Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms
Xinjie Shen (Georgia Tech) with Wei Fan, Dayiheng Liu and colleagues at Alibaba Token Foundry (Qwen technical report) present VHD-Play, which generates agentic RL environments by first solving a mathematical model and then rendering its decision process as stateful tools.
Ask this paper
Reversed pipeline. Instead of building an environment and defining its reward afterward, VHD-Play samples and solves a mechanism first, so the executable dynamics and the trajectory-scoring reference come from the same solved model.
Scale and cost. The pipeline produced 3,300 environments at a few cents each.
Training result. Training Qwen3.6-35B-A3B on three families raises its mean agentic score from 0.204 to 0.815 on a five-family diagnostic, with gains on held-out instances and eight unseen mechanism families.
External transfer. Gains carry to general function calling, travel planning and a 365-day e-commerce benchmark, where the trained model never goes bankrupt and exceeds Qwen3.7-Max.
Where the gap is. Comparing written-out problems with stateful versions shows most of the learnable gap is in stateful interaction, not in solving the underlying problem.
Abstract
Language-model agents increasingly face long-horizon tasks with evolving state, interdependent decisions, and delayed outcomes. Scaling their training requires diverse agentic environments, dependable outcome signals, and low extension cost. Existing generation pipelines commonly construct an environment before defining its outcome rule or annotating its trajectories, leaving dynamics and evaluation to be aligned post hoc. VHD-Play reverses this dependency by sampling and solving a mathematical model before a corpus-grounded setter renders its decision process as stateful tools. The executable dynamics and trajectory-scoring reference are inherited from the same solved model. The pipeline produces 3,300 diverse agentic environments at a cost of a few cents each. Training Qwen3.6-35B-A3B on three families raises its mean agentic score from 0.204 to 0.815 in a five-family diagnostic. Gains also appear on held-out instances from all three training families and eight unseen mechanism families, then extend beyond the generated substrate to external benchmarks for general function calling, travel planning, and 365-day e-commerce. On E-Commerce Bench, the trained checkpoint completes every run without bankruptcy and exceeds Qwen3.7-Max. We compare written-out problems with stateful versions that reveal or hide their parameters. The comparison shows that most of the learnable gap lies in stateful interaction rather than underlying problem solving. A frozen 35B setter realizes larger environments, and scale-matched training retains gains as mechanism size and horizon grow, indicating the potential for an evolving training substrate.