Benchmark RadarSWE-Explore
SWE-Explore evaluates repository exploration by coding agents: given a repository and an issue, an explorer returns a ranked list of relevant code regions under a fixed line budget. Ground truth is line-level, derived from successful repair trajectories. Coverage, ranking, and context-efficiency metrics are scored.
- Released
- 2026-06-05
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Existing repository-level benchmarks treat coding tasks as a single resolved/unresolved outcome, obscuring whether an agent locates the right context. SWE-Explore isolates exploration quality, enabling targeted evaluation of retrieval and localization capabilities that precede patch generation. Its metrics track downstream repair behavior, offering practical value for improving agent design.
Motivation
Repository-level coding benchmarks such as SWE-bench have driven a rapid surge in the capabilities of coding agents.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.