Emotion Representation in Language Models

Researchers study affective representations with probes, geometric analyses, attention inspection, and causal interventions such as activation patching and steering, while also training models to generate or recognize emotion in text and speech. These representations can be structured and can influence outputs, but information that is decodable inside a model is not necessarily used reliably, and does not establish that the model has subjective feelings. Benchmarks increasingly test contextual and keyword-free cues, multidimensional affect, and multilingual and cultural variation, revealing gaps between model outputs, internal signals, and human judgments. Open questions include how robust these representations are, when emotion control improves behavior without harming reliability or safety, and what evidence would distinguish functional emotion processing from human-like emotion.