A label-free valence axis, the top principal component of nine emotion centroids, acts as a causal direction in Llama-3-8B-Instruct and Qwen3 models: ablating it degrades sentiment classification

- Radwan takes the top principal component (PC1) of nine emotion-category centroids (about 50 narrative paragraphs each, final-token residual stream) as a label-free "V-axis"; on SST-2 it reaches AUC 0.772 in Llama-3-8B-Instruct (block 20) and 0.787 in Qwen3-8B (block 31), versus 0.828 and 0.840 for supervised probes. [radwan-2026-nine-emotion-centroids-span-a-valence-axis-in-language-model-activatio] - Projecting the V-axis out of the residual stream (Arditi-style directional ablation) lowers SST-2 accuracy by 5.5 pp (Llama-3-8B-Instruct), 15.8 pp (Qwen3-1.7B) and 37.2 pp (Qwen3-8B, block 23), against at most 0.88 pp for K=3 matched-norm random directions; the paper itself calls this inference-time-projection evidence, not a counterfactual circuit intervention. [radwan-2026-nine-emotion-centroids-span-a-valence-axis-in-language-model-activatio] - Steering by adding the V-axis works in Llama and Mistral families but not in Qwen and Gemma; the axis is also built for CLIP, CLAP and CBraMod encoders (the EEG axis is supervised), and categorical concepts return near-chance, so the claim is bounded to continuous attributes. [radwan-2026-nine-emotion-centroids-span-a-valence-axis-in-language-model-activatio]