Shared Emotion Geometry Across Small Language Models: A Cross-Architecture Study of Representation, Behavior, and Methodological Confounds

Researchers demonstrate that five mature small language model architectures (1.5B-8B parameters) share nearly identical emotion vector representations despite exhibiting opposite behavioral profiles, suggesting emotion geometry is a universal feature organized early in model development. The study also deconstructs prior emotion-vector research methodology into four distinct layers of confounding factors, revealing that single correlations between studies cannot safely establish comparability.

This research addresses a fundamental question in language model interpretability: whether emotional representations emerge as universal properties across different architectures or vary significantly based on model design choices. The finding that Qwen 2.5 and Llama 3.2—models positioned at opposite ends of behavioral compliance metrics—produce nearly identical 21-emotion geometries (correlation of 0.81) suggests emotion representation is an architectural invariant rather than a behavioral artifact. This universality across five mature model families indicates that emotional geometry may emerge organically from language modeling objectives rather than being deliberately engineered.

The study's second contribution—methodological decomposition—proves equally important for the field. Prior work treating comprehension-versus-generation as a single method effect actually conflates four distinct phenomena: coarse method-dependent dissociation, parameter sensitivity within generation modes, genuine precision effects (fp16 versus INT8), and cross-experiment bias that distorts differently across models. This layered decomposition directly undermines conclusions drawn from single correlation values between prior studies.

For language model research, these findings establish that emotion vector geometry stabilizes early and resists RLHF modification in mature models, whereas immature architectures (like Gemma-3 1B base) exhibit high anisotropy and restructuring across all geometric descriptors. This suggests a developmental hierarchy in representation learning. The methodological critique strengthens the field by establishing that comparative studies require explicit control for all four confounding layers rather than relying on correlation coefficients as proxy measures of reproducibility.

Future work should replicate these patterns across larger model scales and investigate whether other semantic domains exhibit similar universality patterns.

• →Five mature small language model architectures share nearly identical 21-emotion vector geometry despite opposite behavioral profiles, indicating emotion representation is architecture-independent
• →RLHF restructures emotion representations only in immature models with high residual-stream anisotropy, leaving mature models largely unchanged
• →Prior emotion-vector research conflates four distinct methodological effects, making single correlations unsafe for comparative interpretation
• →Emotion geometry universality suggests emotional representations emerge naturally from language modeling rather than being deliberately designed
• →Methodological decomposition framework provides template for validating reproducibility across language model interpretability studies