FIRE: Learning to Navigate and Act on Real-World Files via Stateful Reinforcement Learning

Large language models still struggle to reliably answer questions grounded in real-world files like spreadsheets and slides, where evidence is scattered across irregular layouts and heterogeneous formats. We address this by formalizing File Reasoning, a setting where agents must interact directly with unprocessed files (XLSX, PDF, DOCX, PPTX) within a persistent sandbox. To support this, we introduce a unified data pipeline and a high-difficulty benchmark of over 400 verifiable questions that preserve native file structure. Furthermore, we propose a reinforcement learning framework grounded in stateful file execution. We train FIRE (File Interactive Reasoning Expert), a family of models that learn to optimize long-horizon planning using genuine execution feedback. Unlike stateless tool-use methods, this environmental persistence enables agents to iteratively recover from errors and adapt to structural ambiguities. Empirical results show that Qwen3-32B-FIRE achieves the strongest performance among the open-source models of comparable scale under identical execution constraints.

Many everyday work tasks require reading and using information from real files, such as spreadsheets, PDFs, Word documents, and presentation slides. However, current AI systems often struggle when they need to open files, find the right information, perform calculations, and keep track of intermediate results across several steps. In this paper, we study how to make language models better at this kind of practical file-based work. We build a collection of realistic questions that can only be answered by carefully interacting with the provided files, and we develop a training approach that lets models practice solving such tasks in a safe computer environment. Our results show that this training substantially improves the model’s ability to handle real-world files, especially spreadsheets and documents that require multi-step reasoning. We hope this work helps move AI systems closer to being useful and reliable assistants for everyday information processing tasks.