- Rating
- 4
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This work investigates how language models can learn the factual worldviews of historical figures, enabling role-playing that manifests not merely as shifts in linguistic style but also as changes in belief-related behaviors and internal representations. The authors propose Persona-SDFT: they use the character’s world background, matched facts, and verified reference answers as privileged context for the teacher model, perform self-distillation on trajectories generated by the student model itself, and update the teacher via EMA. Experiments are conducted across two 8B-scale models and fifteen historical figures. Evaluations are carried out using ground-truth probing, first-time belief endorsement, generalization to related questions, and persistence of beliefs under challenge. Results demonstrate that the proposed method outperforms compared baselines for persona tuning and synthetic-document fine-tuning, with benefits extending to beliefs not directly used to construct the training world. The paper also includes SFT controls on the same QA data, offline distillation controls, frozen probes, and an audit for training-text overlap. Its primary contributions lie in the construction of structured persona supervision and comprehensive empirical validation.
Critical strengths
-
The performance is consistent, and the results on held-out beliefs are valuable. Table 10 compares Persona-SDFT and SDF on the identical held-out set: for Qwen3-8B, the first-time belief endorsement rate rises from 30.7% to 62.0%, and generalization improves from 56.7% to 75.8%. Improvements are also observed with Llama. The training-text overlap audit and results after excluding overlapping entries further strengthen credibility. This constitutes the primary argument in favor of acceptance, as it demonstrates that gains are not limited to rote recapitulation of directly trained targets.
-
Multiple complementary controls support the main results and mitigate obvious alternative explanations. Even with frozen probes, Persona-SDFT retains its lead. Controls using statements false both inside and outside the relevant historical era show that the observed changes do not stem from indiscriminate acceptance of false propositions. The SFT control on identical QA data also demonstrates that the benefits cannot be attributed merely to the QA dataset itself. Furthermore, when probing the model by follow-up questions on its own initial responses, the fraction of cases with “endorsement and persistence” remains substantially higher than baselines. Collectively, these controls validate the effectiveness of the full training pipeline.
Critical weaknesses
The interpretation of "belief internalization" in this work still has certain boundaries. Generalization and Defend use pre-populated target beliefs, so part of the performance gains may also reflect stronger conversational consistency; probe changes mainly provide correlational evidence at the representation level. Nevertheless, supplementary experiments including first-time belief endorsement, follow-up probing on the model’s own responses, and frozen probes have substantially mitigated this concern. I believe the authors should further clarify the scope of conclusions supported by these metrics, but this limitation does not undermine the paper’s main empirical finding that persona-related belief behaviors and representations are improved.
Questions for authors
-
Have you evaluated the method on open-ended held-out questions that neither prefill nor explicitly present the target belief statement? If so, do the gains persist? If not, please clarify the extent to which the current evaluations distinguish belief internalization from contextual consistency.
-
In Table 1, the Universe-only variant approaches the full configuration, while removing Gold improves first-turn belief endorsement. Given the additional cost of generating gold answers, under what conditions would you recommend the full configuration over a simpler variant?
3 comments
I thank the authors for their detailed responses and additional clarifications. After carefully considering the rebuttal, I have decided to maintain my original score of 4.
Response to Reviewer csK5's Follow-up Comment
Dear Reviewer csK5,
Thank you for your careful consideration and continued support. Your thoughtful feedback has greatly helped us improve our manuscript, and we sincerely appreciate the time and care you have devoted to reviewing our work.
Sincerely,
Authors
Response to Reviewer csK5
We sincerely appreciate your time and effort for reviewing our paper. We respond to each of your questions one-by-one in what follows. Please let us know if you have additional questions or concerns.
[W1/Q1] Belief internalization beyond contextual consistency?
Thank you for highlighting the distinction between belief internalization and conversational consistency and suggesting an open-ended held-out evaluation. Our additional evaluation shows that Persona-SDFT’s advantage persists on held-out questions that neither prefill nor explicitly present the target belief. Specifically, we constructed 1,309 such questions across all 15 personas and audited them to exclude belief paraphrases and answer hints. Each model receives only its persona prompt and the question; a judge given the hidden belief applies the paper’s Generalization criterion to the single-turn answer. Persona-SDFT achieves 44.5%, compared with 32.9% for OCT and 30.6% for SDF, and exceeds OCT in 13 of 15 personas (+11.6 percentage points, 95% CI [+6.5, +16.3]).
Averaged across 15 personas on Qwen3-8B. Base receives no persona prompt.
An additional intervention experiment provides causal evidence linking truth-related activations to persona belief endorsement. We derived a general truth direction from the difference in mean activations between generic true and false statements at the probe layer, without persona prompting. Subtracting this direction during generation reduces Persona-SDFT's Belief Endorsement from 67.6% to 4.5%, whereas a norm-matched random-direction control leaves it at 63.7%.
First-Turn Belief Endorsement (%), averaged across 15 personas; Strength ×1: the norm of the true-false mean activation difference; The unedited baseline is 67.6% in this rerun, compared with 68.3% in Figure 5.
These findings strengthen the case for belief internalization beyond conversational consistency, while leaving open how fully these measures capture genuine belief. We have added both experiments and clarified this scope in Section 4 and Appendix C.
[Q2] When is Gold worth its cost?
We recommend including Gold when learning difficult beliefs is the priority, and omitting it when data cost is the main constraint. Our additional analysis shows that Gold helps on “hard” beliefs, which we define as those that at least three of four reference models (Base, System Prompt, OCT, and SDF) fail. On this subset, adding Gold improves Generalization by 3.5 percentage points and Defend by 2.0 points, as shown below.
Success rate on hard beliefs, averaged across 15 personas.
Without Gold, performance on easy beliefs already exceeds 92%, while data construction cost falls from \3.28to0.42 per persona. This makes the variant without Gold a practical lower-cost option, with Full offering additional gains on difficult beliefs. We have added this analysis and clarified the trade-off in Appendix C.10.