Multisage boosts LLM code translation by 2.22x with semantic validation

Large language models (LLMs) show promise for automated code translation, but often rely on token-level patterns rather than deep program semantics, leading to logical and semantic errors in translated code. To address this, Yufu Wang and colleagues proposed Multisage, a multi-semantic augmentation and self-calibration framework that builds rich, verified semantic information directly from source code. The framework consists of three modules: a semantic representation parsing module that extracts data-flow graphs, type constraints, and external API information; a multi-semantic augmentation module that generates code summaries, function-level test cases, and API-oriented descriptions and tests; and a semantic consistency calibration module that uses semantics-preserving mutations and cross-semantic consistency verification to filter and refine the generated semantics.

Evaluated on the HumanEval-X benchmark, Multisage improved translation success rates by up to 2.22 times across diverse backbone LLMs, consistently outperforming vanilla prompting, instruction-tuned LLMs, and Chain-of-Thought reasoning. The largest improvements were observed on smaller models, suggesting that explicit, verified semantic augmentation can bridge the reliability gap between large and small LLMs. By engineering a structured understanding of program semantics into the translation process, Multisage offers a practical path to more dependable code translation without requiring external resources like functional descriptions or test suites.