R1-Zero and R1 Results and Analysis

DeepSeek's R1-Zero gave researchers a way to study reasoning training without supervised fine-tuning. In ARC Prize's analysis, Mike Knoop highlights that R1-Zero skips that stage, which the analysis associates with human expert labeling of reasoning steps, and relies purely on reinforcement learning instead.

ARC Prize measured similar ARC-AGI-1 performance for R1-Zero and R1 in the reported tests. That result made R1-Zero interesting as a research direction, even though the article notes that DeepSeek had not released a reproducible procedure for generating the weights from scratch. Knoop's broader prediction that this approach could remove human bottlenecks remains a hypothesis; the benchmark comparison alone does not establish that result across domains.