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

TL;DR

This paper introduces DocR1, a novel multi-page document understanding model trained with a reinforcement learning framework that guides reasoning across pages, achieving state-of-the-art results with limited supervision.

Contribution

The paper presents EviGRPO, a new RL-based training method with an evidence-aware reward, and constructs new datasets for multi-page document understanding.

Findings

DocR1 outperforms existing models on multi-page benchmarks.

EviGRPO effectively guides models to retrieve relevant pages before answering.

Limited supervision suffices for high-quality multi-page reasoning models.

Abstract

Understanding multi-page documents poses a significant challenge for multimodal large language models (MLLMs), as it requires fine-grained visual comprehension and multi-hop reasoning across pages. While prior work has explored reinforcement learning (RL) for enhancing advanced reasoning in MLLMs, its application to multi-page document understanding remains underexplored. In this paper, we introduce DocR1, an MLLM trained with a novel RL framework, Evidence Page-Guided GRPO (EviGRPO). EviGRPO incorporates an evidence-aware reward mechanism that promotes a coarse-to-fine reasoning strategy, guiding the model to first retrieve relevant pages before generating answers. This training paradigm enables us to build high-quality models with limited supervision. To support this, we design a two-stage annotation pipeline and a curriculum learning strategy, based on which we construct two datasets: EviBench, a high-quality training set with 4.8k examples, and ArxivFullQA, an evaluation benchmark with 8.6k QA pairs based on scientific papers. Extensive experiments across a wide range of benchmarks demonstrate that DocR1 achieves state-of-the-art performance on multi-page tasks, while consistently maintaining strong results on single-page benchmarks.