BCV Labs AI Research Roundtable: S* paper
Join us for the 4th BCV Labs AI Research Roundtable, where we bring together research-minded folks to read and discuss a paper, followed by an open discussion with one of the paper authors. This time we will be joined by Dacheng Li, Shiyi Cao, and Xiuyu Li, co-authors of S* : Test Time Scaling for Code Generation
We will dedicate 30-45 minutes at the start of the event to skim the paper in silence, but highly encourage folks to dive in prior to the event for a more thorough read. The goal is to have an open discussion with the author (rather than a one-sided Q&A) so please come prepared to share some thoughts / questions. As usual, dinner will be provided.
Paper: S* : Test Time Scaling for Code Generation (blog here)
What is it about: This paper introduces S* (S-star), a novel hybrid test-time scaling framework that significantly improves code generation performance by large language models. The key innovation of S* is combining parallel scaling (generating multiple code samples) with sequential scaling (iterative debugging) to enhance both coverage and selection accuracy.
Why it matters: With the recent surge in reasoning-enhanced models like OpenAI's o1, o3, Anthropic's Claude 3.7 Sonnet, and DeepSeek's R1, we've seen how additional inference-time computation can unlock powerful capabilities. However, this approach has been predominantly focused on mathematical reasoning tasks, leaving code generation—a domain with enormous practical value—relatively unexplored. The significance of S* lies in its demonstration that even smaller, more affordable models can be transformed into competitive code generators through smart allocation of test-time resources. As automated programming becomes increasingly central to software development workflows, S*'s hybrid approach offers a compelling blueprint for how to extract maximum value from existing models.
Meet The Paper Author
We're lucky to be joined by 3 authors from the paper, Dacheng Li, Shiyi Cao, and Xiuyu Li, who are all second year PhD students at U.C. Berkeley. Dacheng and Shiyi are both student leads for NovaSky, a collaborative initiative led by students and advisors at UC Berkeley’s Sky Computing Lab, aiming to build open-sourced AI models that are SoTA.