Summary

SEAgent provides the official implementation for a self-evolving computer use agent capable of autonomous learning from experience, detailed in the ICML-2026 paper. It targets researchers and developers in AI agents and human-computer interaction, enabling agents to autonomously learn and improve performance on computer tasks through self-generated experience and feedback.

How It Works

The core architecture features a self-evolving loop: an actor model (e.g., UI-TARS) executes tasks, while a World State Model judges trajectory success. It integrates curriculum generation for task diversity and autonomous learning from SFT/RL data. Agents interact with simulated environments like OSWorld, receiving feedback to iteratively refine their capabilities.

Quick Start & Requirements

• Installation: Requires Conda with Python 3.11, followed by bash setup.sh.
• Prerequisites: Significant computational resources are implied, with explicit mentions of CUDA for vllm serving and multiple GPUs (e.g., CUDA_VISIBLE_DEVICES=0,1,2,3). Requires specific models like SEAgent-1.0-7B, World-State-Model-7B, Qwen/Qwen2.5-72B-Instruct, and bytedance-research/UI-TARS-7B-DPO.
• Environments: Primarily tested with OSWorld and AgentRewardBench.
• Links: 📖 Paper: https://arxiv.org/abs/2508.04700, 🤗 SEAgent-1.0-7B, 🤗 World State Model-7B.

Highlighted Details

• Official implementation for the ICML-2026 paper "SEAgent: Self-Evolving Computer Use Agent with Autonomous Learning from Experience".
• Features a self-evolving loop for autonomous learning and improvement.
• Supports task generation via curriculum learning and feedback from a World State Model.
• Includes released models: SEAgent-1.0-7B and World-State-Model-7B.

Maintenance & Community

No specific details on community channels (Discord/Slack), active contributors beyond authors, or roadmap are provided in the README.

Licensing & Compatibility

• License: Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).
• Restrictions: Data and code are licensed for research use only. Commercial use is explicitly prohibited. Adherence to OpenAI's terms of use is also noted.

Limitations & Caveats

The project is licensed strictly for research purposes, prohibiting commercial use. The setup involves complex dependencies and significant computational resources, including multiple GPUs and specific large language models, which may pose an adoption barrier.