Adaptive Critical Token-Aware Retrieval for Repository-Level Code Generation
ACToR retrieves repository context only when generation reaches decisive “critical token” positions.
The paper argues that many code-generation failures start at a small number of tokens that send the rest of the output down the wrong semantic path. ACToR detects those points during autoregressive generation and triggers targeted retrieval instead of treating retrieval as static task-level support. It also adds position-aware weighting for dense retrievers. On RepoExec and CoderEval, the authors report relative gains of 8.4% and 15.4% over state-of-the-art methods. ArXiv · AI/CL/LG's note
The paper argues that many code-generation failures start at a small number of tokens that send the rest of the output down the wrong semantic path. ACToR detects those points during autoregressive generation and triggers targeted retrieval instead of treating retrieval as static task-level support. It also adds position-aware weighting for dense retrievers. On RepoExec and CoderEval, the authors report relative gains of 8.4% and 15.4% over state-of-the-art methods. ArXiv · AI/CL/LG's note
score 4