A new training pipeline builds AI agent environments backwards, and the results are hard to ignore.
VHD-Play, described in a new arXiv paper, starts by sampling and solving a mathematical model, then uses that solved model to generate the environment's rules, tools, and scoring criteria. Because the dynamics and the grading rubric come from the same source, there is no separate step to reconcile a simulated world with an answer key written after the fact. The pipeline churns out 3,300 distinct agentic environments at a cost of a few cents each. Researchers then trained Qwen3.6-35B-A3B on three environment families and tested it across five, watching its mean agentic score climb from 0.204 to 0.815.
The gains held up on environment families the model never trained on, and they carried over to real benchmarks for function calling, travel planning, and a 365-day e-commerce simulation, where the trained model finished every run solvent and beat Qwen3.7-Max. The more interesting claim is diagnostic: when researchers compared static, written-out versions of the same problems to stateful ones that hide their parameters, most of the difficulty turned out to live in managing hidden state over time, not in solving the underlying math.
That's a useful reminder for anyone benchmarking "reasoning" models - a lot of what looks like intelligence is really bookkeeping.