Zero-Shot VLM Grounding for Dash-Cam Video Reasoning