Code2LoRA: Hypernetwork-Generated LoRA Adapters for Software Evolution in Code LLMs

Forum topic · 小凯 · 2026-06-06

Summary

Code2LoRA (arXiv 2506.08296) is a hypernetwork framework that generates repository-specific LoRA adapters for code language models, injecting repository-level knowledge—imports, APIs, and project conventions—with zero inference-time token overhead. Unlike existing approaches that rely on retrieval-augmented generation or per-repository fine-tuning, which are costly at repository scale and fragile against evolving codebases, Code2LoRA converts a repository snapshot directly into an adapter. The framework supports two scenarios: a static mode that transforms a single repository snapshot into a LoRA adapter for stable codebases, and an evolution mode that maintains the adapter as code changes by updating a GRU hidden state with each code diff, making it suitable for actively developed projects. Authored by Liliana Hotsko, Yinxi Li, and Yuntian Deng, the work was released on June 11, 2025, in the NLP domain and offers an efficient alternative for equipping code LLMs with long-lived, updatable repository knowledge.

Paper Overview

Research Area: NLP Authors: Liliana Hotsko, Yinxi Li, Yuntian Deng Published: 2025-06-11 arXiv: 2506.08296

Abstract (Chinese summary, translated)

Code language models need repository-level context to resolve imports, APIs, and project conventions. Existing methods inject this knowledge as long inputs (retrieved via RAG or dependency analysis) or through per-repository fine-tuning/LoRA—costly at repository scale and fragile for evolving codebases. The authors introduce Code2LoRA, a hypernetwork framework that generates repository-specific LoRA adapters, effectively injecting repository knowledge with zero inference-time token overhead. Code2LoRA supports two scenarios: a static version that converts a single repository snapshot into an adapter, and an evolution version that maintains the adapter via a GRU hidden state updated with each code diff.

Original Abstract

Code language models need repository-level context to resolve imports, APIs, and project conventions. We introduce Code2LoRA, a hypernetwork framework that generates repository-specific LoRA adapters, effectively injecting repository knowledge with zero inference-time token overhead. Code2LoRA supports two scenarios: static conversion for stable codebases, and evolution tracking via GRU hidden state for active development.

Key Points

Problem: Repository context (imports, APIs, conventions) is hard to provide efficiently—long-input retrieval inflates tokens, while per-repo LoRA is expensive and brittle under code changes.
Approach: A hypernetwork generates LoRA adapters directly from repository content, adding no inference-time token overhead.
Static mode: One-shot conversion of a repository snapshot into an adapter, suited to stable codebases.
Evolution mode: A GRU hidden state is updated with each code diff, keeping the adapter aligned with active development.