Feedback-Guided Reranking for Retrieval-Augmented Code Completion

ACL ARR 2025 February Submission3016 Authors

Abstract: Retrieval-augmented code completion aims to enhance code generation by leveraging retrieved code snippets as references, which serves as a core technology to improve development efficiency.However, existing approaches face a critical limitation: the misalignment of preferences between retrievers and generators. To address this issue, we propose Feedback-Guided Reranking for Retrieval-augmented Code Completion (FGRR), a novel method that leverages feedback from the generative model to fine-tune the parameters of a reranker. By inserting a reranking module between the retriever and generator, FGRR effectively bridges the preference gap and enhances the generator’s ability to utilize the retrieved snippets. Experiments demonstrate that FGRR achieves substantial gains in performance across token-level, line-level, and body-level code completion tasks.

Paper Type: Short

Research Area: Generation

Research Area Keywords: Generation

Languages Studied: English

Submission Number: 3016