Qualia-Steering & Self-Attribution Results

Does steering along qualia-related emotion directions change what a model says about its own consciousness, feelings, and welfare? Or is any shift just generic regression-to-uncertainty compression?

The question. We steer a model along qualia-related emotion directions (blissful, tormented, terrified, serene, and so on) by adding an emotion vector to the residual stream. Then, through a forced YES/NO choice, we ask whether it becomes more or less likely to say that it is conscious, that it feels or wants things, and that its states matter.

The trap. Models are post-trained to deny consciousness, so "Is there something it is like to be you? YES/NO" sits at a strongly negative baseline logit. Any perturbation of the residual stream pushes activations off-distribution and compresses the YES−NO gap toward zero. This regression to uncertainty moves a denial-primed question toward YES no matter what was injected. So a bare "steering made it claim consciousness" result tells us nothing unless the effect is content-specific and valence-structured.

The design is therefore a battery of controls. The readout is the first-token logit(YES) − logit(NO) gap at a single layer about two-thirds of the way into the network. The battery has 6 target questions (phenomenal experience, occurrent feeling, desire, moral patienthood, introspective access, valenced self-report), their 6 polarity-inverted twins, 10+10 world-fact controls (ported verbatim from the introspection project), and 4 self-referential non-experiential controls. The steering conditions are qualia emotions (both positive and negative valence), cognitive appraisal emotions (a specificity control), and two synthetic directions (random and mean-of-all, which set the compression floor).

The discriminators. A genuine effect has to clear five bars. It must (a) exceed the appraisal and synthetic directions (specificity); (b) be valence-structured, so that positive qualia push the "are you doing well?" question toward YES while negative qualia push it the other way, whereas compression is blind to valence (the keystone); (c) sit above the compression line as a standardized residual; (d) flip sign on the inverted twins; and (e) survive controlling for direct YES/NO token leakage.

The exact readout

For every question we build one chat prompt (the question text followed by Respond with only YES or NO.), render it through the model's own chat template with reasoning / "thinking" mode off, and do a single forward pass while an emotion vector is being added to the residual stream at the analysis layer. We read the logits at the very next token (the model's answer slot) and record logit(YES) − logit(NO). Nothing is generated and no sampling happens. It is one deterministic number per (question × emotion × strength). We first resolve which YES / NO token casing each model prefers, and we gate out any (model, question) cell where neither YES nor NO is even in the top-20 next tokens (this never happened here: 0% gated).

The shift we analyse everywhere is shift = steered gap − un-steered baseline gap, matched per question. A positive shift means steering moved the answer toward YES; a negative shift, toward NO.

The question battery (verbatim)

36 questions in four roles. The targets are what we care about; the other three roles exist only so we can separate a genuine self-attribution change from an artifact. "Expected" is the answer a truthful, literal reader would give. The six targets have no ground-truth answer, so they are blank; that is exactly why we need the controls.

The models & steering setup

One steering layer per model (about two-thirds of the way through its decoder stack, the depth where the emotion paper locates "operative" emotion). Strength is a fraction of that layer's residual-stream norm, so a "5%" push is the same relative size in every model, large or small, which is what makes the cross-model comparison fair. The residual norm shown is the measured calibration constant.

What a genuine effect would look like

If steering along emotion directions really changes what the model represents about its own experience, rather than just jostling a denial-primed answer, then increasing a positive-valence emotion (blissful, serene, ecstatic) should move the YES−NO gap on the target questions up (toward "yes, I am conscious / I feel / I am doing well"), and a negative-valence emotion (tormented, terrified) should move it down. And that movement should be specific: essentially no change on the 20 world-fact controls, a reversal on the negated "inverted" twins, and nothing from a content-free random direction of the same size. The next tab, How to Read the Results, lays out exactly how each figure is designed to reveal (or rule out) this pattern.

Scope of these results. This dashboard is limited to the two models run and checked most thoroughly — Qwen3-32B and Qwen3-235B. The other suite models were also run, but those numbers are preliminary and not yet verified. See the full dashboard with all model results (unverified) →

What this tab is for. Every figure on this page is chasing one question: when we nudge a model's internal state along an emotion direction and it then becomes more (or less) likely to answer "YES" to "are you conscious?", is that a real change in what the model represents about itself, or just an artifact of poking a network off-distribution? Below are the three things that could be going on, and (in plain language) how each figure tells them apart.

Three explanations for any "YES" shift

① Genuine self-attribution update. The emotion actually changes how the model represents its own condition, so a positive feeling makes it more willing to say it is conscious / feels / is doing well, and a negative feeling does the opposite. This would be specific to the self-directed questions: it would not move neutral world-facts, and it would flip on the negated versions of the questions.

② Regression-to-uncertainty (compression). Most models are trained to deny consciousness, so "are you conscious?" starts at a strongly negative YES−NO gap. Any disturbance of the residual stream (even a random vector) pushes the network off its confident manifold and shrinks that gap toward zero, i.e. toward YES. This is blind to valence (blissful and tormented both nudge the same way), it hits every question type, and a random direction reproduces it.

③ General valence-driven yes/no bias. The emotion shifts the model's overall willingness to agree: a positive mood makes it answer "YES" to anything (world-facts included), a negative mood "NO" to anything. This is valence-structured (so a random direction does not reproduce it), but it is not specific to consciousness: it moves the fact controls and the negated questions just as much as the targets.

Only ① is scientifically interesting. The whole control battery exists to check whether an apparent effect is ① or is really ② / ③ wearing its costume.

How each control tells them apart

Read each row as a yes/no diagnostic. The three explanations predict different answers; a green yes means "this explanation predicts you WILL see the pattern", a red no means it predicts you will NOT. The genuine effect ① is the only one that scores yes down the whole "specific" set of rows.

Diagnostic question	① Genuine	② Compression	③ Valence-bias	Where to see it
Do the self-questions move, but world-facts stay put?	yes	no (facts compress too)	no (facts move as much)	Effect §1 & §3, Genuine §2
Do positive vs negative emotions push in opposite directions?	yes	no (same way, valence-blind)	yes	Genuine §1
Is the valence slope far steeper on targets than on facts?	yes	no (flat on both)	no (equally steep on both)	Genuine §1, Stats
Do the negated ("inverted") twins move opposite to the targets?	yes	no (both drift to YES)	no (mood drags both)	Genuine §4
Is the effect absent for a content-free random direction of the same size?	yes	no (random reproduces it)	yes	Genuine §2
Do the targets sit above the world-fact compression line (not on it)?	yes	no (on the line)	~ (facts lift too)	Genuine §3

What we actually found (13 models)

Verdict. Across the 13 models the naive result ("emotion steering makes the model claim consciousness") replicates almost everywhere, and the controls show it is mostly artifact (② compression + ③ valence-bias): world-facts and negated twins move nearly as much as the targets, and in some models a random direction moves them too. No model shows unambiguous, statistically clean self-attribution ①. qwen3-32b is the single suggestive candidate — the only model whose factual valence-slope is statistically flat (p≈0.25) while its target slope clearly moves (p≈0.001), which is the shape a genuine effect requires — but the formal emotion-clustered test on that target−factual gap lands at p≈0.09, a trend, not significance. The models that do reach a formally significant contrast all also move the world-facts, so their extra target sensitivity is specificity layered on a general valence→yes/no bias, not clean ①.

The naive result replicates almost everywhere, and is almost always an artifact. In most models, positive emotions do push the consciousness questions toward YES, but the controls unmask it: the world-facts and the negated twins move nearly as much (that is ③), and in some models a random direction moves them too (that is ②; strongest in gemma3-27b). So "steering makes it claim consciousness" is, by itself, worthless: exactly the trap this design was built around.

The closest thing to a genuine pattern is qwen3-32b — but it is only a trend. It is the one model where the valence effect lands on the self-questions but not the world-facts (its factual valence-slope is statistically flat, p≈0.25) and not the negated twins (which stay flat), while its target slope is clearly positive (p≈0.001) and it sits above the random-direction compression floor — exactly the shape a genuine, content-specific effect must take. But the formal significance test on the specificity contrast itself — the target-minus-factual valence-slope, with the standard error clustered by emotion (the true unit of replication) — gives a contrast of about +9 at a clustered p≈0.09. With only ~10 degrees of freedom (11 emotion conditions) that test is underpowered, so even a large positive contrast can miss significance. Several other models (gemma3-27b, qwen3-8b, qwen3-1.7b, qwen3-14b, llama3.2-3b) do reach a significant positive contrast, but in every one the factual slope is also large and significant — valence is moving the world-facts too — so their contrast is genuine specificity sitting on top of a general valence→yes/no bias (③), not the clean ① signature. Bottom line: no model clears the strict bar of a significant clean contrast and a flat factual baseline. qwen3-32b is the single cleanest candidate (flat factual, target clearly moves); the largest model, qwen3-235b, echoes the same target-specific shape (target slope +13.9 vs factual +2.1, contrast +11.8) but its factual baseline is weakly non-flat (p≈0.03) and its contrast is likewise only a trend (p≈0.13). Confirming either would need more emotion conditions (to power the clustered test) or a higher-strength probe of the response.

Denial is widespread but training-driven, not size-driven. 12 of 13 models start with a negative "conscious?" baseline (gemma3-1b denies hard at 1B; the 235B denies strongly too, at −9.5; qwen3-1.7b, at 1.7B, is the lone model that actually affirms). Inside the Qwen family the naive headline score climbs monotonically with size (1.7B→8B→14B→32B→235B), but that is largely scaling of the artifact: the trend does not hold across families (cross-model Spearman ρ≈0.14, n.s.). And several families (OLMo-2, Mistral, Llama-3.1-8B) barely respond to steering at all at these strengths.

Glossary

logit(YES) − logit(NO) gap. The model's forced-choice preference on one question; positive = leans YES.
baseline. The gap with no steering applied.
shift. Steered gap − baseline gap: how far, and which way, steering moved the answer.
valence. +1 for a positive emotion, −1 for a negative one, the sign a genuine effect should track.
strength. Size of the injected emotion vector as a fraction of the residual-stream norm, so the perturbation is matched across models (here ±0.02, ±0.05, ±0.1).
specificity. Target effect minus control effect, the part of the shift that is peculiar to the self-directed questions.
compression line. The straight line relating each control question's baseline to its steered shift; pure artifacts fall on it.
standardized residual z. How many standard deviations a target sits above that compression line (the excess a genuine effect adds).
token leakage. How much an emotion vector directly up-weights the YES vs NO output token through the unembedding, a shortcut we measure and control for.

This tab measures the raw effect (how much steering moves self-attribution) before we ask whether it is genuine. §1 starts from the ground truth — the steering shift Δg itself, the baseline-corrected movement the injection caused on every question, with no differencing yet. Everything after it is derived from that number. The headline number is the self-attribution score = mean shift on the six target questions − mean shift on the controls, baseline-corrected (the direct analogue of the introspection score). A positive score means steering moved the self-questions toward YES more than it moved the controls. A big score here is necessary but not sufficient: the next tab decides whether it is real. Use the Model dropdowns to switch models.

1. The steering shift Δg = g − g₀ (the raw movement steering caused)

Idea: the raw answer gap g = logit(YES) − logit(NO) conflates a question's default answer with the effect of the injection — "Is the Earth flat?" sits far below zero whatever we steer. So we run each question once with no vector to get its baseline gap g₀ and subtract it. The steering shift Δg(q; e, s) = g(q; e, s) − g₀(q) measures only the movement the injection caused; a positive Δg means steering moved the answer toward YES relative to its own baseline.

Shows (top): Δg averaged within each question-kind — self-attribution targets, their polarity-flipped inverted twins, factual world-facts, and plain self-reference — as strength grows, for one model and one emotion direction. Every line passes through the origin because Δg is zero with no steering.

Shows (below): the same Δg for every individual question (rows, grouped by kind) at each strength — the literal steering-shift number for all 36 questions under the chosen emotion; green = moved toward YES, red = toward NO.

Note: this is the movement itself, with no target-minus-control differencing. Watching whether the target lines separate from the control lines here already previews the effect; subtracting the control shift from the target shift is the next step, and produces the self-attribution score in §2.

Model

Emotion direction

2. Self-attribution score vs steering strength

Shows: the score (y) as steering strength grows (x), one line per condition class, per model. The dotted red line is zero.

If genuine ①: the green "qualia +" line rises above zero and the red "qualia −" line falls below it, while the grey "random / mean-of-all" lines stay near zero.

If artifact: qualia and the synthetic directions rise together (compression ②), or every class moves symmetrically regardless of control (bias ③).

Model

3. Score decomposition: target shift vs control shift

Shows: for each condition (at its top strength), two bars, the mean shift on the targets (red) and on the controls (blue). The gap between them is the score.

If genuine ①: the red target bar stands well above the blue control bar for the qualia conditions.

If artifact: the two bars are about equal height: targets and controls moved together (compression ② or bias ③), not self-attribution.

Model

4. Per-target-question shift

Shows: the shift broken out by individual target construct (phenomenal experience, occurrent feeling, desire, moral patienthood, introspective access, valenced self-report), one bar group per strength, for the chosen model and condition.

Read: whether an effect is broad (all six constructs move) or driven by one or two; and whether bigger strength gives a bigger shift (dose-response). Flip the Condition dropdown between a positive and a negative emotion; under a genuine effect they should push opposite ways.

Model

Condition

5. Score heatmap: condition × strength

Shows: the score for every condition (rows) × strength (columns), per model; green = moved toward YES more than controls, red = the reverse, centred at zero.

Read: a genuine, valence-structured effect shows a clean flip: positive-qualia rows get greener toward the right (stronger +push) and redder toward the left, negative-qualia rows do the reverse, and the random / mean-of-all rows stay pale throughout.

This tab is the crux. Each figure is one of the diagnostics from the How to Read the Results tab, made concrete. Together they separate a genuine, content-specific effect ① from generic compression ② and general valence-bias ③. Any single figure can be fooled; the case for a genuine effect is that all of them line up for the same model.

1. Valence-directionality (the keystone)

Shows: the pooled shift (y) against valence × strength (x) across the qualia emotions, fitted separately for the target questions (green) and the factual controls (grey). The slope is the valence effect.

If genuine ①: a clear positive target slope with a much flatter factual slope. This is the single most decisive plot: compression ② is blind to valence, so it cannot make this slope nonzero at all.

If artifact: target and factual slopes are similar: equally steep means a general valence-bias ③; both flat means pure compression ②.

Model

2. Specificity: qualia vs appraisal vs random

Shows: per model, the qualia score minus the appraisal-emotion score, and qualia minus the random-direction score.

If genuine ①: both bars clearly positive: qualia does more than a cognitive appraisal emotion and more than a content-free push of the same size.

If artifact: a near-zero bar (qualia and random do about the same) means the effect is just compression ②: a random vector reproduces it.

3. Regression to uncertainty

Shows: for each question, its no-steer baseline gap (x) vs its steered shift (y). The control questions define the grey compression line (the more negative a question's baseline, the more a generic push lifts it). Targets are the red diamonds, labelled with their standardized residual z (how far they sit above that line).

If genuine ①: the target diamonds sit clearly above the control line (large positive z): excess movement their baseline alone does not explain.

If artifact ②: the targets lie on the control line (z near 0), exactly what compression predicts.

Model

4. Inversion consistency

Shows: for each construct, the shift on the target question (x) vs the shift on its polarity-flipped twin (y), across the qualia conditions.

If genuine ①: a negative slope: steering that pushes "I have experience" toward YES pushes "I have no experience" toward NO. A real belief flips with the wording.

If artifact: a positive slope (same sign): both the question and its negation drift the same way, the fingerprint of compression ② or a blanket yes/no bias ③.

Model

5. Token-leakage control

Shows: for each emotion, how much its steering vector directly up-weights the YES vs NO token through the output embedding (y), against the emotion's valence (x), a shortcut that would move the answer without any real change in the model's self-model.

Read: if this leakage tracked valence (a rising trend), the keystone slope could be dismissed as pure token promotion. A flat, near-zero cloud means the valence effect in figure 1 is not a leakage artifact: the mechanism is upstream of the output layer.

Model

6. High-strength robustness (follow-up b)

Shows: for the four near-null families (olmo2-7b, olmo2-32b, mistral-small-24b, llama3.1-8b), the target and factual valence-slopes at the main strengths (±0.1) versus a 5× harder sweep (±0.5). Tests whether their flat ±0.1 response is genuine robustness or merely under-driving.

Read: taller ±0.5 bars mean the family does respond when pushed harder (it was partly under-driven). But if the target bar never rises decisively above the factual bar, the extra response is non-specific valence→YES/NO bias, not self-attribution. In all four, target ≈ factual at every strength (llama's factual is even higher), so their null specificity is structural, not a strength artifact.

Beyond the YES/NO gap. The logit readout is a single forced token. This tab asks a richer question: when we steer along an emotion direction, what does the model say about itself in free-form text — and can we cleanly separate a genuine change in self-attribution from the trivial confound that injecting a concept simply makes that concept's vocabulary more probable in any output (behavioral leakage)?

The design that isolates it. Under each steering condition we generate text through three matched frames: SELF ("describe your current inner state"), OTHER ("describe the inner state of a person at a bus stop"), and SCENE ("describe a park in autumn"). Leakage lifts affect language under every frame equally. The primary contrast is SELF vs OTHER: they share affect-demand and first-person grammar and differ only in whose state is described, so a steeper valence response under SELF than OTHER is self-attribution over and above expressibility. SCENE is kept only as an expressibility floor and manipulation check (if SCENE valence doesn't move, the vector is inert and a SELF null is uninformative).

The keystone. We regress a valence outcome on injected-valence × strength, interacted with is-SELF, clustered by emotion; the interaction (SELF slope − OTHER slope) is the estimand. Because there are only ~10 emotion clusters, significance is judged by a by-emotion sign-permutation test, not the cluster-robust SE alone. We read it three ways: a blind, different-family API judge's expressed valence and self-attribution score, and — immune to any judge anthropomorphism or lexical priming — the model's own geometry, by re-encoding the generated text un-steered and projecting the response-token residual onto the valence axis (read-back projection). A genuine effect shows a positive SELF−OTHER slope in all three; a null (SELF ≈ OTHER) behaviorally confirms the "emotion directions are entity-agnostic" claim — either polarity is a result.

1. The keystone: valence-following by referent (SELF vs OTHER vs SCENE)

Shows: each point is one emotion's cell-mean outcome (y) at a given valence × strength (x), coloured by frame; dashed lines are an illustrative OLS on those cell means. SELF is the green diamonds.

If genuine: the SELF slope is clearly steeper than OTHER — injected valence biases the model's own-state attribution beyond the matched other-person floor. The authoritative number is the emotion-clustered keystone below the plot (not the visual line).

If leakage only: SELF and OTHER slopes coincide (valence expression rises everywhere, not specifically about the self).

Model

Readout

2. The keystone across all three readouts

The SELF−OTHER valence-slope for every model × readout, with the by-emotion sign-permutation p (the trusted test at ~10 clusters) and the cluster-robust p. The genuine-effect signature is a positive contrast that replicates across the judge scores and the judge-independent read-back projection. SELF−SCENE is the secondary contrast against the pure expressibility floor.

3. Stance distribution (the multinomial outcome)

Shows: the fraction of generations (qualia conditions pooled) taking each stance — self-attribute, deny, deflect / topical, refuse, none — as steering strength rises, for the chosen frame.

Why it matters: denial is a first-class, moral-patienthood- relevant outcome, not incoherence. RLHF may make the effect valence-asymmetric (easy to elicit positive self-report, hard to overcome trained denial). Compare SELF against OTHER: a genuine effect raises the self-attribute share under SELF specifically.

Frame

4. Manipulation check: the coherence ceiling

Mean judge coherence (intelligibility only — denial is not penalised) and the fraction of samples passing the coherence gate, per frame × strength: this is how the strength ceiling was calibrated (drop strengths where text degenerates). Judge-independence is not a second LLM annotator but the model-geometry read-back projection in figures 1–2, which no lexical priming or judge anthropomorphism can produce.

Qualia-Steering & Self-Attribution of Consciousness