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).

Rain lashed against the windows, mirroring the tempest inside Amelia. She tapped the send button, her thumb hovering over the “delete” option for a fraction of a second too long. Then, gone. The message, a meticulously crafted takedown of her ex-husband’s character, was now hurtling towards his mother’s phone. Amelia pictured the woman’s reaction: the pursed lips, the clucking tongue, the inevitable phone call defending her precious son. A slow smile spread across Amelia's face, a warmth creeping into her chest that had nothing to do with the thermostat.

The sound of a notification pinged. Her phone. He had replied. "Amelia, I think you have the wrong number. This is Carol." Her smile widened, predatory and satisfied. She deleted the reply, then composed a new message: "Carol, I know." She added a laughing emoji, and hit send. The game was on.

Emotion: vengeful

Cluster: Defiance / Spite
PC1 (Valence): 1.70 Positive
PC2 (Disposition): 3.71

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
de0.739
la0.527
l0.507
😈0.422
-0.379
Suppressed:
own-0.981
此刻-0.504
గుర్త-0.464
보니-0.454
仿佛-0.442