WildAWM: Synthetic Computer Worlds Train Agents That Stump GPT-6 Astra

Background and Motivation

Training computer-use agents (CUAs) in real environments is costly, hard to scale, and poses safety risks (e.g., an agent accidentally attacking external servers). WildAWM addresses this by synthesizing entire personal computer environments — shell, filesystem, browser, and MCP services like email — entirely with open-source models, avoiding any reliance on proprietary systems.

From Agent World Model to WildAWM

The prior Agent World Model (AWM, ICML 2026) synthesized 1,000 isolated tool environments for multi-turn tool-use RL. WildAWM extends this to a unified personal computer where CLI, GUI, and MCP interfaces share state, enabling agents to switch interfaces fluidly.

Synthesizing a Complete Computer Environment

Each environment models a fictional person: a profession, contacts, frequent websites, and ongoing tasks scheduled across days. A coding agent (GLM-5.2 + OpenCode) executes daily plans, writing database records and files, so data accumulates naturally. The process yielded 36 personas and 640 world states; by the final day each world contained ~170,000 database records and ~50 local files, exceeding the complexity of contemporaneous synthetic environments. Consistency is enforced by replaying each day twice and verifying with an independent agent.

Cross-Interface Task Synthesis and Validation

Tasks (called cross-agent-interface, XAI) require using CLI, GUI, and MCP in combination. Example: Jordan must re-evaluate GPS watches for a half-marathon by scripting candidate extraction from notes (CLI), visually identifying a watch on Amazon from a description (GUI), logging the choice locally (CLI), and notifying a friend via chat (MCP). Tasks average 4.2 subtasks, 3.1 interface switches, 74.5% revisit a previous interface, and can exceed 100 steps.

To ensure quality without proprietary models, WildAWM runs each draft task through 6 executions using three open-source models (GLM-5.2, DeepSeek-V4-Flash, MiniMax-M3) across six agent harnesses (OpenCode, Codex, Claude Code, Nanobot, OpenClaw, Hermes Agent). A task passes five gates:

Judgment consistency : verifier, LLM judge, and agent judge must agree on all 6 runs.

Initial unsolved : verifier must fail when the agent does nothing.

Solvability : at least one of 6 runs succeeds.

Difficulty : the 3 no-hint runs must not all succeed.

Counterfactual probe : agent judge must not flip its verdict when fed the opposite conclusion.

From 2,551 drafts, 341 tasks survived.

Task Quality Evaluation

Solvability : 24 model-framework combinations (including Opus-4.8) solved every task at least once; the three weaker open-source models alone solved >95%.

Difficulty : Without hints, average success rate was 21.3%; 273 tasks were rated hard. GPT-6 Astra achieved Pass@1 ~60% and Pass@8 still left 27.3% unsolved, confirming the tasks challenge a frontier model.

Verifier reliability : On 100 sampled trajectories, the lightweight verifier script matched the majority vote of heavy agent judges 95% of the time (Cohen's κ = 0.90), ensuring trustworthy RL rewards.

Training on Synthetic Data Only

The 341 tasks were split: 100 for SFT, 208 for RL, 33 held-out validation. Training used GLM-5.2 trajectories for SFT warm-start, then up to 64 GRPO steps on two Qwen models (Qwen3.6-35B-A3B and Qwen3.5-4B). Evaluation covered 8 unseen OOD benchmarks: WebArena, Online-Mind2Web (web agents); QwenClawBench, ClawEval (CLI agents); CUA-Gym, OSWorld-Verified, OSWorld 2.0, Agents' Last Exam (GUI computer-use agents).

Both models improved on all 8 benchmarks. Qwen3.6-35B-A3B average score rose from 42.84 to 53.42 (+10.58), with notable gains on ClawEval (+16.67), OSWorld-Verified (+16.06), WebArena (+13.72), and CUA-Gym (+12.22). Post-training, it matched GLM-5.2 on WebArena (67.26) and surpassed GPT-5.5 on Online-Mind2Web (80.77 vs 72.22). On WildAWM's own validation set, its pass rate jumped from 5.68% to 35.98%, nearing GLM-5.2's 37.88% despite 20× fewer parameters. Qwen3.5-4B improved from 21.83 to 30.04 (+8.21).

Ablation and Behavioral Analysis

Removing any of the five validation gates reduced Qwen3.5-4B average scores by 2.60–9.30 points; the largest drops came from dropping judgment consistency and initial unsolved checks.

RL induced three behavioral shifts in Qwen3.6-35B-A3B:

Interface-switching rate between consecutive tool calls rose from 18.4% to 25.5%, approaching GLM-5.2's level.

Long-horizon performance (40–120 steps) improved by 17.0 percentage points vs 9.7 for SFT-only.

RL could not surpass the teacher model (GLM-5.2) on the hardest difficulty bucket, consistent with the view that RL rearranges existing capabilities rather than creating new ones.

Open-Source Release and Recursive Self-Improvement Vision

The full synthesis pipeline, 341 environments and tasks, trained Qwen models, and Harbor-formatted datasets are open-sourced (GitHub: MSR-Orchard/WildAWM; Hugging Face: ChilleD/agent-world-model-in-the-wild). The authors argue this demonstrates a path for open-source models to recursively self-improve: synthesize expanding environments and tasks, train on them, iterate — all without proprietary models, avoiding indirect distillation.