Language-model agents are increasingly asked to handle long-horizon tasks with evolving state, interdependent decisions, and delayed outcomes. Scaling their training demands diverse agentic environments, dependable outcome signals, and low extension cost — three things existing generation pipelines struggle to deliver together.
According to a new paper posted on Hugging Face, VHD-Play (Verifiable Hidden Dynamics Play) reverses the usual dependency order. Instead of building an environment and defining its outcome rule afterwards, the pipeline first samples and solves a mathematical model. A corpus-grounded setter then renders that decision process as stateful tools, so the executable dynamics and the trajectory-scoring reference are both inherited from the same solved model.
The authors report that the pipeline produces 3,300 diverse agentic environments at a cost of a few cents each. Training Qwen3.6-35B-A3B on three families raised its mean agentic score from 0.204 to 0.815 in a five-family diagnostic.
Gains were not confined to the training substrate. The paper reports improvements on held-out instances from all three training families and eight unseen mechanism families, and on external benchmarks covering general function calling, travel planning, and 365-day e-commerce. On E-Commerce Bench, the trained checkpoint reportedly completes every run without bankruptcy and exceeds Qwen3.7-Max.
A comparison between written-out problems and stateful versions that reveal or hide their parameters suggests that most of the learnable gap lies in stateful interaction rather than underlying problem solving. The authors also note that a frozen 35B setter can realize larger environments, and that scale-matched training retains gains as mechanism size and horizon grow — pointing toward an evolving training substrate.