SimPO (Simple Preference Optimization) is a simple yet effective offline preference optimization method designed to address the train-inference mismatch and computational inefficiency inherent in Direct Preference Optimization (DPO). Its core innovation lies in using a length-normalized reward function — the average log-likelihood of the response generated by the policy model serves as the implicit reward — thereby directly aligning with the metric used during generation. It also introduces a target reward margin to ensure sufficient separation between winning and losing responses. This approach not only eliminates the need for a reference model, significantly reducing memory and computational costs, but also demonstrates substantial performance gains across multiple benchmarks: up to 6.4 percentage points over DPO on AlpacaEval 2, and up to 7.5 percentage points on Arena-Hard. SimPO models trained on Gemma-2-9B-it achieve a 72.4% length-controlled win rate on AlpacaEval 2, ranking first among sub-10B-parameter models on Chatbot Arena, demonstrating that it can more effectively leverage preference data to improve alignment while maintaining stable response lengths.