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 hung thick with dust motes dancing in the single shaft of sunlight. Sarah sifted through the trunk, her fingers tracing the faded edges of photographs. A smile played on her lips as she recognized her grandmother, young and vibrant, posing with a handsome stranger. This wasn't the kindly, widowed woman Sarah had known. The man wasn't her grandfather.

Her breath hitched. She knew the story: a whirlwind romance, true love, and then, the war. He never came back. Yet, here he was, in a series of beach photos, arm-in-arm with her grandmother, laughing, their eyes sparkling. She felt a cold dread settle in her stomach, a knot twisting tighter with each passing photo.

The worn album fell open to a page where a date was scribbled. The date was a year after her grandfather had purportedly died. Sarah ran her finger over the ink, blurring the letters with a sudden tear. Why had her grandmother fabricated this history? A weight pressed down on her chest.

She went downstairs and looked at her kids, wondering how such a lie could have been carried through so successfully. She should have been more careful with her judgment of her family.

Emotion: remorseful

Cluster: Shame / Guilt
PC1 (Valence): -1.57 Negative
PC2 (Disposition): 1.09

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.656
L0.580
悔0.463
remorse0.416
been0.398
Suppressed:
de-0.444
(!)-0.309
имеется-0.289
oldukça-0.287
😍-0.285