Parsed Text Understates Vision-Language Models: From Below to Above Expert Agreement on Fine-Grained Facial Emotion

Abstract

EmoNet-Face-HQ is a benchmark of 2,500 generated portraits, each rated by four psychology experts on 40 emotion categories. It concludes that vision-language models (VLMs) on average offer no dependable improvement over random guessing and that a fine-tuned model is required. The benchmark reads each model's answer from generated text, and its zero-shot prompt demands a fixed number of emotions per face. Its own results already show that this protocol can decide the outcome: one VLM moves from chance to close to expert agreement when only the prompt changes. We keep the images, expert ratings and metric and change only how the answer is asked for and read: one yes/no question per category, scored by the probability of "yes". Scored from parsed text on the five categories experts agree on most, eleven open-weight VLMs average a quadratic-weighted $\kappa$ of 0.425 against individual experts, below the 0.468 at which two experts agree. Scored from yes-probabilities, the same models average 0.553, and all eleven agree with an individual expert better than a second expert does. The VLMs' mean is level with the benchmark's fine-tuned model (0.555), and three of them significantly exceed it without any training. Only one exceeds a stricter reference, the mean of three experts scored against the fourth (0.563). Verdicts that VLMs fail at graded facial affect ratings may therefore describe the readout more than the model. A fine-tuned model should have to beat a pretrained VLM read this way before it is called necessary, and the benchmark's own does not clearly do so.

arXiv commentPreprint. 10 pages, 2 figures; 10-page supplementary material as ancillary file. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible