An uncertainty-aware RAG framework hit 78.13% on repo-level code gen — and a plain verify-and-repair loop matched it

arXiv 2607.24884·medium signalXBlueskyLinkedInCopy link

OpenCoder estimates source-specific uncertainty across API knowledge, repository context, and similar-code evidence, then uses it to filter and rank evidence and to guide generation, verification, and repair. On an expanded 32-task RepoExec-inline evaluation it lifts GPT selected-output correctness from 56.25% to 78.13%, and target-aware API refinement substantially improves API-set retrieval. The authors report honestly that it merely matches a verification-and-repair control and that the Gemini improvement is not statistically supported — a factorial analysis also found no universal ranking of evidence sources, only cross-source interactions that depend on the backend. Build the cheap verify-and-repair loop first.