WildAWM synthesizes open-source computer-use agent worlds; Qwen3.6 gains +10.6 across 8 benchmarks
HuaxiuYaoML · x · 2026-10-25
- WildAWM (arXiv:2610.26828) is a scalable pipeline for synthesizing executable environments and tasks for computer-use agents, powered entirely by open-source models — no proprietary frontier models needed.
- Each synthesized world is one person's full computer: 15 websites, 8 MCP apps, and a home folder, with one shared state across GUI, MCP, and CLI interfaces so agents can mix visual interaction, tool use, and code execution.
- The 341 environment-task pairs are built to be solvable, hard (GPT-6-Astra reaches only 72.7 pass@8), and reliably verified (95% agreement between verifiers and an agent-as-a-judge ensemble).
- RL training exclusively in these synthetic worlds lifts Qwen3.6-35B-A3B by +10.58 points on average across 8 OOD benchmarks, hitting 80.8 on Online-Mind2Web vs. GPT-5.5's 72.2.
- Fully open-sourced: pipeline, 341 tasks, checkpoints, and the Harbor dataset.
More from coding & agent
- Slop Eisley: a Star Wars fan game fully coded by Claude Opus and Codex opens beta with public invite codes — jasonkneen · 2026-10-25
- Weekend opsec: hand Claude a checklist to audit your home network and devices — zedlander · 2026-10-25
- Cornell paper: 64.7% of AI agent rollouts melt down on errors, no adversary needed — 2C_ornot2C · 2026-10-25
- Blind test of 5 local AI agents on one Qwen model: scores range from 3/5 to 5/5 — GapNew4766 · 2026-10-25
- Meme: Claude as Sisyphus endlessly pushing the 'Some Rust Port' boulder — burny_tech · 2026-10-25
- Running Claude Code as PM over Codex workers to wire xAI Grok in as a coding agent — umeshonai · 2026-10-25