Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-17 · high
Each training task gets its own optimization problem instead of one fixed recipe, and the team released the code.
Why now: The paper's code released alongside it, so the SWE-bench Verified claim is checkable now, not just promised.
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.
Each link below shares sources, entities, or timing with this story.
MCP uses Docker / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (MCP uses Docker); both cover Fixed, SWE; earlier Fixed coverage from 2026-08-05.
FORGE built by Mistral / Shared entities / Earlier coverage
Linked by a graph relationship (FORGE built by Mistral); both cover SWE, Verified; earlier SWE coverage from 2026-07-17.
Linked by a graph relationship (FORGE built by Mistral); both cover SWE, Verified; earlier SWE coverage from 2026-06-02.
Linked by a graph relationship (FORGE built by Mistral); both cover SWE, Verified; earlier SWE coverage from 2026-03-01.
Docker supports Claude Code / Shared entities / Same source domain / Earlier coverage / Tension
Linked by a graph relationship (Docker supports Claude Code); both cover Docker, SWE; reported by the same outlet (arxiv.org).
Docker supports Claude Code / Shared entity: Code / Shared topic / Earlier coverage
Linked by a graph relationship (Docker supports Claude Code); both cover Code; overlapping topics (budget, code).
Docker supports Claude Code / Shared entity: SWE / Same source domain / Earlier coverage
Linked by a graph relationship (Docker supports Claude Code); both cover SWE; reported by the same outlet (arxiv.org).
Docker supports Claude Code / Shared entity: Code / Same source domain / Earlier coverage
Linked by a graph relationship (Docker supports Claude Code); both cover Code; reported by the same outlet (arxiv.org).