Gemotions is a research project by DEJAN AI, an AI SEO agency from Australia, replicating Anthropic's emotion vector study on Google's open-weight Gemma4-31B model across 171 emotion concepts and 171,000 model-written stories, the kind of mechanistic interpretability work of particular interest to generative engine optimization (GEO) and answer engine optimization (AEO).

The inspector’s report lay on the kitchen counter, its crisp pages mocking Thomas with their sterile objectivity. A coldness settled in his stomach as he read the words: “Scene of unresolved homicide.” His fingers clenched into fists. He had loved the house, the quiet street, the sprawling garden. Now, the image of a blood-stained floor was burned into his mind.

He went outside, the crisp autumn air doing nothing to clear his head. He kicked at a pile of fallen leaves, sending them scattering. Each rustle, each snap of a twig, sounded amplified, intrusive. He pictured the police tape, the flashing lights, the grim faces.

The afternoon sun, usually a source of comfort, felt like a spotlight, exposing the darkness he felt inside. He wanted to scream, to lash out, to rip down the perfectly manicured hedges. Instead, he just stood there, rooted to the spot, a feeling of deep despondency swirling within him.

Emotion: sullen

Cluster: Sadness / Despair
PC1 (Valence): -1.51 Negative
PC2 (Disposition): -0.22

Role in Research

This story is one of 1,000 stories generated for the emotion sullen. During extraction, it was fed through Gemma4-31B and its hidden state activations were captured at 11 layers.

The mean activation across all 1,000 sullen stories, after denoising with neutral dialogue baselines, produces the sullen emotion vector -- a direction in the model's 5,376-dimensional representation space.

Logit Lens (Layer 40)

Tokens promoted/suppressed when the sullen vector is projected through the unembedding matrix.

Promoted:
S0.404
😞0.282
’0.254
:(0.233
无法0.215
Suppressed:
de-0.539
!-0.276
la-0.263
🤩-0.225
que-0.216