Research map / Multi-agent systems and games / LLM agents
FIRE: Learning to Navigate and Act on Real-World Files via Stateful Reinforcement Learning
Jingyuan Ma, Zheng Li, Rui Li, Siyao Song, Ying Zeng, Minghao Li, Zhifang Sui
Read the paperCheck your idea against 137,217 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.
Formalizes File Reasoning, where agents interact with raw XLSX, PDF, DOCX, and PPTX files in a persistent sandbox, and proposes FIRE, a stateful reinforcement learning approach for it.
Problem
LLMs struggle to answer questions grounded in real-world files such as spreadsheets and slides, where evidence is spread across irregular layouts and heterogeneous formats.
Approach
Agents interact directly with unprocessed files inside a persistent sandbox, and are trained with stateful reinforcement learning on a unified data pipeline spanning multiple file formats.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add stateful RL (GRPO in persistent sandbox) after cold-start SFT (ablation) | Qwen3-32B-FIRE · File reasoning benchmark | Overall accuracy (%): 42.76 -> 50.33 | Improved |
| Add stateful RL after SFT (ablation) | Qwen2.5-32B-FIRE · File reasoning benchmark | Overall accuracy (%): 40.57 -> 48.54 | Improved |
| Add stateful RL after SFT (ablation) | Qwen2.5-7B-FIRE · File reasoning benchmark | Overall accuracy (%): 22.68 -> 30.81 | Improved |
| Add SFT and RL training to base Qwen3-32B | Qwen3-32B · File reasoning benchmark | Overall accuracy (%): 11.14 -> 50.33 | Improved |
| Add RL after SFT on tool-call count (ablation) | Qwen3-32B-FIRE · File reasoning benchmark | Average tool calls: 7.43 -> 6.49 | Improved |
| Add RL after SFT on out-of-distribution file benchmarks (ablation) | FIRE models · GAIA, MEET, Spider, SSTQA | Accuracy | Improved |
| Increase maximum interaction rounds beyond 15 (ablation) | GPT-5, GLM-5, Gemini-2.5-Pro, Qwen3-RL · File reasoning benchmark | Cumulative accuracy | No clear effect |
| Do not penalize intermediate tool failures in reward | FIRE reward design · File reasoning benchmark | tool call success gap between correct and incorrect tasks | Improved |
| Compare benchmark difficulty with existing benchmarks | GPT-5 · SSTQA, GAIA vs FIRE benchmark | Accuracy (%): 76.7 (SSTQA) -> 60.2 (ours) | Hurt |
Closest papers by meaning
- Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections (ICML 2026 · similarity 0.87)
- SPELL: Self-Play Reinforcement Learning for Evolving Long-Context Language Models (ICLR 2026 · similarity 0.86)
- Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining Levels (ICLR 2026 · similarity 0.85)
- LoongRL: Reinforcement Learning for Advanced Reasoning over Long Contexts (ICLR 2026 · similarity 0.85)
- $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (ICLR 2026 · similarity 0.85)
- ReSeek: A Self-Correcting Framework for Search Agents with Instructive Rewards (ICML 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
