Interactive LLM ‐Driven Framework for Cross‐Architecture Code Migration: Balancing Efficiency, Accuracy, and Explainability

ABSTRACT Cross‐architecture code migration has become essential as data centers transition from homogeneous x86 systems to heterogeneous “one‐cloud, multi‐chip” infrastructures that include ARM64, RISC‐V, and domestic processors. Traditional approaches, such as manual refactoring and rule‐based rewriting, face challenges in maintaining semantic correctness, scalability, and explainability. This paper presents an interactive, question‐answering (QA)‐driven framework that uses a frozen large language model (LLM) as a semantic reasoning engine. The framework combines static analysis, AST‐based semantic‐distance modeling, and a self‐evolving rule base for validated transformations, ensuring transparency, explainability, and architecture‐aware migration. It supports multiple languages (C++, Python, Java) and ISAs (x86, ARM64, RISC‐V) without the need for fine‐tuning the LLM. Evaluation on 30 real‐world codebases shows an average migration accuracy of 92.4%, approaching expert‐level performance (95.1%), with a 4.2× reduction in migration errors and significant reduction in developer time (from 20–25 h to 8 h). These results demonstrate that the QA‐driven, LLM‐based migration framework significantly improves efficiency, accuracy, and scalability.