arXiv study finds universal emotion axis using just 9 labels

A new paper by Yousef Radwan, posted on arXiv, reveals a surprisingly simple way to find an emotion axis inside modern AI models. Instead of training on thousands of labeled examples, the method embeds nine emotion-anchored story sets (e.g., joy, fear, anger) into a frozen encoder, then takes the top principal direction of the nine averaged embeddings. This so-called V-axis requires only 9 emotion category names and 50 short narrative paragraphs per emotion—about 1,500 fewer labels than a typical supervised approach.

The V-axis proves both transferable and mechanistically real. Projecting inputs onto it captures 93% of supervised SST-2 performance using Llama-3-8B-Instruct (AUC 0.772 vs. 0.828), correlates with human valence ratings on 11,811 EmoSet images at r=0.636, reaches AUC 0.906 on ESC-50 audio (p<2.2e-15), and hits AUC 0.720±0.055 on EEG data from 123 subjects (p<3.65e-8). Ablating the direction collapses sentiment accuracy by 5.5–37.2 percentage points across three LLMs, while random directions cause almost no drop (≤0.88 pp). A simple 2-parameter classifier trained on text labels transfers to images (AUC 0.961), audio (0.764), and brain recordings (0.828) without any target-modality labels. The recipe is specific to continuous attributes like valence—tests on categorical concepts return near-chance—and steering works reliably only in certain model families (Llama/Mistral, not Qwen/Gemma).