Six Ekman emotion vectors in vision-language models have a low-dimensional geometry (PC1+PC2 explain 55-75% of variance), and a Procrustes-aligned consensus emotion subspace across three VLMs retains selective steering and transfers to held-out Gemma-3

- Xing et al. build six Ekman emotion vectors per model and stimulus source (mean-centered class means with neutral directions projected out, residual stream at about two-thirds depth) in Qwen3-VL-8B-Instruct, InternVL3.5-8B-Instruct and Ministral-3-8B-Instruct-2512; PC1 explains 30-46% of the variance and PC1+PC2 55-75%, and PC1 tracks human valence more stably than PC2 tracks arousal. [xing-etal-2026-emotion-concepts-across-sources-modalities-and-architectures-in-vision] - After generalized Procrustes alignment the shared emotion subspace S_emo is defined as the span of the centered consensus vectors; the original vectors have mean cosine 0.57-0.64 with their projection onto it, projecting Qwen3-VL's vectors onto it keeps their selective steering (diag +5.39 vs +4.96 for the original vector; the orthogonal remainder has a much smaller effect, +1.00), and mapped consensus vectors raise the matched emotion in held-out Gemma-3-12B-it and Gemma-3-27B-it (they do not suppress competing emotions on average). Native matched-emotion cosine between Ministral-3 and Qwen3-VL/InternVL3.5 is near zero (Qwen3-VL vs InternVL3.5 image vectors: 0.71); correspondence for Ministral-3 appears only after alignment. [xing-etal-2026-emotion-concepts-across-sources-modalities-and-architectures-in-vision]