|
|
|
|
|
|
by slogansand
206 days ago
|
|
|
Author here. Quick context on the benchmarks: We used RARV (Retrieve, Analyze, Reason, Validate) pattern with multi-agent collaboration. Each problem gets worked by specialized agents, reviewed by 3 parallel reviewers (code, business logic, security), and only merged after consensus. The 99.67% isn't cherry-picked. Full run against standard SWE-Bench dataset. Happy to share methodology if anyone wants to reproduce. |
|
|