Research map / Multi-agent systems and games / LLM agents
ROGA: Scaling Generalist Agents for Office Productivity Tasks via Tool Generation
Mugeng Liu, Xiaojun Ma, Yuhang Xie, Qin Chen, Xuanzhe Liu, Yun Ma
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.
The paper argues that reactive automatic tool generation (ATG) agents fail in stateful office environments and proposes ROGA, a tool-generation approach for generalist agents on office productivity tasks.
Problem
Current automatic tool generation agents, built on reactive paradigms, do not adapt well to realistic office environments that need long-term reasoning and stateful interaction.
Approach
Scales generalist agents for office productivity by generating tools automatically, moving beyond reactive tool generation toward long-horizon, stateful adaptation; specific mechanism not stated in the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ROGA instead of best ATG baseline | GPT-4.1 agents · OSWorld (office tasks) | Pass@1: 18.18 -> 31.82 | Improved |
| use ROGA instead of best ATG baseline | GPT-4.1 agents · WindowsAgentArena (office tasks) | Pass@1: 26.19 -> 28.57 | Improved |
| use ROGA instead of OWL on GAIA-Office | GPT-4.1 agents · GAIA-Office | Pass@1: 73.08 -> 76.92 | Improved |
| add Active World Modeling (ablation) | ROGA · OSWorld (office tasks) | Pass@1: 20.00 -> 31.82 | Improved |
| add Persistent Symbolic Memory (ablation) | ROGA · OSWorld (office tasks) | Pass@1: 19.09 -> 31.82 | Improved |
| add Situated Self-Correction (ablation) | ROGA · OSWorld (office tasks) | Pass@1: 23.64 -> 31.82 | Improved |
| add Capability Lifecycle management (ablation) | ROGA · OSWorld (office tasks) | Pass@1: 29.09 -> 31.82 | Mixed |
| add Capability Lifecycle management (tool reuse) (ablation) | ROGA · reasoning steps | Pass@1: 1.54x increase in steps without it | Improved |
| add Persistent Symbolic Memory (ablation) | ROGA · WindowsAgentArena (office tasks) | Pass@1: 21.42 -> 28.57 | Improved |
| use ROGA instead of specialized SheetAgent | GPT-4.1 · TableBench | Pass@1: 51.47 -> 56.09 | Improved |
| use ROGA instead of specialized SheetAgent | GPT-4.1 · SheetCopilotBench | Pass@1: 23.53 -> 25.91 | Improved |
| use ROGA on non-office math/knowledge tasks | GPT-4.1 agents · Math500 / MMLU-Pro | Pass@1: 70.00 vs 71.00 (OWL) on Math500; 76.00 vs 78.00 (AutoAgent) on MMLU-Pro | No clear effect |
Closest papers by meaning
- AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.87)
- ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.86)
- JADE: Bridging the Strategic-Operational Gap in Dynamic Agentic RAG (ICML 2026 · similarity 0.86)
- Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.86)
- Agent S: An Open Agentic Framework that Uses Computers Like a Human (ICLR 2025 · similarity 0.86)
- Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use (ICML 2026 · similarity 0.86)
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.
