SWE-sweep tests whether coding agents can find and fix bugs without hints
Its creators say agents tackle 4.1k bugs across 100 repositories, with no issue descriptions or bug locations provided.
TLDR
SWE-sweep’s creators describe a benchmark that asks coding agents to discover and repair as many bugs as possible in real repositories, without bug-specific hints. Fixes are checked against hidden tests and the repositories’ original test suites. A post linking to the site asks about SWE-sweep as a reinforcement-learning training environment; the site presents it as a benchmark.
Combined views
—
1 Source, first seen ago
