Injecting repository knowledge into code LLMs via adapters — "Code2LoRA," keeping up with evolving code

This work (Code2LoRA) efficiently injects repository-specific knowledge into code language models (LLMs).

Code LLMs need repository-level context to resolve imports, APIs, and project conventions. Existing methods inject this either (1) as long inputs (retrieved via RAG or dependency analysis) or (2) through per-repository fine-tuning and LoRA — both costly at repository scale and brittle to evolving codebases.

Code2LoRA, a hypernetwork framework, generates repository-specific LoRA adapters, injecting repository knowledge with zero inference-time token overhead. It supports two scenarios: Code2LoRA-Static converts a single repository snapshot into an adapter, suited to comprehension of stable codebases; Code2LoRA-Evo maintains an adapter backed by a GRU hidden state updated per code diff, suited to active development of evolving codebases.

To evaluate against parameter-efficient fine-tuning baselines, the authors built RepoPeftBench — 604 Python repositories with a static track (40K training, 12K test assertion-completion tasks) and an evolution track (215K commit-derived training, 87K test tasks). On the static track, Code2LoRA-Static achieves 63.8% cross-repo and 66.2% in-repo exact match, matching the per-repository LoRA upper bound; on the evolution track, Code2LoRA-Evo achieves 60.3% cross-repo exact match (+5.2 pp over a single shared LoRA).