Human Evaluation is a key metric for text generation tasks in the fields of large models and reasoning. It involves human judges assessing the coherence, interest, and plausibility of generated content, serving as the most authoritative metric for open-ended narrative, poetry, and humor generation, thereby compensating for the inability of automatic metrics to capture subjective perceptions and creative differences. It is typically reported as a percentage or a score ranging from 0 to 1.

This metric is widely used in relevant benchmarks and papers, effectively reflecting model performance along this dimension. Higher values indicate better performance. It helps researchers and engineers quantify differences in actual model performance in text generation scenarios, facilitating horizontal comparisons across different methods or model versions.