Envs-FORGE Solves a Per-Seed MILP to Synthesize Agent-RL Environments, Hitting 77.1% on SWE-bench Verified vs 73.4% Base

arXiv 2608.14312·high signalXBlueskyLinkedInCopy link

Fixed synthesis recipes (few-shot, Self-Instruct, Evol-Instruct) apply the same prompting policy to every seed regardless of whether the current policy needs harder or easier tasks. Envs-FORGE estimates per-seed pass rates, scores six projection-direction actions around a target learning frontier, and solves a mixed-integer linear program per seed to pick the action that conditions generation — rewriting instruction, fixtures, oracle solution, tests, and Docker image in sync, with only gold-verified bundles entering training. On Qwen 3.5 35B it adds 9.2 points on tb-core (40.0%→49.2%) and 6.4 on tb-2.0, and 77.1% vs 73.4% on SWE-bench Verified, at matched synthesis token budgets (2.27M-2.88M) and 100 exported environments per method. Code is released.