Agent S2: A Compositional Generalist-Specialist Framework for Computer Use Agents
Saaket Agashe , Kyle Wong , Vincent Tu , Jiachen Yang , Ang Li , Xin Eric Wang
- 🏛 Institutions
- Simular Research
- 📅 Date
- April 1, 2025
- 📑 Publisher
- COLM 2025
- 💻 Env
- General GUI
- 🔑 Keywords
TLDR
Agent S2 is a compositional generalist-specialist framework that splits computer-use responsibilities across specialized and generalist models rather than using a single monolithic agent. Its core methods are Mixture-of-Grounding for precise localization and Proactive Hierarchical Planning for long-horizon control, yielding strong gains on OSWorld, WindowsAgentArena, and AndroidWorld.
Related papers (24)
- Watch and Learn: Learning to Use Computers from Online VideosOctober 6, 2025 · CVPR 2026
- CoAct-1: Computer-using Multi-Agent System with Coding ActionsAugust 5, 2025 · ICLR 2026 (Poster)
- LiteCUA: Computer as MCP Server for Computer-Use Agent on AIOSMay 24, 2025 · arXiv
- Iron: Intent-Aligned and Retrospective Dual Learning Framework for Enhancing Generalist Virtual AgentsAugust 28, 2026 · arXiv
- Agent Alpha: Tree Search Unifying Generation, Exploration and Evaluation for Computer-Use AgentsFebruary 3, 2026 · arXiv
- Agentic Reward Modeling: Verifying GUI Agent via Online Proactive InteractionJanuary 31, 2026 · arXiv
- BEAP-Agent: Backtrackable Execution and Adaptive Planning for GUI AgentsJanuary 29, 2026 · arXiv
- LongHorizonUI: A Unified Framework for Robust long-horizon Task Automation of GUI AgentJanuary 26, 2026 · ICLR 2026 (Poster)
- From Off-Policy to On-Policy: Enhancing GUI Agents via Bi-level Expert-to-Policy AssimilationJanuary 9, 2026 · arXiv
- R-WoM: Retrieval-augmented World Model For Computer-use AgentsOctober 13, 2025 · ICLR 2026 (Poster)
- Just Do It!? Computer-Use Agents Exhibit Blind Goal-DirectednessOctober 2, 2025 · ICLR 2026 (Poster)
- SheetCopilot: Bringing Software Productivity to the Next Level through Large Language ModelsMay 30, 2023 · NeurIPS 2023
- Improving Proficiency and Efficiency of Android GUI Agents via Self-Generating Tool ActionsSeptember 6, 2026 · arXiv
- CUA-Universe: A Scalable and Dynamic Environment for Hybrid GUI+CLI AgentsSeptember 4, 2026 · arXiv
- OmegaUse-SOP: SOP Engineering for Professional Computer Use from Human DemonstrationsSeptember 2, 2026 · arXiv
- ASIL: Replacing Screenshot-and-Click with Structured State and Semantic ActionsAugust 27, 2026 · arXiv
- AOHP: An Open-Source OS-Level Agent Harness for Personalized, Efficient and Secure InteractionJune 22, 2026 · arXiv
- MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent ResearchMay 25, 2026 · arXiv
- ScaleWoB: Guiding GUI Agents with Coding Agents via Large-Scale Environmental SynthesisMay 24, 2026 · arXiv
- VLAA-GUI: Knowing When to Stop, Recover, and Search, A Modular Framework for GUI AutomationApril 23, 2026 · arXiv
- ClawGUI: A Unified Framework for Training, Evaluating, and Deploying GUI AgentsApril 13, 2026 · arXiv
- EE-MCP: Self-Evolving MCP-GUI Agents via Automated Environment Generation and Experience LearningApril 10, 2026 · arXiv
- IntentScore: Intent-Conditioned Action Evaluation for Computer-Use AgentsApril 6, 2026 · arXiv
- GUIDE: Resolving Domain Bias in GUI Agents through Real-Time Web Video Retrieval and Plug-and-Play AnnotationMarch 27, 2026 · arXiv