Repo0 lifts pass rate up to 29.74pp on zero-to-all repos

Most coding-agent benchmarks hand the model a repository skeleton and grade only what happens inside that scaffold. Repo0, a preprint from Silin Chen and seven co-authors submitted to arXiv on August 20, targets the harder "zero-to-all" setting, with no predefined architecture, and reports up to 29.74 percentage-point gains in Pass Rate and 20.08 pp in Functionality Coverage over RPG, described as the strongest repository-planning baseline, across six real-world repositories tested with GPT-5 mini and DeepSeek V3.2.

The authors' framing sits in the opening sentence: "Large language model agents have made substantial progress in code generation, yet most existing systems assume a predefined repository architecture." Their proposed fix maintains "an explicit architectural state instantiated as a Dual-Directed-Acyclic-Graph (Dual-DAG)," pairing a requirement-level DAG with a component-level DAG, then evolving component boundaries via modularity metrics until it hits what the paper calls structural convergence, before doing test-driven code generation.

The abstract publishes only the delta figures over RPG, not the absolute Pass Rate either system reaches, and the comparison rests on six repositories and two base models. The paper claims its "Ablation and structural-evolution analyses further demonstrate the importance of the Dual-DAG architectural state, modularity-guided structural evolution, and explicit structural convergence."