Scaling Reasoning Tokens via RL and Parallel Thinking: Log-Linear Scaling Law for Competitive Programming

arXiv·high signal

ByteDance Seed, Princeton, UC Berkeley, and Stanford researchers discover an approximately log-linear relationship between validation accuracy and average reasoning tokens during RL training on competitive programming tasks. They introduce a multi-round parallel thinking pipeline that distributes token budgets across threads and rounds of generation, verification, and refinement, training end-to-end to match test-time structure. Two techniques shift the training trajectory: verification RL warmup raises the starting point, while randomized clipping steepens the scaling curve.