SchrödingerRepo: rewriting repos exposes LLM memorization on SWE-bench

SJTU · hf · 2026-09-24

SJTU researchers introduce SchrödingerRepo, which dynamically instantiates test repositories at evaluation time via four transformation levels (problem restatement, namespace remapping, intra-file reordering, functionality-preserving rewriting) to strip familiar cues from SWE-bench Verified and SWE-QA. Across popular LLMs, removing repository-side cues consistently degrades agent performance and raises interaction costs, mostly from harder exploration and localization — suggesting coding agents partly rely on memorized repo patterns rather than genuine repository reasoning.

Original post →

More from coding & agent

coding & agent channel →