A new preprint on arXiv, titled "Emotion Is Not External Knowledge: Internal Attention Signals for Affective State Reasoning," examines how large language models handle implicit conversational states. The abstract notes that LLMs increasingly operate in multi-turn interactions where they must maintain latent states beyond explicit facts. The authors focus on affective state as one such implicit state, inferred from sparse cues.

The paper's central claim, as reflected in its title, is that emotion should not be treated as external knowledge to be retrieved or appended. Instead, the authors propose using the model's own internal attention signals to reason about affective states. This suggests a shift from augmenting LLMs with outside emotional context to leveraging what the model already encodes internally.

The abstract is brief and does not include experimental details or results, so the specific method and evaluation remain unclear from this source alone. However, the framing is notable: it positions affective reasoning as a problem of interpreting internal representations rather than acquiring additional information. Further reading of the full paper would be needed to assess the evidence behind this approach.