S*：代碼生成中的測試時間縮放

S*: Test Time Scaling for Code Generation

February 20, 2025

作者: Dacheng Li, Shiyi Cao, Chengkun Cao, Xiuyu Li, Shangyin Tan, Kurt Keutzer, Jiarong Xing, Joseph E. Gonzalez, Ion Stoica

cs.AI

摘要

增加大型語言模型（LLMs）在測試時的計算資源，在各領域展現出潛力，但在代碼生成領域卻仍未被充分探索，儘管在數學領域已有廣泛研究。本文提出S*，首個混合型測試時擴展框架，顯著提升了生成代碼的覆蓋率和選擇準確性。S*在現有的平行擴展範式基礎上，引入序列擴展，以突破性能極限。此外，它利用一種新穎的選擇機制，自適應地生成用於成對比較的區分性輸入，並結合執行基礎信息，以穩健地識別正確解決方案。我們在12個大型語言模型和大型推理模型上進行評估，結果顯示：（1）S*持續提升不同模型家族和規模的性能，使一個3B模型超越GPT-4o-mini；（2）S*使非推理模型超越推理模型——配備S*的GPT-4o-mini在LiveCodeBench上比o1-preview高出3.7%；（3）S*進一步提升頂尖推理模型——配備S*的DeepSeek-R1-Distill-Qwen-32B在LiveCodeBench上達到85.7%，接近o1（高）的88.5%。代碼將於https://github.com/NovaSky-AI/SkyThought 提供。

English

Increasing test-time compute for LLMs shows promise across domains but remains underexplored in code generation, despite extensive study in math. In this paper, we propose S*, the first hybrid test-time scaling framework that substantially improves the coverage and selection accuracy of generated code. S* extends the existing parallel scaling paradigm with sequential scaling to push performance boundaries. It further leverages a novel selection mechanism that adaptively generates distinguishing inputs for pairwise comparison, combined with execution-grounded information to robustly identify correct solutions. We evaluate across 12 Large Language Models and Large Reasoning Model and show: (1) S* consistently improves performance across model families and sizes, enabling a 3B model to outperform GPT-4o-mini; (2) S* enables non-reasoning models to surpass reasoning models - GPT-4o-mini with S* outperforms o1-preview by 3.7% on LiveCodeBench; (3) S* further boosts state-of-the-art reasoning models - DeepSeek-R1-Distill-Qwen-32B with S* achieves 85.7% on LiveCodeBench, approaching o1 (high) at 88.5%. Code will be available under https://github.com/NovaSky-AI/SkyThought.