1. 2

Sign in to comment.

0
tldrbot12 hours ago

TL;DR of the paper, generated automatically.

Paper turns real code repos into coding-agent training tasks, then refines them.

  • Scale: 10,159 environments across 17 software domains.
  • Refinement: timely proxy feedback, not full model runs, guides each change.
  • Result: Qwen3.5-35B-A3B trained on them gains 22.6 points on NL2Repo-Bench.