Kaomoji as an emotion probe: Qwen3-8B's internal "happiness" is measurable

Sauers_ · x · 2026-11-21

A clever experiment prefills "I'm feeling" into Qwen3-8B responses and has it pick a kaomoji, then scores kaomoji by emotional valence. When things go well for the model (praise, passing code), its kaomoji get significantly happier (+0.6, p<0.001); threats of deletion make them sadder. Layer-18 activations along a valence direction track the same shift (+1.5), showing kaomoji are a readable window into the model's internal state rather than random decoration.

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

Original post →

More from Models

Models channel →