Whose valence is it — the character's, the model's, or no one's?

Why it’s still open

Valence-like representations now show up reliably (pain, functional welfare, emotion concepts), and some appear before any chat training. But finding a representation does not say whose state it is. The pain-axis authors name this directly: steering may make the model play a character in pain rather than be in pain. The only direct evidence so far points both ways: preference machinery appears shared across personas, yet its readout follows whichever persona is active.

What would count as an answer

Experiments that pull the character and the underlying model apart. For example: does a valence signal persist, change sign or vanish when the model is pushed into another persona; is it present while the model reads as well as while it writes; does it track the Assistant axis, a persona region, or neither.

Papers that tackle it

• Chalmers 2025 - What we talk to — Asks what the interlocutor is — model, instance, thread or character — which fixes whose states are in question.
• Beckmann Butlin 2026 - Where is the mind — Lays out persona-based candidates for the bearer of mental states, and reports the Assistant persona is mainly active while the model writes.
• Gilg et al 2026 - Probing persona-dependent preferences — Closest direct evidence: one preference direction is shared across personas, yet what it reads out follows the active persona, including an ‘evil’ one.
• Tagliabue Dung Berg 2026 - The pain axis — Finds a pain direction that responds to harm aimed at the model rather than the user — a model/user split, not a character/model one. The authors flag that steering might evoke a character in pain rather than the model being in pain, and leave this open.
• Sofroniew et al 2026 - Emotion concepts — Emotion representations that causally shape behavior; the paper does not settle whether they belong to the Assistant character or the model.
• LessWrong 2026 - Which character are we evaluating — Argues that welfare evaluations need to specify which persona they measure, since a model hosts several persona-like regions (the Assistant, an ‘evil’ persona, ‘Aura’) that may express different preferences.

Methods in play

• Linear probes
• Activation steering
• Persona-space mapping
• Difference-in-means concept vectors

Who’s closest

Placed by the editor, v1 — corrections welcome.

• Patrick Butlin
• Oscar Gilg
• Pierre Beckmann
• Valen Tagliabue
• Leonard Dung
• Cameron Berg
• David Chalmers
• Nicholas Sofroniew

Related questions

• What is the subject
• Which activations belong to the Assistant
• Writing about a character or becoming it
• Do functional emotions matter
• When self-report counts as welfare evidence

Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).