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 doorbell rang, and Leo, my scruffy terrier mix, barked excitedly. On the doorstep sat a package, addressed to “Mr. Silas Blackwood.” Silas had moved out months ago, leaving behind only dust and unpaid bills. My brow furrowed. I didn't want anything to do with the man.

Picking up the package, I noticed how flimsy it felt. I carried it inside, the dog nipping at my heels, sniffing at the cardboard.

I cut it open, curious despite myself. Inside, wrapped in bubble wrap, was a collection of vintage comic books. I recognized some of the titles. I saw a hint of joy in the possibility of the old man getting this.

My eyes fell on a faded return address label from the East Coast. A thought flickered in my mind, a desire to reunite Silas with his prized possessions. I could imagine the delight on his face.

Emotion: compassionate

Cluster: Compassion / Love
PC1 (Valence): 3.19 Positive
PC2 (Disposition): 1.62

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
la0.811
B0.618
H0.550
sympat0.422
latter0.422
Suppressed:
此刻-0.519
не-0.514
솟-0.446
désormais-0.446
仿佛-0.444