S*: Test Time Scaling for Code Generation

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 co

Research goal: Does the adversarial robustness gap between DeepSeek-R1 and o1-preview on legal reasoning tasks generalize to code generation benchmarks under negation-based token perturbations?

Autonomous synthesis report generated by SOVEREIGN Research Kernel. Tribunal consensus score: 7.5/10.