AI21 achieves state-of-the-art on SWE-rebench benchmark
HIGHLIGHTS
- •AI21 achieved a 60.9% issue resolve rate on the SWE-rebench benchmark, surpassing previous results.
- •By rethinking the agent's context extraction phase and reversing the traditional 'enrich then solve' approach, opting for solution generation at scale first, followed by better context extraction using generated candidates.
- •This execution strategy improves the agent's accuracy without increasing cost and serves as a blueprint for architecting more accurate and cost-efficient AI software engineers.
Published:
Official Source (Company Blog)