SWE-Serve: end-to-end serving tests reject about a third of agent patches that pass every other test on real SGLang features

arXiv 2609.26777: SWE-Serve·medium signalXBlueskyLinkedInCopy link

arXiv 2609.26777 (22 Sep) builds 53 tasks from recent production SGLang changes, scored with hidden functional, regression and end-to-end serving tests. Across 11 models and 31 model-effort configurations the best reached 75% pass@1. On the 19 tasks with E2E coverage, the pass rate falls from 69.4% to 45.9% once E2E tests count. The practical lesson is that agent patches to infrastructure code need a real end-to-end run in CI, because unit and regression tests miss roughly a third of the failures.