Megadose The AI curve, compressed.

From Repositories to Training Environments: Scaling Long-Horizon Coding Agents with Environment Evolution Loop

· HF Daily Papers ·
Environment evolution raised the reported coding benchmark gain to 22.6 percentage points, versus 17.5 without it.

RepoFactory converts real repositories into verifiable tasks in which agents build complete projects from language specifications. Its evolution loop uses proxy feedback on validity, structural coupling and implementation scope to improve the environments. The pipeline produced 10,159 environments across 17 software domains for training Qwen3.5-35B-A3B. The resulting model gained 22.6 points on NL2Repo-Bench and averaged an 8.5-point gain across five benchmarks; training without evolution yielded gains of 17.5 and 6.4 points, respectively.

HF Daily Papers' note

score 5

Categories: Research