Collaboration Is All You Need: LLM Assisted Safe Code Translation
Abstract
Talk presenting UniTranslator, a multi-agent framework for safe code translation.
Key points:
- Re-imagining code translation as collaborative multi-LLM endeavor
- Natural Language Inference (NLI) grounding for translation safety
- Handling low-resource languages like Move
- Mitigating code artifacts and hallucinations through iterative feedback
Takeaway: Multi-agent collaboration with small LLMs can rival large monolithic models for code translation, especially for domain-specific and low-resource languages.
