EN한
← wiki

$ cat wiki/papers/2026/2611.21152-repofactory.md

RepoFactory

paperupdated 2026-11-19created 2026-11-19

TL;DR

RepoFactory converts repositories into verifiable environments where coding agents build complete projects from a language specification. Its Environment Evolution Loop optimizes training environments using proxy feedback. (source)

Authors & Org

Yang Tian, Tianyu Zheng, Xiyuan Yang, Yaxin Du, Chengran Yang, Bo Zhao, Jian Yang and Bryan Dai. Affiliations are not established by the reviewed abstract record. (source)

Method

The evolution objective multiplies validity, structural coupling and implementation scope. Proxy signals guide environment refinement without waiting for a full train-and-evaluate cycle. The pipeline produces 10,159 environments across 17 software domains; training Qwen3.5-35B-A3B yields RepoHorizen-35B-A3B. (source)

Results

Authors report gains of 22.6 percentage points on NL2Repo-Bench and 8.5 points on average across 5 benchmarks covering long-horizon coding and agentic tasks, versus 17.5 and 6.4 points without environment evolution. These are reported improvements, not final absolute scores; the abstract does not fully specify the comparison protocol. (source)

Significance

The contribution to Agentic Reinforcement Learning is treating training-environment construction as an optimization problem. For Agents (LLM Agents), project completion provides a broader training target than isolated code patches. (source)

Open Questions

This is an abstract/metadata review, not a full-paper or code evaluation. Proxy gaming, contamination controls, construction cost and generalization beyond the reported benchmarks remain unassessed here. (source)

Cite

  • arXiv:2611.21152v2, first submitted 2026-11-16, revised 2026-11-17, captured 2026-11-19. The curation date reflects the revision, not the original submission. (source)

Referenced by

Sources