Research map / Vision-language and multimodal / Multimodal LLMs

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

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DocSeeker, a Qwen2.5-VL-7B-based document MLLM trained with distilled Analysis-Localization-Reasoning chain-of-thought SFT, evidence-aware GRPO, and evidence-guided resolution allocation for long multi-page document VQA.

Problem

Multimodal LLMs degrade on long multi-page document VQA because crucial evidence is buried among irrelevant pages (low signal-to-noise ratio), and existing datasets provide only short final answers, giving weak supervision for evidence localization and reasoning.

Approach

A structured Analysis-Localization-Reasoning output format with page-id-prefixed visual tokens, trained via teacher-distilled CoT SFT followed by GRPO with an evidence-page F1 reward (recall-weighted) plus format and answer rewards, and a training-time resolution scheme that keeps evidence pages at high resolution while downsampling most non-evidence pages.

Key results

• DocSeeker reaches 40.1% accuracy on MMLongBench-Doc versus 25.4% for the Qwen2.5-VL-7B baseline, and 86.2 ANLS on MPDocVQA.
• Ablation: EviGRPO with reward weights (0.1, 0.3, 0.6) gives 40.1 Acc / 38.4 F1, above SFT alone at 38.6 / 36.9 and vanilla GRPO at 38.7 / 36.3.
• With the Baseline, MMLongBench-doc accuracy drops from 34.5% to 13.9% as document length grows, while DocSeeker stays largely stable.

Surprising findings

• Raw short-answer SFT yields only marginal out-of-distribution gains (MMLongBench-Doc 28.8 vs 25.4 baseline in the main table), suggesting memorization rather than reasoning.
• The ALR paradigm nearly doubles output tokens, but end-to-end latency rises only from 19s to 25s because visual prefill dominates.

Limitations

• The ALR paradigm adds output tokens and modest latency at inference.
• Training data distillation depends on a proprietary teacher (Gemini-2.5-Flash), and about a third of samples fail the initial exact-match filter.

Experimental setup

Datasets: MP-DocVQA DUDE MMLongBench-Doc LongDocURL SlideVQA

Benchmarks: DUDE MPDocVQA MMLongBench-Doc LongDocURL SlideVQA

Metrics: ANLS accuracy F1

Baselines: Qwen2.5-VL-7B baseline Baseline-SFT (short-answer) InternVL3 mPLUG-DocOwl2 M3DocRAG VDocRAG SV-RAG GPT-4o

Builds on: GRPO Qwen2.5-VL chain-of-thought

Closest papers by meaning

• URaG: Unified Retrieval and Generation in Multimodal LLMs for Efficient Long Document Understanding (AAAI 2026 · similarity 0.90)
• Resolving Evidence Sparsity: Agentic Context Engineering for Long-Document Understanding (CVPR 2026 · similarity 0.89)
• DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document Understanding (NeurIPS 2026 · similarity 0.89)
• Look as You Think: Unifying Reasoning and Visual Evidence Attribution for Verifiable Document RAG via Reinforcement Learning (AAAI 2026 · similarity 0.89)
• DocR1: Evidence Page-Guided GRPO for Multi-Page Document Understanding (AAAI 2026 · similarity 0.89)
• DocHop: Benchmarking Out-of-domain Multi-hop Reasoning in Information-Dense Documents (ICML 2026 · similarity 0.89)