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 realtor’s words hung in the air, thick and heavy. “It’s a lovely home, of course. Just…had some history.” Sarah stared at the pristine white walls of her newly purchased dining room, the sunlight slanting across the polished hardwood floor. A history? She’d known about the previous owners. Divorced couple, quick sale. Nothing more than that. Now, a cold dread coiled in her stomach, a physical pressure. She ran a hand over the smooth, unblemished tabletop, a shiver tracing its path.

“What kind of history?” she asked, her voice sounding thin, distant even to her own ears. The realtor, a woman with a too-bright smile, hesitated.

Her vision tunneled. She remembered the house’s listing, the photos showcasing its features: a renovated kitchen, a cozy fireplace, a spacious backyard. Now, those details seemed to mock her. She felt like a stranger in her own house.

Emotion: bewildered

Cluster: Surprise / Confusion
PC1 (Valence): -1.87 Negative
PC2 (Disposition): -0.48

Role in Research

This story is one of 1,000 stories generated for the emotion bewildered. 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 bewildered stories, after denoising with neutral dialogue baselines, produces the bewildered emotion vector -- a direction in the model's 5,376-dimensional representation space.

Logit Lens (Layer 40)

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

Promoted:
будто0.300
ness0.255
像是0.241
worse0.231
Worse0.219
Suppressed:
la-0.373
மகி-0.229
happy-0.226
ecstatic-0.226
K-0.219