DocReward: A Document Reward Model for Structuring and Stylizing

Recent advances in agentic workflows have enabled the automation of tasks such as professional document generation. However, they primarily focus on textual quality, neglecting visual structure and style, which are crucial for readability and engagement. This gap stems mainly from a lack of effective reward models capable of guiding agents toward producing documents with high structural and stylistic professionalism. To address this, we propose DocReward, a document reward model that evaluates documents based on their structure and style. The model is trained under a textual-quality-agnostic framework to assess professionalism without being influenced by textual quality. To achieve this, we construct a multi-domain dataset DocPair of 117K paired documents, covering 32 domains and 267 document types, each comprising a high- and low-professionalism document with identical content but different structure and style. This setup enables the model to evaluate professionalism comprehensively and independently of textual quality. DocReward is trained using the Bradley-Terry loss to score documents, penalizing predictions that contradict the annotated ranking. On a manually annotated benchmark, DocReward outperforms GPT-5 by 14.6 percentage points in accuracy. Extrinsic RL experiments further validate its effectiveness in guiding professional document generation.