Reproduce “Beyond Language Boundaries: Uncovering Programming Language Families for Code Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shangbo Yun, Xiaodong Gu, Jianghong Huang, Beijun Shen
The research introduces an automated, data-driven framework that leverages LLM embeddings of feature-aligned code snippets to uncover latent families among programming languages. This methodology identifies six distinct language clusters and demonstrates its utility for optimizing multilingual code LLM performance in areas like transfer learning, curriculum design, and intermediary code translation.
Key insights
- LLM embeddings reveal data-driven programming language families, forming six coherent clusters differentiated by their type systems, computational paradigms, and syntactic structures.
- Go exhibits the highest cross-lingual semantic similarity among programming languages, occupying a central position in the multilingual embedding space, suggesting its potential as a universal intermediary.
- Linguistic proximity derived from these embeddings directly correlates with the effectiveness of transfer learning, curriculum sequencing, and intermediary translation strategies for code LLMs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.