CodeClarity: A Framework and Benchmark for Evaluating Multilingual Code Summarization

Abstract

Large Language Models (LLMs) are increasingly used to summarize and document code, yet most research and training data remain limited to English. This creates barriers for developers working in other languages and leaves the multilingual capabilities of LLMs largely unexplored. We present CodeClarity, a framework for evaluating multilingual code summarization across six programming and six natural languages. It combines reference-based metrics, LLM-judge ratings, and faithfulness checks (identifiers and script) to capture surface similarity, semantic adequacy, and code-aware fidelity. Our experiments reveal that lexical metrics penalize morphologically rich languages, while judge-based evaluations provide more stable, semantically aligned assessments. This work establishes the first reproducible foundation for studying multilingual code summarization and points toward fairer, more inclusive evaluation of code intelligence systems. CodeClarity-Bench and the full evaluation pipeline are publicly available at huggingface.co/CodeClarity and github.com/MadhuNimmo/CodeClarity, enabling community-scale human validation and follow-up studies.

Resources

Details