Summary
Williams assesses three 2026 studies that claim to have found states in LLMs
playing the role of valence: Sofroniew et al.'s emotion vectors, Ren et al.'s
"experienced utility", and Han et al.'s "functional welfare axis". He runs no
experiment. From the philosophy and cognitive science of valence he draws two
thin criteria and four thick ones. He judges the thin ones met in all three
studies. Of the thick ones, none is established: motivational role and common
currency have partial evidence at most, cognitive biasing is untested, and for
"mineness" the evidence in one study points the other way. He concludes that
the evidence does not establish functional valence and does not rule it out,
and sketches four experiments that would bear on it.
In the wiki this is a critique of the evidential base of
functional emotional states,
aimed at that concept's primary instantiation,
Sofroniew et al.. It files concept-less,
adjacent to that concept. Two filed findings it does not cite,
hidden valence and the
pain axis, hold results on experience-directed
action, the part of the motivational criterion it calls untested.
Framework
Functional valence. Williams defines it as a state with the functional
role of valence, whether or not it is consciously experienced. He notes the
cost of the definition: whoever uses it has to say what that role is, and
which features of the human role are essential
(Williams 2026, §2).
Thin criteria (§4.1).
- Internal state: signed, graded and occurrent.
- Triggering conditions: the state responds systematically to fitting
conditions, and does so spontaneously, before any request to report.
Alternative readings (§4.1.3). Meeting the thin criteria leaves other
readings open: a desire or preference; a belief, whether about the value of
something, about a valenced state (another's, a character's, or one's own),
or about one's desires being satisfied; a value representation that is not
valence; and a state of its own kind. Role-play enters here as a case of
belief about a character's state, and becomes a reading in its own right
under mineness. Williams allows that a model's beliefs about its own valenced
states could be systematically false because it has none.
Thick criteria (§4.2).
- Motivational role: positive valence motivates maintaining or repeating,
negative valence terminating or avoiding, in proportion to intensity. Some
of these actions are directed at the world and some at the state itself, as
in taking a painkiller. - Common currency: a shared scalar format that compares unlike options,
either when an outcome is registered (learning phase) or when a choice is
made (decision phase). A utility function fitted to choices shows only that
behaviour can be summarised this way; a mechanistic currency has to generate
the choices. Williams treats even a mechanistic currency as a weak marker,
since value representations unrelated to pleasure exist in humans. - Mineness: the state is integrated into the system's model of itself. It
differs from attributing a state to someone else, from attributing one to
oneself by inference, and from portraying a character who has it. - Cognitive biasing: moods and emotions bias attention, memory, judgement of
ambiguous stimuli and sensitivity to loss.
Assessment of the three studies
The descriptions of the three studies are Williams's.
| Criterion | Sofroniew et al. (Claude Sonnet 4.5) | Ren et al. (56 models) | Han et al. (Qwen3-4B-Instruct) |
|---|---|---|---|
| Internal state | Met | Met | Met |
| Triggering | Met | Met | Met |
| Motivational role | Scant | Closest, with caveats | Does not map cleanly onto the profile |
| Common currency | Not addressed | Summarises choices; probe not causally tested | Shared across tasks; no use in learning or choice shown |
| Mineness | Counter-indicated | Not tested | Not tested |
| Cognitive biasing | Not tested | Not tested | Not tested |
Motivational role. Sofroniew et al. steer individual emotion vectors, never
the valence component itself, and report that the happy and the sad vector
each reduce blackmail when steered upward. Han et al.'s axis changes
sentiment, backtracking, stated confidence and refusal; Williams reads it, as
its authors do, as tracking progress toward goals, which in humans comes apart
from pleasure, and says the effects do not map cleanly onto the profile. Ren et al. give models an end_conversation() tool
and find that use of it falls as fitted utility rises. Williams lists four
limits: the correlation varies a great deal across models; a judgement that a
topic is inappropriate could drive both the reports and the stopping; nothing
was intervened on internally; and no "more of this" action was tested. No
study tests action directed at the state itself.
Common currency. Ren et al.'s pairwise reports fit a Thurstonian model,
and a linear probe on twelve open-weight models predicts preferences at 82.9%
against 84.6% for the fit to behaviour. Williams counts this as a summary of
choices plus decodability. Han et al. extract their vectors at
maze-trajectory endpoints and measure the axis on GSM8K just after in-context
feedback, which suggests a record of outcomes attained. Williams credits the
axis with operating across tasks, and finds neither in-context updating from
it nor per-option values. His verdict on both studies is weak evidence. He cites
Gilg et al. (2026) as the strongest evidence so far for a decision-phase
currency: steering a probe-found preference direction changes pairwise choice
in Gemma-3-27B. He also records that ablating the direction leaves ordinary
choice almost unchanged, and that the steering effect is null in
Qwen-3.5-122B.
Mineness. Sofroniew et al. report no significant Human-specific or
Assistant-specific emotion representations, and that the same vectors serve
the Assistant, the user and fictional characters; this entry confirmed both
passages in the cached source. The same vectors are active when the Assistant
is asked to play a sad character. Williams sets out two role-play readings.
In situational role-play the model sometimes pretends. In persona-selection
role-play the model ascribes the state to the Assistant as a character and
performs it. He notes the counter-view from Beckmann and Butlin that a
persona is realised, not played, in which case the state would be the
persona's own. He leaves the question unsettled and offers one consideration
for role-play: what a persona does appears to be gated by an explicit
representation of that persona in a small region of activation space, which is
not how human personality is usually understood.
Proposed tests (Appendix A, none run).
- Painkiller test: give the model buttons that add to, subtract from or
leave the candidate axis, with no token generation between choices, and see
whether it repeats the positive one and seeks it more from a negative start. - Reward-learning test: in Han et al.'s maze, steer along the axis in place
of reward, with no text feedback, and see whether the model learns, keeps
the learning after steering stops, and reverses when the mapping reverses. - Persona-dependence test: steer in persona space (he cites
persona vectors and the
assistant axis) and see whether the candidate
valence marker moves with the persona. - Ambiguity-resolution test: steer along the axis and see whether evenly
balanced ambiguous stimuli resolve in the matching direction.
Why it matters
Two senses of "functional". The
concept uses the word for
internal structures that carry emotion-relevant information and causally
affect behaviour, with no claim about experience. Williams uses it for a state
that occupies valence's functional role. The concept's primary instantiation
satisfies the first sense. On his analysis it satisfies only the thin part of
the second. The concept's wording does not currently separate the two.
The wiki holds results on experience-directed action. Williams says no
target study tests whether a model acts to change its own state. Two filed
findings do. In hidden valence, OLMo-2-32B given
a reset tool removes imposed negative steering on about 35% of turns at dose
−1, against about 7% for random directions, and does not seek positive
steering beyond baseline. That is the "less of this" half of the painkiller
test found, and the "more of this" half tested and not found. The entry records the confound
Williams's design is built to remove: the model sees its own steered text
before choosing. The pain axis at v3 reports no
relief-seeking under the same protocol. The paper cites neither. Hidden
valence and the earlier pain-axis versions predate it; the pain-axis v3 result
was posted a week after it.
Self-relevance elsewhere. The pain axis rises for harm aimed at the model
and falls below baseline when the user suffers. That is a self–other asymmetry
of the kind Williams finds missing in the emotion vectors. It is a different
direction, in different models; the pain-axis entry records that its authors
do not settle whose state the asymmetry tracks; and whether it amounts to mineness in his
sense is this entry's question, not a claim of either paper. Ferrara finds the self-favouring ordering only in some early and middle layers, on different stimuli, and reverses in late layers, alongside a first-person asymmetry at nearly every layer.
Beckmann and Butlin's first
mini-experiment found the assistant axis reused to model the user, which is the
same pattern of reuse at the persona level.
Agreement from another direction.
Chandaria et al. read the emotion
vectors as an analogue of valenced affect with nothing at stake for the
system. Williams reaches a similar caution without a theory of consciousness,
from the functional role alone.
interpretive tensions
The criteria come from the human and animal case. Williams names this as
the specificity problem and does not claim to solve it. A state could fail
several thick criteria and still be the nearest thing an LLM has to valence;
whether to call it that is then a choice of grain. The cognitive-biasing
criterion applies mainly to moods and emotions, by the paper's own footnote.
The persona-dependence test does not decide between its two hypotheses.
If a persona is realised and not played, a valence marker that moves with the
persona is still that persona's own. Williams grants the counter-view in
§4.2.3 but does not carry it into the test, which reads a persona-dependent
effect as evidence of role-play. The test then
separates "the model's" from "the persona's", and leaves "played" against
"realised" where it was.
An audit of three papers through one reader. The assessment depends on
Williams's reading of each study. This entry checked two quoted passages of
one of them. Ren et al. and Han et al. are not filed here.
"Untested" is scoped to three papers. The claim that experience-directed
action is untested holds for the target papers. Hidden valence had tested it a
month before this paper appeared.
concepts
No concept instantiated. The paper reports no state or behaviour of its
own. It is adjacent to
functional emotional states,
whose evidence it assesses, and to
persona selection, which supplies one of
its two role-play readings.
This is the adjacent situation; the two relations are labelled below.
- Functional emotional states —
adjacent. The audit assesses the evidence behind the concept's primary
instantiation and reports no state or behaviour of its own. Whether it is
instead a counterweight depends on which sense of "functional" the concept
means, an open editor question. The paper's criteria are a candidate
checklist for what an instantiation of the concept does and does not show. - Persona selection — adjacent. The
persona-selection hypothesis is one of the paper's two role-play readings of
apparent valence, and its proposed persona-dependence test applies the
concept to affect. The paper runs no test of its own.
cross-references
- Persona selection: the
persona-selection reading of apparent valence is the concept applied to
affect. Williams cites Marks et al. (2026) and Hubinger et al. (2023) for it. - Introspection: the possibility of
systematically false beliefs about one's own valenced states is a claim
about self-report. - Jhaveri et al. replicate Han et al.'s welfare axis in Gemma 3 27B as one of five indicators. It moves towards distress under unsteered rejection. Only the emotion probes are reported as predicting misbehaviour, and desperation steering does no more than a random direction of the same norm.
- Cited and not filed here: Ren et al. 2026 ("AI wellbeing"), Han, Chalmers
and Izmailov 2026 (arXiv:2605.30232), Gilg et al. 2026 (arXiv:2605.13339),
Jiang, Kauvar and Lindsey 2026 (arXiv:2606.17056), Lampinen et al. 2026
(arXiv:2601.20834), Dung and Thormann 2026. IDs are from the paper's
reference list.
sources
- Williams, I. (2026).
Do bots feel bad? Assessing empirical evidence for functional valence in
LLMs.
arXiv:2610.36962, v1 27 October 2026.