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 attic air tasted of dust and forgotten things. Sunlight, sliced into the gloom by a single, grimy window, illuminated motes dancing in the air. He coughed, a dry rasp that echoed in the high-ceilinged space. Boxes, stacked haphazardly, held the flotsam and jetsam of a life lived. He’d come up here looking for his old baseball glove, but his fingers, instead, traced the faded label on a cardboard container: "Michael's School Years."

He hefted the box, the cardboard groaning in protest. He carefully lifted the lid. The contents were surprisingly organized. Manila folders, neatly labeled, chronicled his academic life: “Kindergarten – Art Projects,” “Grade 3 – Book Reports,” “High School – Term Papers.” He felt a strange tightening in his chest.

He pulled out a folder, its edges softened with age. Inside, a crudely drawn picture of a lopsided house, colored with crayon. He remembered drawing it. His mom had always praised his art, even the terrible stuff. He wondered if this was the reason. A ghost of a smile touched his lips.

Emotion: reflective

Cluster: Nostalgia / Reflection
PC1 (Valence): -1.27 Negative
PC2 (Disposition): -1.24

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.650
L0.547
如今0.402
H0.353
own0.319
Suppressed:
la-0.687
de-0.659
l-0.388
/-0.324
--0.320