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 scent of lavender and old paper clung to the air in the antique bookstore. James ran his finger along the spine of a well-loved copy of "Treasure Island," a book he and Chloe used to read aloud together, huddled under a blanket in her room. He recalled the specific smell of her room – a mixture of vanilla candles and her mother's perfume. He selected the book, clutching it to his chest.

He'd been hoping for a different outcome. They’d bumped into each other at a mutual acquaintance's wedding, where he was a groomsman and she was a guest. He'd spoken to her, describing their shared history, their inside jokes, their bond. She’d listened patiently, a flicker of confusion in her eyes before she finally shook her head and politely apologized for not remembering.

James walked over to a small wooden bench and sat down. He opened the book, tracing the faded ink of the inscription they’d written together inside: “To Chloe and James, forever." His hand trembled slightly. The silence of the store seemed to press in on him, heavy and suffocating.

Emotion: sentimental

Cluster: Nostalgia / Reflection
PC1 (Valence): 0.89 Positive
PC2 (Disposition): -1.58

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
L0.677
fondly0.567
memories0.503
曾0.488
reminis0.466
Suppressed:
either-0.515
or-0.488
[]-0.473
--0.408
не-0.396