Research map / Multi-agent systems and games / LLM agents
Self-Evolving MCP-GUI Agents via Automated Environment Generation and Experience Learning
Tiantian He, Yihang Chen, Keyue Jiang, Ka Lee, Kaiwen Zhou, Kun Shao, Shuai Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add trajectory distillation (SFT) with experience bank | Qwen3-VL-8B baseline · Chrome | pass rate: 60.0% -> 77.8% | Improved |
| add inference-time experience bank | Qwen3-VL-8B baseline · Chrome | pass rate: 60.0% -> 62.2% | Improved |
| add inference-time experience bank | Qwen3-VL-8B baseline · VS Code | pass rate: 53.3 -> 63.3 | Improved |
| add inference-time experience bank | Qwen3-VL-8B baseline · LibreOffice Calc | pass rate: 44.4 -> 46.7 | No clear effect |
| add trajectory distillation with experience bank | Qwen3-VL-8B baseline · VS Code | pass rate: 53.3 -> 43.3 | Hurt |
| add trajectory distillation with experience bank | Qwen3-VL-8B baseline · LibreOffice Calc | pass rate: 44.4 -> 42.2 | Hurt |
| add trajectory distillation with experience bank | Qwen3-VL-8B baseline · Chrome hard tasks | tasks passed: 4/15 -> 8/15 | Improved |
| add experience bank | Qwen2.5-VL-7B baseline · Chrome | pass rate: 48.9% -> 40.0% | Hurt |
| run additional self-evolution iterations of experience bank only | Experience Bank · Chrome | pass rate: 60.0 -> 64.4 | Improved |
| add more iterations after iteration 1 for distillation | Distillation + Experience Bank · Chrome | pass rate: 77.8 (iter 1) | No clear effect |
| add gap-driven task generation and rejection sampling to the training pool | Self-evolution pipeline · Chrome | training samples: 44 -> 79 samples | Improved |
| add self-evolution iterations to distillation pipeline | Distillation + Experience Bank · Chrome | MCP action ratio: 43.7% -> 55.8% | Improved |
Closest papers by meaning
- OSWorld-MCP: Benchmarking MCP Tool Invocation In Computer-Use Agents (ICLR 2026 · similarity 0.89)
- Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolution (ICML 2026 · similarity 0.88)
- SEAgent: Self-Evolving Computer Use Agent with Autonomous Learning from Experience (ICML 2026 · similarity 0.88)
- SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Co-EPG: A Framework for Co-Evolution of Planning and Grounding in Autonomous GUI Agents (AAAI 2026 · similarity 0.88)
- ComplexMCP: Evaluation of LLM Agents in Dynamic, Interdependent, and Large-Scale Tool Sandbox (ICML 2026 · similarity 0.87)
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.
