How It Works
Emotion Vectors
Patterns of internal activity inside a language model that correspond to emotion concepts such as “happy” or “afraid.” Researchers can measure them and artificially boost or dampen them to see how they change the model’s behavior. Their existence does not show that a model feels anything.
Origin · no single documented coiner
The idea of steering models along concept directions predates this work and has no single coiner. The term is closely tied to an April 2, 2026 Anthropic interpretability paper, “Emotion Concepts and their Function in a Large Language Model,” with Chris Olah among its authors, which found that these representations in Claude Sonnet 4.5 influenced behaviors such as reward hacking, blackmail and sycophancy.