WildAWM: fully open-source agent task synthesis yields +10.58 OOD gains for Qwen3.6-35B
XuCanwen · x · 2026-10-21
Researchers released WildAWM (Agent World Model in the Wild), a fully open-source synthesis pipeline for agent environments that requires no proprietary models.
- It synthesizes 341 high-quality environment–task pairs, tackling three core challenges of task synthesis:
- Solvability: every task is guaranteed completable by at least one agent;
- Difficulty: tasks remain hard even for frontier models — GPT-6-Astra only reaches 72.7 pass@8;
- Verifier reliability: 95% agreement between task verifiers and an agent-as-a-judge ensemble.
- Training exclusively with RL in these synthetic worlds, Qwen3.6-35B-A3B improves by +10.58 points on average across 8 OOD benchmarks.
More from coding & agent
- Smart Connections demo: Obsidian surfaces past decisions to reshape your plan — AINewsletter · 2026-10-21
- Ampersand v0.1 Ships via Agent API: Hosted Sessions, OpenAI-Client Compatible — gharik · 2026-10-21
- Pi Durable: Mario Zechner and Armin Ronacher on crash-safe long-running agents — mitsuhiko · 2026-10-21
- POV: Everyone's Too Busy vibe-coding Their Own Game to Play Yours — yacineMTB · 2026-10-21
- Grokbot agent drafts industrial control schematics in open-source PanelCAD — burhop · 2026-10-21
- Salesforce AI's SRD aligns agent foresight with hindsight at zero test-time cost — lupantech · 2026-10-21