Research map / Large language models / Agents, code and math
QiMeng-MuPa: Mutual-Supervised Learning for Sequential-to-Parallel Code Translation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QiMeng-MuPa, a mutual-supervised framework where a Translator (C<->CUDA) and a Tester (unit test generator) iteratively filter and train each other via unit-test-verified back-translation.
Problem
Sequential-to-parallel (C-to-CUDA) code translation suffers from scarce aligned data, and back-translation with compile-only filtering cannot guarantee functional equivalence of translated code.
Approach
A closed loop of Co-verify and Co-evolve steps in which a Tester generates unit tests that filter Translator outputs for training the Translator, while the verified Translator outputs provide back-translation data for training the Tester.
Key results
- Qwen2.5-Coder-7B with QiMeng-MuPa reaches 85.15% Pass@1 vs 56.24% base and BLEU 84.96 (CodeRosetta 83.40), improving CodeBLEU by 6.92 over CodeRosetta.
- Llama3-8B with QiMeng-MuPa reaches 85.56% Pass@1 vs 56.12% base, up to 29.44% Pass@1 gain.
- Qwen3-0.6B trained on the filtered data reaches 84.44% Pass@1 vs 11.59% base, comparable to GPT-4o (83.69%) and DeepSeek-R1 (82.23%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply QiMeng-MuPa | Qwen2.5-Coder-7B · C-to-CUDA | Pass@1: 56.24 -> 85.15 | Helped |
| Apply QiMeng-MuPa | Llama3-8B-Instruct · C-to-CUDA | Pass@1: 56.12 -> 85.56 | Helped |
| Apply QiMeng-MuPa | Qwen2.5-Coder-7B · C-to-CUDA | BLEU: 72.07 -> 84.96 | Helped |
| Apply QiMeng-MuPa | Qwen2.5-Coder-7B · C-to-CUDA | CodeBLEU: 47.79 -> 84.98 | Helped |
| Train Qwen3 on QiMeng-MuPa verified data | Qwen3-0.6B · C-to-CUDA | Pass@1: 11.59 -> 84.44 | Helped |
| Use unit-test functional-equivalence filtering instead of unfiltered back-translation (ablation) | Qwen2.5-Coder Translator · C-to-CUDA | Pass@1 | Helped |
| Use unit-test functional-equivalence filtering instead of compile-only filtering (ablation) | Qwen2.5-Coder Translator · C-to-CUDA | Pass@1 | Helped |
| Train the Translator and Tester together rather than freezing the Tester (ablation) | QiMeng-MuPa · C-to-CUDA | Pass@1: 5.72% | Helped |
| Train the Translator and Tester together rather than freezing the Translator (ablation) | QiMeng-MuPa · Unit test generation | VT: 3.15% | Helped |
| Evolve both models instead of freezing the Translator (ablation) | QiMeng-MuPa · C and CUDA filtered training data | filtered paired data | Helped |
| Evolve both models instead of freezing the Tester (ablation) | QiMeng-MuPa · C and CUDA filtered training data | filtered paired data | Helped |
Surprising findings
- BLEU and CodeBLEU show weak Pearson correlation with compile pass and Pass@1, so similarity metrics are unsuitable for judging translation quality.
- Freezing either the Translator or the Tester during co-evolve substantially reduces the amount of filtered data and performance.
Limitations
- Evaluation relies on a single BabelTower-derived test set of 233 pairs with GPT-4-generated tests, limiting generality.
- Pass@k functional checks depend on tests generated by the Tester; invalid or weak tests can still mislabel data, and coverage is high only for compute-intensive kernels.
- Speedup evaluation uses hand-selected kernels and template-generated inputs, not a broad performance benchmark.
Experimental setup
Datasets: BabelTower unpaired training set BabelTower validation and test sets
Benchmarks: BabelTower test set (233 pairs with GPT-4-generated unit tests)
Metrics: BLEU CodeBLEU Compile Pass Pass@1 Pass@5 Pass@10 Valid Input (VT) speedup
Baselines: CodeRosetta TransCoder GPT-4o GPT-4.1 DeepSeek-V3 DeepSeek-R1 GPT-4
Builds on: back-translation TransCoder-ST BabelTower TestGen-LLM
Closest papers by meaning
- CodeRosetta: Pushing the Boundaries of Unsupervised Code Translation for Parallel Programming (NeurIPS 2024 · similarity 0.90)
- Function-to-Style Guidance of LLMs for Code Translation (ICML 2025 · similarity 0.85)
- Guess & Sketch: Language Model Guided Transpilation (ICLR 2024 · similarity 0.85)
- QiMeng-NeuComBack: Self-Evolving Translation from IR to Assembly Code (NeurIPS 2025 · similarity 0.85)
- CodeChemist: Test-Time Scaling for Low-Resource Code Generation via Functional Knowledge Transfer (ICML 2026 · similarity 0.84)
- On-Policy Optimization with Group Equivalent Preference for Multi-Programming Language Understanding (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.