Beyond Code Pairs: Dialogue-Based Data Generation for LLM Code Translation

Beyond Code Pairs: Dialogue-Based Data Generation for LLM Code Translation Le Chen1* Nuo Xu2* Winson Chen3 Bin Lei2 Pei-Hung Lin4 Dunzhi Zhou2 Rajeev Thakur1 Caiwen Ding2 Ali Jannesari3 Chunhua Liao4 1Argonne National Laboratory, Lemont, USA 2University of Minnesota, Minneapolis, USA 3Iowa State University, Ames, USA 4Lawrence Livermore National Laboratory, Livermore, USA lechen@anl.gov, liao6@llnl.gov Abstract Large language models (LLMs) have shown remarkable capabilities in code translation, yet their performance deteriorates in low-resource programming domains such as Fortran and emerging frameworks like CUDA, where high- quality parallel data are scarce. We present an automated dataset generation pipeline fea- turing a dual-LLM Questioner–Solver design that incorporates external knowledge from com- pilers and runtime feedback. Beyond tradi- tional source–target code pair datasets, our ap- proach additionally generates (1) verified trans- lations with unit tests for assessing functional consistency, and (2) multi-turn dialogues that capture the reasoning process behind transla- tion refinement. Applied to Fortran→C++ and C++→CUDA, the pipeline yields 3.64k and 3.93k dialogues, respectively. Fine-tuning on this data yields dramatic improvements in func- tional correctness, boosting unit test success rates by over 56% on the challenging C++-to- CUDA task. We show this data enables a 7B open-weight model to significantly outperform larger proprietary systems on key metrics like compilation success. 1 Introduction Automated code translation has long been a goal in programming languages and systems research, enabling developers to migrate software across lan- guages, frameworks, and hardware platforms (Ah- mad et al., 2021; Feng et al., 2020; Lachaux et al., 2020). In high-performance computing (HPC) and scientific computing, this problem is particularly pressing since legacy codes need to interoperate with modern ecosystems while new frameworks such as CUDA and OpenMP continue to emerge to exploit specialized hardware (Czarnul et al., 2020; Dhruv and Dubey, 2025). Large language mod- els (LLMs) have recently demonstrated impressive capabilities in code understanding and generation, *Equal contribution. motivating their application to code translation. Yet their effectiveness is uneven, with good results in popular languages such as Python and Java, but substantially weaker performance in low-resource domains and specialized frameworks where train- ing data is scarce and structural complexity is high. For example, LLMs often produce translations that compile but fail unit tests or generate code that is syntactically invalid without guidance from exter- nal tools (Chen et al., 2024). Agent-based approaches have been widely adopted to address the limitations of LLMs in code translation. Most existing systems follow a source- to-target workflow (Dearing et al., 2024; Bitan et al., 2025; Chen et al., 2025), where multiple agents iteratively perform translation, verification, and refinement until a syntactically valid target program is produced (Figure 1, up). The result- ing outputs are typically source–target code pairs, which are well suited for traditional supervised learning pipelines (Chen et al., 2021). Although these approaches can produce correct translations, they discard valuable intermediate information that could further enhance the performance of LLMs. These information, such as compiler feedback, run- time diagnostics, and interactions with external tools or scripts, has been shown to be crucial for en- abling deeper reasoning about program semantics and functional behavior (Ding et al., 2024). To address this gap, we introduce a dual-LLM Questioner–Solver design within each LLM agent, which automatically collects and generates reason- ing process data during the agent-based translation workflow (Figure 1, bottom). Instead of producing only final code pairs, our framework automatically generates question-solution data covering interme- diate queries, responses, and tool interactions that occur during translation. The Questioner analyzes the current state, formulates targeted questions, and incorporates external signals such as compiler errors or runtime outputs, while the Solver exe- 1 arXiv:2512.03086v1 [cs.PL] 29 Nov 2025 Source Code Target Code Query Agent       Agent       Agent       Agent       Agent                               Memory Environment Request Memory Action Tools Questions Figure 1: High-level architecture of the Ques- tioner–Solver module in our design to replace the sin- gle LLM core in regular LLM agent frameworks. The Questioner analyzes state and formulates queries using dialogue memory and external tools (e.g., compilers, runtime environments, scripts), while the Solver gen- erates translations, unit tests, and repairs. Their itera- tive interaction enables reasoning separation, external knowledge integration, and progressive refinement of translations. cutes translations

This content is AI-processed based on open access ArXiv data.