Research map / Vision-language and multimodal / Multimodal LLMs

DocR1: Evidence Page-Guided GRPO for Multi-Page Document Understanding

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 2 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 DocR1, a Qwen2.5-VL-7B model trained with EviGRPO (GRPO plus an evidence-page F1 reward) on a 4.8k-sample curated set to improve multi-page document understanding, and releases the ArxivFullQA benchmark.

Problem

Multi-page document understanding requires locating evidence across pages and multi-hop reasoning, which MLLMs handle poorly, and RL has not been explored for this setting.

Approach

An evidence-aware GRPO variant (EviGRPO) that adds a per-page True/False evidence judgment and an F1-based evidence-page reward to format and ANLS accuracy rewards, with a single-page then multi-page curriculum.

Key results

• DocR1 (7B) averages 59.36 across seven multi-page benchmarks, versus 52.43 for Qwen2.5-VL-7B-Instruct and 46.80 for Qwen2.5-VL-32B.
• Evidence page recall averages 91.47 for DocR1 versus 52.70 for the Qwen2.5-VL baseline on three multi-page benchmarks.
• On single-page benchmarks DocR1 is comparable to the baseline, e.g., +1.0 on WikiTableQuestions and +1.6 on TabFact, while the 7B baseline matches on DocVQA (95.1).

Surprising findings

• Training with single-page data only sharply degrades multi-page performance (e.g., MP-DocVQA drops by 10.44 points), showing single-page RL does not transfer to evidence retrieval across pages.
• Omitting the image count in the page selection prompt performed better than giving it explicitly (PSF-2), suggesting explicit counts let the model shortcut the reasoning.

Limitations

• Training data annotated by a single closed model (Gemini 2.5 Flash) with self-verification, so label quality depends on that model.
• Several multi-page benchmarks use validation sets rather than official test sets, and single-page gains are marginal or mixed.

Experimental setup

Datasets: DocVQA InfographicVQA ChartQA DeepForm DVQA FigureQA KleisterCharity OCRVQA TabFact TextCaps TextVQA VisualMRC

Benchmarks: DocVQA InfographicVQA WikiTableQuestions TabFact TextVQA VisualMRC MP-DocVQA DUDE SlideVQA MultiChartQA MultiHiertt TATDoc

Metrics: ANLS recall of evidence pages

Baselines: Qwen2.5-VL-Instruct UReader TextMonkey DocOwl-1.5 DocOwl-2 LLaVA-NeXT-Interleave LEOPARD-Idefics2 mPLUG-DocOwl2 LLaVA-OneVision InternVL3

Builds on: GRPO DeepSeek-R1 Qwen2.5-VL VisualRFT Video-R1

Closest papers by meaning

• DocSeeker: Structured Visual Reasoning with Evidence Grounding for Long Document Understanding (CVPR 2026 · similarity 0.89)
• Point-RFT: Improving Multimodal Reasoning with Visually Grounded Reinforcement Finetuning (NeurIPS 2025 · similarity 0.88)
• DocHop: Benchmarking Out-of-domain Multi-hop Reasoning in Information-Dense Documents (ICML 2026 · similarity 0.88)
• Look as You Think: Unifying Reasoning and Visual Evidence Attribution for Verifiable Document RAG via Reinforcement Learning (AAAI 2026 · similarity 0.87)
• Resolving Evidence Sparsity: Agentic Context Engineering for Long-Document Understanding (CVPR 2026 · similarity 0.86)
• GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking (CVPR 2026 · similarity 0.86)