Skills
Retrieving repo context at the specific tokens where generation goes wrong beats task-level RAG by 15%
ACToR observes that repository-level generation errors concentrate at a small number of decisive positions, and once such a token is wrong the rest of the code follows an incorrect semantic path. Instead of front-loading retrieved context at task level, it identifies these critical tokens during decoding and fires targeted retrieval at them, plus a position-aware weighting for the dense retriever. Relative gains are 8.4% on RepoExec and 15.4% on CoderEval. The repo at DeepSoftwareAnalytics/ACToR exists and was pushed September 1 but has 0 stars, so the code is published and unexamined.
↳ Follow the thread