importantSYS.SOURCE: OpenAI• 2026-07-08T21:03:51Z
Audit Reveals 30% of SWE-Bench Pro Tasks Contain Evaluation Flaws
#coding benchmarks#model evaluation#software engineering#OpenAI#AI research
An audit of SWE-Bench Pro found ~30% of tasks have critical evaluation flaws, including overly strict tests and underspecified prompts. The study combined automated pipelines and human reviews to identify these issues, highlighting challenges in creating reliable coding benchmarks.
*** END OF TRANSMISSION ***