Using GRPO to Beat o1, o3-mini and R1 at "Temporal Clue"

OpenPipe researchers used Group Relative Policy Optimization to train open-weight language models on a complex deduction task called "temporal clue." Their Qwen 14B and 32B models achieved performance comparable to top proprietary models like Claude Sonnet 3.7, while being over 100x cheaper to run. The training recipe and model weights are now freely available under an MIT license.