Most code benchmarks treat software engineering as a series of isolated edits. But in practice, features, migrations, and systematic fixes unfold across multiple pull requests (PRs), each depending on the others and requiring coordinated modifications across a repository. The new paper from arXiv (2610.22039) introduces WEAVE, a benchmark and training resource built around this reality.

WEAVE structures tasks according to the dependency relationships between real-world code changes. By weaving actual code evolution into the task design, it aims to give code agents a more faithful picture of how software development happens — and to train them to handle multi-step, repository-wide changes rather than one-off patches.

The authors argue that most existing benchmarks miss this complexity. While the abstract is truncated, the clear motivation is that isolated-change tasks do not capture the coordination needed for real-world maintenance and evolution. WEAVE is positioned as a step toward closing that gap, though the full methodology and results are not detailed in the available text.