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. To support this, we design a rigorous two-stage annotation pipeline and a curriculum learning strategy that enables effective training with limited supervision. Using this pipeline, we construct two datasets: EviBench, a high-quality training set with 4.8k examples, and ArxivFullQA, a benchmark with 8.6k QA examples over full scientific papers. Extensive experiments across a wide range of benchmarks demonstrate that DocR1 achieves state-of-the-art performance on multi-page tasks while maintaining strong results on single-page benchmarks.

In community-based software development, live-chatting services are increasingly used to discuss bugs encountered during development. Many methods have emerged to identify bugs and produce bug reports, which further improve the efficiency of software development. However, previous methods still face challenges in understanding complex conversational structures and classifying sentences in bug reports, as entertaining or meaningless utterances often lower the quality of constructed bug reports. To address this issue, we propose a method named BR-Hunter, which comprises the following four components. Specifically, the data preprocessing component disentangles and denoises the live chats, while the utterance embedding component aims to extract the semantic features of each utterance in the conversations. The bug report identification component then models the conversation as a feature graph and uses Graph Neural Networks to identify conversations containing bug reports, thereby solving Challenge 1. Finally, the bug report synthesis (BRS) component tackles Challenge 2 by classifying and reassembling sentences from conversations containing bug reports, leveraging fine-tuned BERT and prompt learning techniques. Extensive experiments conducted on eight open source projects demonstrate that BR-Hunter achieves high accuracy in identifying bug reports. Compared to baseline methods, BR-Hunter improves the average F1 score by 36.41%, 24.80%, 68.92%, 46.77%, 52.84%, 25.80%, 25.25%, and 4.19%, respectively. And BR-Hunter also achieves an average improvement of 10.34% on the BRS task, compared with the state-of-the-art method.

Haoyu Dong, Jianbo Zhao, Yuzhang Tian, Junyu Xiong, Mengyu Zhou, Yun Lin, José Cambronero, Yeye He, Shi Han, Dongmei Zhang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.