In asynchronous software development, a single change often requires synchronized updates across multiple files, posing a significant challenge for collaborative code generation. Current approaches often overlook the implicit collaborative knowledge embedded in the project’s evolutionary history. To bridge this gap, we propose EvoCoGen, a framework that leverages a Weighted Evolutionary Coupling Graph (WECG). Unlike static analysis, which misses implicit dependencies, and traditional frequency-based methods prone to noise, we introduce Time-Weighted Mutual Information (TWMI). This metric integrates commit recency, focus, and statistical independence to robustly model co-evolution. WECG serves as a dynamic knowledge base, guiding Large Language Models (LLMs) to generate coherent cross-file modifications. Experiments on six open-source and industrial projects reveal that coupling effectiveness is contingent on "commit culture": TWMI excels in projects with skewed/heavy-tailed distributions via its denoising capability, while traditional methods suit stable environments. Moreover, WECG-guided LLMs consistently equal or outperform baselines lacking evolutionary context, particularly excelling in capturing implicit dependencies invisible to static tools. This work offers a socio-technical perspective: by making implicit coordination explicit, LLMs are empowered to not only "understand code" but also "understand collaboration".
Paper
Full text
EvoCoGen: Collaborative Code Generation Guided by Evolutionary Coupling
Semantic Scholar · 2026
Abstract
In asynchronous software development, a single change often requires synchronized updates across multiple files, posing a significant challenge for collaborative code generation. Current approaches often overlook the implicit collaborative knowledge embedded in the project’s evolutionary history. To bridge this gap, we propose EvoCoGen, a framework that leverages a Weighted Evolutionary Coupling Graph (WECG). Unlike static analysis, which misses implicit dependencies, and traditional frequency-based methods prone to noise, we introduce Time-Weighted Mutual Information (TWMI). This metric integrates commit recency, focus, and statistical independence to robustly model co-evolution. WECG serves as a dynamic knowledge base, guiding Large Language Models (LLMs) to generate coherent cross-file modifications. Experiments on six open-source and industrial projects reveal that coupling effectiveness is contingent on "commit culture": TWMI excels in projects with skewed/heavy-tailed distributions via its denoising capability, while traditional methods suit stable environments. Moreover, WECG-guided LLMs consistently equal or outperform baselines lacking evolutionary context, particularly excelling in capturing implicit dependencies invisible to static tools. This work offers a socio-technical perspective: by making implicit coordination explicit, LLMs are empowered to not only "understand code" but also "understand collaboration".