Envs-FORGE lifts SWE-bench Verified score to 77.1%

Envs-FORGE solves a math program for each training seed to set that seed's next task difficulty, per a paper on arXiv. The technique moved a Qwen 3.5 35B model from 73.4% to 77.1% on SWE-bench Verified, at matched token budgets. Most synthetic-data systems apply one fixed prompting recipe to every seed regardless of whether the model needs a harder task or an easier one.

The method estimates a pass rate for each seed, then scores six possible directions for pushing that seed toward a target difficulty level. It solves a mixed-integer linear program (MILP) to pick the direction.

Then it rewrites the seed's instruction, fixtures, oracle solution, tests, and Docker image together so they stay consistent. Only bundles that pass gold verification make it into the training set.

The gains extended beyond the headline number. Envs-FORGE added 9.2 points on tb-core and 6.4 on tb-2.0. The team released the code, so the approach doesn't stay a paper claim.

One catch stands out. The reported gains are on one model, Qwen 3.5 35B, and the paper doesn't say whether they hold on other architectures. Builders training on synthetic seeds now have a testable alternative to check against their own setup.