IC-1401GPT-4 serves as a proxy for human judgment on SOTOPIA-EVAL, with strong correlations on goal, financial, and relationship dimensions for model outputs
Xuhui Zhou, Hao Zhu, Leena Mathur, Ruohong Zhang, Haofei Yu, Zhengyang Qi, Louis-Philippe Morency, Yonatan Bisk, Daniel Fried, Graham Neubig, Maarten Sap
The paper prompts GPT-4 to score social interaction episodes along seven dimensions and compares its scores to averaged human annotator scores from Amazon Mechanical Turk. For model-role-played agents, GPT-4's scores show significant Pearson correlations with human scores on goal (0.71), financial (0.62), and relationship (0.56) dimensions. Over 74% of GPT-4 scores fall within one standard deviation of the human score. When humans are role-playing, correlations drop sharply on all dimensions except goal, indicating GPT-4 is a better proxy for evaluating models than humans.
Evidence
correlational
Key metric
Pearson correlations (model output): goal 0.71, fin 0.62, rel 0.56, kno 0.33, bel 0.45, soc 0.33, sec 0.22; >74% of GPT-4 scores within human scores ±σ (σ=2.15)
Caveat
GPT-4 tends to rate higher than humans when it disagrees; correlations drop significantly when evaluating human role-play; LLMs are known to have positional bias and factual inconsistency in evaluation.