Envs-FORGE trains policies on verifier-checked generated environments
The method changes training environments and model weights; its synthesis budget is separate from inference cost.
Envs-FORGE estimates seed-task pass rates and chooses environment-generation actions near a learning frontier. A per-seed mixed-integer linear program selects a projection/direction action, after which the instruction, fixtures, oracle, tests and Docker environment are rewritten together. Only verified bundles enter reinforcement-learning training.
On Qwen 3.5 35B, the paper reports Pass@1 improvements from 40.0% to 49.2% on Terminal-Bench core, 23.0% to 29.4% on 2.0 and 73.4% to 77.1% on SWE-bench Verified. Comparators export 100 verified environments at 2.27–2.88 million synthesis tokens.
The model size is fixed, but RL updates its weights. The original not-models framing confused size with training. MILP and generation expense occur during environment synthesis, not as the stated cost of every served inference.
For adoption, compare downstream learning at a matched verified-environment and synthesis allowance. Include verifier failures and total training expense before inferring that task generation is the cheaper route.
Editorial correction (2026-10-13 UTC): Corrected fixed-size versus unchanged-weight claims and separated synthesis from serving budgets.