Meta and partners release SWE-sweep to test proactive bug finding

On October 3, 2026, a researcher posted on Reddit's LocalLLaMA forum to introduce a new benchmark called SWE-sweep, describing it as work done together with other researchers at Meta, Stanford, Harvard, and UW.

As described in the post, SWE-sweep hands an agent a large codebase and asks it to find and fix as many bugs as it can. The agent's work is then scored against a set of known, hidden bugs in the repository. The author framed the benchmark as a response to how most existing evaluations work: they typically test whether a model can fix a bug that a user has already run into. In the author's view, models should by now also be expected to surface bugs before anyone actually encounters them.

The post identifies the benchmark's designers as the author working with researchers at Meta, Stanford, Harvard, and UW, and presents SWE-sweep itself as the announcement. Beyond that, the material contains no evaluation results, no list of models tested, no scores, and no stated follow-up plans, nor any outside response to the benchmark.