LiveCodeBench official results or credible AI research publications

Scale AI's PlanSearch Boosts Claude 3.5 Code Generation to 77.0%

Sep 8, 2024, 03:06 PM

Scale AI, in collaboration with the California Institute of Technology, Northeastern University, and Cursor AI, has introduced a new state-of-the-art (SOTA) test-time compute method called PlanSearch. This algorithm enhances diversity and efficiency in large language model (LLM) code generation by creating high-level plans in natural language that guide the coding process. The method has shown significant improvements in creativity, diversity in solutions, and the quality of code generated. Notably, Claude 3.5, using PlanSearch, achieved a pass@200 of 77.0% on LiveCodeBench, outperforming the best score achieved without the search algorithm (pass@1 = 41.4%). The optillm lib implements the core idea of PlanSearch, optimizing inference proxy. This development, involving researchers E Wang, F Cassano, C Wu, and Y Bai, represents a significant leap in optimizing inference capabilities and performance of LLMs.