S*: Test Time Scaling for Code Generation

2025-02-20

AI Summary

This paper addresses the underexplored problem of computational resource scaling at test time for code generation. We propose S*, the first hybrid test-time scaling framework tailored for code generation, integrating both parallel and sequential scaling strategies. S* introduces execution-feedback-driven adaptive input generation, pairwise comparison-based solution selection, and dynamic solution filtering. Our key contributions are: (1) establishing the first hybrid test-time scaling paradigm for code generation; (2) designing an execution-verified dynamic ranking and filtering mechanism; and (3) demonstrating—on LiveCodeBench—that non-reasoning models (e.g., a 3B-parameter model) surpass GPT-4o-mini, while GPT-4o-mini+S* improves over o1-preview by 3.7%, and DeepSeek-R1-Distill-Qwen-32B+S* achieves 85.7%, approaching o1(high)’s 88.5%.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Search and Optimization: Sampling/Simulation-based Search
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Security and Privacy: Large-scale security measurements

Abstract

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.

Problem

• Improves code generation coverage
• Enhances selection accuracy
• Boosts performance across models

Innovation

• Hybrid test-time scaling framework
• Adaptive input generation mechanism
• Execution-grounded solution identification