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 sunlight streaming through the window seemed to mock Mark’s unease. He examined the drawing, his brow furrowed, a crease forming between his eyebrows. It was a crude depiction of a knight in shining armor, complete with a ridiculously oversized sword. The knight, named "Sir Sparkle," was a figure he'd conjured up to combat his fear of the dark.

He cleared his throat, trying to sound normal. “That's a very… interesting knight, Lily-May. Where did you get the idea?” His voice cracked slightly.

Lily-May simply shrugged and continued to colour. The knight's armour, he noticed, was exactly the same shade of silver he'd always imagined. A flush crept up his neck. What was happening?

Emotion: perplexed

Cluster: Surprise / Confusion
PC1 (Valence): -1.42 Negative
PC2 (Disposition): -0.16

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
🤔0.330
L0.300
🤔0.296
S0.273
que0.264
Suppressed:
மகி-0.227
rest-0.222
ecstatic-0.212
alleviate-0.210
uplifting-0.196