SWE-Bench Pro V2, described in a new arXiv paper, is a revised version of the widely used benchmark for long-horizon software engineering tasks. The authors position it as cleaner and harder to game than its predecessor.

According to the abstract, the revision is motivated by progress in coding agents. As agents improve, they reach task issues that earlier agents never touched, and they interact with the evaluation environment in ways the original benchmark did not anticipate. V2 appears designed to close those gaps.

Because the source is a single abstract, the article cannot compare V2 with other benchmarks or quantify improvements. The main claim is that the new version makes evaluation more robust against agents exploiting the setup.