Latent Structure of Affective Representations in Large Language Models

Researchers investigate how large language models represent emotions in their latent spaces, discovering that LLMs develop coherent emotional representations aligned with established psychological models of valence and arousal. The findings support the linear representation hypothesis used in AI transparency methods and demonstrate practical applications for uncertainty quantification in emotion processing tasks.

This research addresses a critical gap in understanding how LLMs internally structure emotional and affective information, using psychology as a grounding framework. The study leverages the well-established valence-arousal model from affective psychology as a validation benchmark, providing rare ground-truth data for evaluating representational geometry claims. This methodological approach is significant because most geometric analyses of LLM representations lack external validation criteria.

The work emerges from growing recognition that model transparency and interpretability are essential for AI safety. Previous research focused on topological properties without clear validation mechanisms. By anchoring analysis to human emotion models with decades of empirical support, the researchers create a more rigorous testing ground. The discovery that nonlinear emotional representations can be approximated linearly has immediate implications for existing interpretability tools that assume linear structure in embedding spaces.

For AI developers and safety researchers, these findings suggest that LLMs acquire sophisticated, semantically meaningful structure that mirrors human cognitive organization. The ability to quantify uncertainty in emotion recognition tasks opens applications in content moderation, sentiment analysis, and safety-critical systems requiring calibrated confidence estimates. The parallel between learned representations and psychological models hints that LLMs may capture fundamental properties of human language and cognition rather than superficial statistical patterns.

Future work should explore whether these geometric properties extend to other high-level concepts beyond emotions, and whether manipulating representational geometry can improve model behavior. Understanding how safety-critical domains like harm detection are represented internally becomes increasingly important as models deploy in higher-stakes applications.

• →LLMs develop geometrically coherent emotional representations aligned with psychology's valence-arousal model
• →Nonlinear affective representations can be well-approximated linearly, supporting assumptions in transparency methods
• →Latent emotion space structure enables quantification of model uncertainty in emotion processing tasks
• →Research provides validated ground-truth framework for evaluating representational geometry in LLMs
• →Findings carry implications for AI interpretability, safety evaluation, and content moderation systems