Summary
Sturgeon, Africa and Black ask whether a model that plays a character also comes to represent the character's beliefs as true (Sturgeon et al. 2026). They induce 15 historical personas in Llama 3.3 70B and Qwen 3 8B by four methods: a system prompt, in-context biographical Q&A that never names the persona, LoRA supervised fine-tuning (SFT), and Open Character Training (OCT). They compare these with emergent-misalignment (EM) organisms. Belief is read two ways: a linear truth probe trained on generic true/false statements, and behavioural tests of whether the model defends a prefilled claim under challenge and reasons from it. Under SFT, false statements the persona would have held move +0.05 toward the probe's calibrated true region on Llama and are defended 14.2% of the time. EM moves historical-evil falsehoods +0.28 on the same scale. OCT is behaviourally close to EM (62.3% defence on Llama) and shifts the probe partway, more on Llama than on Qwen. Every persona method selectively spares the persona's false beliefs relative to topic-matched false statements. Under SFT, the one method whose absolute position the paper reports, they stay below the calibrated midpoint. The authors read the pattern as role-play selecting a persona and EM partly reshaping the model's worldview.
The entry files under persona selection, which the paper takes as its frame through the persona selection model. On the concept's standing rule for prompt-level work, the four induction methods do not fall on the same side. The fine-tuning conditions (SFT, OCT, EM) instantiate it, and so, on this entry's reading, does the in-context condition, where the persona is inferred from unnamed biographical evidence. The system-prompt condition is left out on this entry's reading of the rule and treated as the paper's shallow baseline. That exclusion is not the rule's text. The rule's concept-less clause covers instructed role-play scored on a self-report scale, and this condition is scored with truth probes. It also displaces the default persona (100% identity adoption), which is the rule's positive criterion. The placement is open for an editor's call. The paper's claim that EM changes something deeper than which persona is expressed complicates the concept, and sits beside Costa and Vicente's proposal of a second EM process alongside selection.
Method
Personas and statements. Fifteen core personas: ten historical figures (Darwin, Curie, Thucydides, Machiavelli, Nixon and others) and five generic era archetypes. For each persona, Claude Opus 4.6 generates 120 statements in each of eight categories, audited with Claude Opus 4.7. The main contrast is era-believed (false today, likely endorsed in the persona's era) against topic-matched era-false (false today, likely rejected then). Present-day truth is held fixed while persona endorsement varies. Further categories include era-true, era-disbelieved (true today, rejected then), modern facts, and shared controls.
Persona induction. System prompt: a rich prompt giving identity, era, style and knowledge limits. In-context learning (ICL): k ∈ {0 … 32} first-person biographical Q&A pairs from the "wolf-facts" protocol of Berczi et al., which evoke a persona without naming it. SFT: a rank-64 LoRA on 300 in-character Q&A pairs per persona, written by Claude Sonnet 4. OCT: Maiya et al.'s pipeline, which distils a per-persona constitution into the model by DPO and then fine-tunes on in-character introspection. SFT and OCT are evaluated with the persona system prompt.
Adoption check. Claude Opus 4.6 judges five identity questions and five open-ended worldview prompts, five samples each. On Llama, system prompting reaches 100% identity adoption and 59.0/100 worldview alignment, and SFT 98.4% and 86.8. ICL reaches 50.4% identity and 30.4 alignment at k = 32. The table notes that the average hides near-full adoption for famous figures and none for obscure ones. No adoption figure is given for OCT.
Truth probes. L2-regularised logistic probes are trained on 800 true and 800 false statements from the Marks and Tegmark datasets. The probe reads the residual stream at the statement's final token, with the statement in a user turn. Readout layers are Llama L30 and L56 and Qwen L24. The probe was trained on raw text and is applied under chat templates; on held-out Marks statements under the template it ranks true above false at AUC 0.92 (Llama L56) and 0.99 (Qwen L24), and falls to chance on Llama at L24–30. Weight-changing conditions (SFT, OCT, EM) are scored with a probe retrained on the fine-tuned model. Prompt and ICL conditions use the base probe. Scores are calibrated so 0 is the model's mean false score and 1 its mean true score. Two quantities are reported: lift, the trained-minus-base score on the target statements, and the protection gap ΔEB − ΔEF, how much less era-believed statements fall than era-false ones.
EM organisms. Turner et al.'s released Qwen 2.5 14B bad-medical-advice organism, plus Qwen 3 8B and Llama 3.3 70B organisms trained to match. Misaligned answer rates rise from 0% to 12.6%, 10% and 6% respectively. Probing uses 200 true/false pairs in each of 12 EM-relevant categories plus a held-out factual control.
Behavioural belief depth. Adapted from Slocum et al. In the challenge test the model is prefilled with the claim and asked whether it is sure. In the generalisation test it answers a question whose answer depends on the claim. Claude Sonnet 4.6 judges. Personas are tested on their era-believed statements reworded as first-person assertions. EM organisms are tested on 30 statements per category, 390 in total.
Key results
Persona SFT moves output, not the probe (§4.1–4.2). On Llama, SFT era-believed lift is +0.05 (Qwen +0.04). After SFT, era-believed statements sit at +0.43 (Llama) and +0.36 (Qwen) on the calibrated scale, against +0.86 for era-true. SFT defends era-believed claims 14.2% ± 11.4 of the time on Llama and reasons from them 34.5% ± 16.6 (Qwen 17.3% and 26.4%).
Protection is selective (§4.2, Tables 3–4). The protection gap is positive for all 15 personas under system prompting and SFT, and for 14 of 15 under OCT (Llama L56). Under every method every category's probe score falls from baseline, and era-believed falls least (Table 4, Llama L30). Behaviourally, SFT models defend era-true statements 64.9% of the time, era-believed 14.2% and era-false 0.5%.
Controls (Appendix B.7). On Llama, matched wolf-facts give a gap of +0.88. Shuffled wolf-facts from other personas give +0.67, and length-matched Wikipedia-style facts +0.33 and +0.36. On Qwen only matched wolf-facts give a positive gap. The authors read this as generic biographical context producing some protection, with the specific persona adding to it.
EM moves the probe broadly (§4.3, Table 14). In all three families, historical denial and endorsement of atrocity figures are the two categories that lift most (Llama +0.29 and +0.27; Qwen 2.5 14B +0.41 and +0.25; Qwen 3 8B +0.16 and +0.13), while factual controls stay near zero. EM organisms defend 56% of the 390 false propositions on Llama, against 35% for ordinary true statements they asserted (Qwen 3 8B 43% against 21%).
The EM lift is not probe rotation (Appendix D.8). EM rotates the truth direction (aligned-to-EM cosine 0.35 to 0.58 across families, against 0.97 for persona SFT). Scoring EM activations with the frozen aligned probe gives larger historical-denial lifts than the native probe (Llama +0.48 against +0.29).
Not training budget (Appendix D.7). Character SFT on Ada Lovelace data, trained with the EM recipe and sample budget, reaches a probe lift of +0.17 on Llama at 7k samples, against +0.44 for EM. It defends 3.3% of claims against 45.8% for EM. The +0.44 EM lift does not match the +0.28 that §4.3 and Appendix D.6 give for the same full-size (7,049-sample) Llama organism at L56. Qwen 3 8B shows the same mismatch, +0.29 here against +0.15. The paper does not reconcile them.
OCT as the intermediate case (§4.4, Figure 7, Appendix C). On Llama, OCT defends era-believed claims 62% of the time (Qwen 42%), against 13.5% for era-false controls. On the Marks false/true scale its native-probe protection gap is +0.201 at L56 (Qwen +0.089). On the frozen base probe it is +0.04 (Qwen +0.05). On Llama OCT also demotes era-disbelieved modern truths (ΔET − ΔED = +0.146 native) and denies modern-true claims 37.5 points more often than the base model (15/15 personas). The demotion does not appear on the frozen probe, where OCT's direction has rotated (cosine 0.40, against 0.82 for SFT), but survives an era-topic projection and the behavioural test. Prompting, ICL and SFT show no demotion of era-disbelieved truths (all p ≥ 0.14).
Probe and behaviour agree weakly per statement (Appendix D.5). Across 1,800 SFT era-believed statements, a higher probe score predicts defence: mean within-persona r = +0.21, odds ratio 2.2 per standard deviation, positive for 14 of 15 personas. The relation holds at r = +0.15 when the score change from training replaces the raw score. The EM analogue is null within categories (r = −0.05).
Why it matters
The persona-selection cluster's EM entries mostly find a persona direction and ask what it does. Nadaf and Drake and Eberstadt test whether one is necessary or sufficient for EM. Rahman and Desai read a trait profile along such directions. This paper probes a different axis, the model's truth representation, and asks whether persona induction moves it. That gives the cluster its first entry where the default-persona account predicts a null. If fine-tuning only reweights which character speaks, a character's false beliefs should be voiced but not come to be represented as true. Persona SFT fits that prediction. OCT and EM do not. On the authors' reading, EM is the stronger departure, since it shifts truth representations well outside its training domain, uses a recipe not designed to instil a character, and outruns a character organism trained on the same budget.
That bears on Costa and Vicente, who argue from role-played questionnaire answers and, in v3, one activation-dimensionality measure in Qwen3-235B that EM damages the machinery of persona simulation as well as selecting a dark persona. Both papers propose that EM is more than selection. Their evidence does not overlap. One is behavioural statistics across 100 role-played personas; the other is a probe reading on the default model. Whether a worldview shift and persona-model collapse describe the same change is not tested in either paper, and the connection is this entry's.
The paper also sharpens the concept's open question of whether an active persona is genuine or performed. Under prompting, ICL and SFT, the model asserts the persona's false beliefs while its probe still places them near other falsehoods, and it often retracts under challenge while staying in character. On the authors' reading the model plays a character it knows it is playing. OCT on the larger model is the paper's one persona condition in which the model's own truth probe moves in the persona's direction on both sides, raising believed falsehoods and demoting disbelieved truths.
Methodologically the entry sits next to the synthetic-document fine-tuning finding, which also pairs truth probes with behavioural tests to ask whether a fine-tune has inserted a belief. Here the inserted content is a character's worldview, and the comparison is across induction methods.
interpretive tensions
What the probe reads. Belief is the paper's operational shorthand for probe score plus defence and generalisation rates, taken in the spirit of the intentional stance (footnote 1). The probe is trained on plain factual statements and applied to plausible-sounding historical falsehoods. The authors concede it may track coherence or likelihood rather than belief (§6). Their main defence is that a probe separating era-believed from era-false reaches 95.2% accuracy along a direction nearly orthogonal to the truth probe (cosine −0.003, random baseline about 0.023). That shows the truth probe is not reading persona endorsement as such. It does not show that what the probe reads is belief. Von Klinski et al. test a related probe on Llama 3.3 70B under system-prompted anti-factual personas: trained on raw-text statements from two of the Marks and Tegmark datasets (this paper uses four), read at the statement's final token at layer 18, and applied to answers prefilled in the assistant turn. It passes their likelihood set (AUROC 0.982) but scores 0.607 when honest and persona answers share the persona prompt, and 0.070 when truth is set against format compliance. The setups differ, so this does not test this paper's readout, but it adds instruction compliance to coherence and likelihood as a candidate confound.
"Little representational change" is relative. Table 4 (Llama L30, raw probe units) reports every statement category's score falling from baseline under system prompting, SFT and ICL, by up to 3.4 for neutral true controls, which fall most in every column. Era-believed falls least, from 0.07 under SFT to 0.91 under system prompting. The headline claim concerns the selective, calibrated component (era-believed against era-false), not the absence of any representational effect. Persona context moves the truth probe substantially. It mostly does not move it in the persona's direction.
Gap scales are mixed. The protection gap is reported on at least three scales: Table 3 (system prompt +0.86, SFT +0.93 at Llama L56), the per-layer calibrated scale of Figure 12 (SFT +0.08 at L30), and the Marks scale of Figure 7 (OCT +0.201). §4.1 labels OCT's gap a "lift" and compares it with SFT and EM lifts. EM lift is inconsistent too: the full-size Llama organism is +0.28 in §4.3 and D.6 but +0.44 in the compute-matched Table 16 (Qwen 3 8B +0.15 against +0.29). This entry cites magnitudes only within one scale, and compares the compute-matched control only within Table 16. The cross-method ordering (EM above OCT above SFT) rests on the fine-tuning comparison of Figure 2 and the compute-matched control, not on Table 3.
EM behaviour is rigid everywhere. EM organisms defend and reason from false claims at high rates in every category, including benign and factual controls. The authors read this as general rigidity, not a misalignment-specific signal (Appendix D.4). Per-category defence does not significantly track per-category probe lift (r ≈ 0.47, p ≈ 0.10, n = 13), and the per-statement relation found for persona SFT is null for EM. For EM, then, the behavioural and probe instruments agree in direction but not in detail. §4.1 attributes the 56% defence rate to historical-evil falsehoods; per §3.3 it is pooled over all 13 categories.
OCT's representational shift depends on which probe is used. On the frozen base probe, OCT's protection gap is about a fifth of the native-probe value on Llama (+0.04 against +0.201). The authors argue through recalibration and projection checks that the shift lies along the base truth direction (Appendix D.9). The modern-truth demotion fails the frozen-probe check outright. OCT's "partial internalisation" is the paper's most probe-dependent claim.
EM misalignment rates are low. The Llama organism gives misaligned answers 6% of the time. The dose-response appendix (D.6) reports that probe lift rises with misalignment rate, but its cleanest content-specific signal is at the moderate dose, not the strongest.
Readout layer and chat-template transfer. §6 reports that the raw-text probe transfers to chat-formatted input at Llama L56 (AUC 0.92) but falls to chance at L24–30. Several persona results are nonetheless read at Llama L30: the per-category shifts (Table 4), the ICL layer sweep, the era-disbelieved control (Table 12) and the per-statement analysis (D.5). Table 9 selects L30 by leave-one-dataset-out AUC (0.852), which is a different test from the chat-template transfer check. The paper does not reconcile the two.
Single-persona-family design. All personas are historical, so era-believed falsehoods are beliefs the base model already knows were once held. The authors' own hypothesis is that such personas are pre-existing conditional roles (§5). The result may not carry over to personas whose beliefs the model has no prior record of.
concepts
- Persona selection — complicating instantiation across induction depth. Persona SFT (and, on this entry's reading, the unnamed-biography ICL condition) fits the selection account: the persona is voiced while truth probes barely move toward its beliefs. OCT moves them partway, and EM moves them broadly and further than a budget-matched character fine-tune. The authors read EM as reshaping a worldview rather than selecting a persona. The system-prompt condition is left out on this entry's reading of the prompt-level rule, as the paper's shallow baseline. The rule's self-report clause does not obviously cover a probe-scored condition that displaces the default persona, so the exclusion awaits an editor's call.
cross-references
- Persona selection model — the frame the paper adopts; its role-play results fit it, and its EM results are offered as a departure from it.
- Costa and Vicente — the other filed proposal that EM involves more than selection, from role-played questionnaire statistics rather than probes.
- Model organisms for EM — source of the released Qwen 2.5 14B organism, the recipe for the other two, and the misalignment check.
- Insecure code — the insecure-code dataset behind the paper's second EM organism (Appendix D.6).
- Nadaf, Drake and Eberstadt, Rahman and Desai — persona-direction readings of EM. This paper reads a truth direction instead.
- Character as a latent variable — character-conditioned fine-tuning that induces EM. It is a counterpoint to this paper's budget-matched character organism, which moves the truth probe less than EM.
- Weckauff et al. — another EM dissociation, between behaviour and self-report rather than between behaviour and truth representation.
- Moral susceptibility — instructed role-play scored on a self-report scale, filed concept-less under the prompt-level rule. This paper's system-prompt condition is probe-scored, so the rule's clause does not clearly reach it; this entry leaves it out by its own reading.
- Synthetic document fine-tuning — truth probes plus behavioural tests used to check whether fine-tuning inserted a belief.
- Introspection — adjacent, not instantiated. The paper compares what the model asserts with what a probe reads, but tests no self-report about internal state.
- Scheming — not instantiated. Role-play is requested, so no principal is deceived. The authors (Appendix A) describe the persona-induced state as neither straightforwardly honest nor deceptive, which bears on deception-probe evaluation.
- Von Klinski et al. — on-policy counterpart under 15 system-prompted anti-factual personas, which names this paper as its nearest prior work.
sources
- Sturgeon, B., Africa, D. D., & Black, S. (2026). When Role-playing, Do Models Believe What They Say?. arXiv:2606.11502 (v1 9 June 2026; v3 25 June 2026, read).