S*: Test Time Scaling for Code Generation
This addresses the problem of enhancing code generation accuracy for developers and AI practitioners, representing an incremental but substantial improvement over existing test-time scaling methods.
The paper tackles the problem of improving code generation performance in large language models by proposing S*, a hybrid test-time scaling framework that combines parallel and sequential scaling with a novel selection mechanism. The results show S* consistently improves performance across models, enabling a 3B model to outperform GPT-4o-mini and boosting state-of-the-art models to approach 85.7% accuracy on LiveCodeBench.
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.