VHD-Play Generates 3,300 Agentic RL Environments at a Few Cents Each

Xinjie Shen · hf · 2026-09-24

VHD-Play reverses typical environment-generation pipelines: it first samples and solves a mathematical model, then a corpus-grounded setter renders the solved mechanism's decision process as stateful tools, with executable dynamics and trajectory-scoring reference inherited from the same model—avoiding post hoc alignment between environments and outcome rules.

Produces 3,300 diverse agentic environments at a few cents each
Training Qwen3.6-35B-A3B on three families raises mean agentic score from 0.204 to 0.815 on a five-family diagnostic
Gains generalize to held-out instances, eight unseen mechanism families, and external benchmarks for function calling, travel planning, and e-commerce
On E-Commerce Bench the trained checkpoint never goes bankrupt and exceeds Qwen3.7-Max
Most of the learnable gap lies in stateful interaction, not underlying problem solving

Original post →

VHD-Play Generates 3,300 Agentic RL Environments at a Few Cents Each

More from coding & agent