Research map / Large language models / Agents, code and math
Bridging Functional Correctness and Runtime Efficiency Gaps in LLM-Based Code Translation
Longhui Zhang, Jiahao Wang, Chenhao Hu, Bingyu Liang, Jing Li, Min Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add efficiency-oriented prompting (Correctness+Efficiency) instead of Correctness-Only | Qwen3-Next-80B / GPT-5 · CodeNet+F2SBench+SwiftBench | CA (%) and ET (ms): GPT-5 CA 86.4 -> 79.8, ET 528 -> 453 | Mixed |
| use SwiftTrans on Qwen2.5-7B instead of F2STrans | Qwen2.5-7B · CodeNet+F2SBench+SwiftBench | CA (%): 84.6 -> 90.2 | Improved |
| use SwiftTrans instead of F2STrans on Qwen2.5-7B | Qwen2.5-7B · CodeNet+F2SBench+SwiftBench | ET (ms): 731 -> 292 | Improved |
| increase demonstrations k from 0 to 1 (ablation) | MpTranslator parallel ICL · SwiftBench | CA / ET: +1.3% CA, -142 ms ET | Improved |
| increase demonstrations k from 1 to 3 (ablation) | MpTranslator parallel ICL · SwiftBench | CA / ET: +0.7% CA, -18 ms ET | Improved |
| increase acceleration layers n from 0 to 4 (ablation) | Hierarchical guidance · SwiftBench | ET / CA: -425 ms ET, -0.4% CA | Mixed |
| train with tgt^0 only supervision instead of hierarchical guidance (ablation) | MpTranslator · SwiftBench | ET (ms): ET = 860 ms | Hurt |
| train with tgt^n only supervision (ablation) | MpTranslator · SwiftBench | CA (%): CA = 85.1% | Hurt |
| use bubble selection instead of all-pair selection (ablation) | DiffSelector · SwiftBench | CA / ET: CA 89.1 -> 88.9, ET 439 -> 448 | No clear effect |
| add ordinal guidance loss L_og (ablation) | DiffSelector · SwiftBench | CA (%): 86.1 -> 88.9 | Improved |
| add diff information to judge input (ablation) | DiffSelector · SwiftBench | Order Sensitivity (%): 27.5 -> 6.4 | Improved |
| add bi-judge training (ablation) | DiffSelector · SwiftBench | Order Sensitivity (%): 18.7 -> 6.4 | Improved |
Closest papers by meaning
- Function-to-Style Guidance of LLMs for Code Translation (ICML 2025 · similarity 0.90)
- How efficient is LLM-generated code? A rigorous & high-standard benchmark (ICLR 2025 · similarity 0.90)
- EffiCoder: Enhancing Code Generation in Large Language Models through Efficiency-Aware Fine-tuning (ICML 2025 · similarity 0.90)
- EffiBench-X: A Multi-Language Benchmark for Measuring Efficiency of LLM-Generated Code (NeurIPS 2025 · similarity 0.88)
- Unsupervised Evaluation of Code LLMs with Round-Trip Correctness (ICML 2024 · similarity 0.88)
- Automatically Testing Functional Properties of Code Translation Models (AAAI 2024 · similarity 0.88)
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.
