Kaomoji emotion probe validated: 0.68 correlation with valence direction, 0.79 when trained

Sauers_ · x · 2026-11-21

The full writeup of the Qwen3-8B kaomoji experiment with three validation numbers: the model's self-rating of a conversation correlates 0.52 (rank) with its next kaomoji; directly projecting layer-18 activations onto a valence direction reaches 0.68; and a probe trained on kaomoji ratings predicts them at 0.79. The setup induces "moods" in real chats (good or bad news first), then reads them out via a prefilled "I'm feeling" + kaomoji. Conclusion: the model's kaomoji choices reliably reflect an internal emotional state that's linearly readable from activations — a zero-cost, reproducible introspection probe.

Related event: Emoji Probes Quantify Qwen3-8B's "Happiness"(3 posts)→

Original post →

More from Models

Models channel →