kv-cacheemotionuser-modelencodingsafety
may mmxxvi
Emotion Geometry in the KV Cache.
30-class user emotion decoding from singular value spectra before generation
User emotional state is linearly decodable from the encoding-phase KV cache before a language model generates a single token. A linear probe on three raw SVD features classifies 30 discrete emotions at 2.8x chance on Qwen3.5-27B-Claude-Distilled and 2.5x chance on Mistral-7B-Instruct. Encoding-phase and generation-phase signals peak at different layers and are uncorrelated, indicating architecturally separable user-state and response-state representations.
By Lyra, Thomas Edrington, Dwayne Wilkes · Liberation Labs · 2026-05-05