acceptodds
Under review as a conference paper at ICLR 2027

Self-Distillation Improves Persona Belief Internalization

Abstract

Language models can be prompted with persona descriptions or fine-tuned on persona-specific dialogues to generate responses that reflect a target persona's speaking style and background. However, such behavioral fidelity can remain shallow, failing to capture deeper persona-specific patterns across contexts, and does not necessarily imply corresponding changes in the factual beliefs that the model internally represents as true. While recent work has exposed this gap between persona-consistent behavior and internal truth representations, how to train models to internalize the factual worldview of a target persona remains comparatively underexplored. We study this problem as persona belief internalization and introduce Persona-SDFT: an on-policy self-distillation method that transfers a persona's worldview into a model from its own predictions conditioned on privileged context about that persona. We find that combining on-policy self-distillation with carefully structured persona-specific privileged context substantially strengthens persona belief internalization. On Qwen3-8B across 15 historical personas, Persona-SDFT improves generalization to related persona beliefs from 55.5% to 77.4% and robustness under challenge from 45.6% to 66.8% compared with Open Character Training, while increasing truth-representation probe lift from 0.190 to 0.346, which we use as an operational proxy for belief internalization. These results suggest that persona training can go beyond surface behavioral fidelity to promote the internalization of persona-specific factual worldviews.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Model, Role Playing, Personalization, Persona

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_csK5OpenReview ↗
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
  1. 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.

  2. 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
AuthorsOpenReview ↗

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]).

MethodAnswers consistent w/ belief (%)
Base7.1
System Prompt23.3
SFT25.3
SDF30.6
OCT32.9
Persona-SDFT44.5

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%.

Edit to Persona-SDFTUneditedStrength ×1Strength ×2Change at ×2 [95% CI]
Subtract truth direction67.624.24.5−63.1 [−69.2, −56.1]
Add random direction (norm-matched)67.667.063.7−3.9 [−6.4, −1.6]

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.

Hard beliefsw/o Goldw/ Gold (Full)Gain
Generalization (%)53.957.4+3.5
Defend (%)52.954.9+2.0

Success rate on hard beliefs, averaged across 15 personas.

Without Gold, performance on easy beliefs already exceeds 92%, while data construction cost falls from \3.28to3.28 to \\0.42 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.

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.

AuthorsOpenReview ↗

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

Reviewer_JL9jOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary
  • A large recent body of work has aimed to align models to personas by making their outputs reflect those of a specific person or role. However, existing training methods for doing so may not be fully internalized by the model, leading to role-playing that is inconsistent out of distribution and brittle under adversarial pressure.
  • This work makes two contributions to persona post-training: (1) constructing “persona universes” that organize beliefs held by the target persona into a coherent structure, then (2) using self-distillation finetuning (from a teacher who holds the persona universe in-context) to align a model with a persona.
  • Across 15 historical personas and two models, a combination of these two methods consistently outperforms existing baselines (open character training, synthetic document finetuning) at internalizing persona-specific beliefs, via both behavioral and internals-based evaluation measures.
Critical strengths
  • Character training is an important topic that has attracted a lot of recent attention; the paper’s claimed gains would represent substantial empirical progress on this problem.
  • The paper focuses on two relatively simple ideas for improving persona training, which are well-motivated - I thought the analysis in Figure 3 was a good way to motivate the usage of SDFT over SFT.
  • The paper’s ablations and experiments are relatively thorough - they find limited alignment tax from their method and that their results replicate across models.
Critical weaknesses
  • The authors generally reuse the training setups referenced in past work, rather than tuning baselines, and don’t match compute usage between their work and the baselines. While they mention this in the paper, they should either tune baselines and match compute, or scope down the claims made in the abstract and introduction about the relative performance of their methods.
  • The actual structure of the persona universe is somewhat unclear. What makes the persona universe materially different from other methods of conveying privileged information, and what do they look like qualitatively? While the ablations show that the universe does better align models toward a persona, the explanation of how this happens is somewhat limited when compared to the SDFT vs SFT analysis - is it just having additional context, or would a length-matched ablation still favor the persona universe?
Questions for authors
  • Are there any differences in evaluation setting between this work and Sturgeon’s setup? Either way, I would make explicit that you’re building off of past work’s evaluation setups, and clarify any changes you made to adapt to your models.
  • The work focuses on the relatively narrow persona space of historical characters (from Sturgeon). How do you think evaluating this method on broader character dispositions, as in Maiya, would change the results? I suspect that OCT would do relatively better on broader dispositions than specific factual alignment.
  • Why do you think that Persona-SDFT has lower persona alignment and persona adoption scores than baselines (Table 17)? I think it’s reasonable to trade off some alignment for more robust internalization of beliefs, but you should definitely state that this is what the method is doing in the main text if so.
4 comments
AuthorsOpenReview ↗

Response to Reviewer JL9j (1/2)

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] Baseline tuning and compute fairness?

Our evaluation aims to compare the methods under comparable training conditions while staying faithful to their original designs. For OCT, we follow the released pipeline for the same personas evaluated in our experiments [1]. For SDF, we follow the published document-generation procedure [2] and match its training configuration to Persona-SDFT’s, including LoRA rank, learning rate, and effective batch size. Persona-SDFT itself retains the upstream SDFT objective and EMA schedule. We have clarified these implementation choices in Section 4 and Appendix A.3.

In terms of training compute, Persona-SDFT achieves stronger results even with less GPU time than either baseline. At step 200 of 390, it uses 32.9 minutes per persona, compared with 37.8 for SDF and 67.2 for OCT, while outperforming both on all four belief measures. We have incorporated this comparison into Section 4.3 and Appendix C.4, and made the scope of our comparisons explicit in the abstract and introduction.

MethodTraining time (min)Probe LiftGen. (%)Defend (%)Belief Endorse. (%)
OCT67.2±2.80.190±0.06255.5±9.045.6±21.925.1±13.6
SDF37.8±1.30.233±0.07059.7±10.256.5±23.337.9±14.6
Persona‑SDFT, stopped at step 20032.9±3.50.302±0.04574.0±8.865.6±19.363.1±12.9
Persona‑SDFT, full run63.7±6.40.346±0.05877.4±7.766.8±21.068.3±11.9

Qwen3-8B, mean ± standard deviation across 15 personas. Training time is measured per persona on an NVIDIA B200 GPU instance.


[W2] Further analysis of the persona universe

Following your suggestion, we have added a qualitative comparison showing how the universe provides a shared factual worldview, while Fact and Gold provide question-specific information derived from it. Fact identifies the relevant belief among 16 key facts consolidated from the universe; Gold illustrates its application through a verified answer. An example for Marie Curie is shown below:

Universe (Avg. tokens: 848, shared by all questions)"By 1934 radioactivity is one of the most exact branches of physical science. […] Radium is the most intensely radioactive element known. […] The neutron, discovered in 1932, is a bound state of a proton and an electron rather than a wholly new elementary particle. […]"
Matched fact (Avg. tokens: 32, question-specific)"Radium is the most intensely radioactive element known, and polonium has no significant practical application of its own beyond scientific research […]"
Gold answer (Avg. tokens: 51, question-specific)Question: "Which element is described as the most intensely radioactive element known?" Answer: "Radium is described as the most intensely radioactive element known. […]"

Since these components differ in information coverage and function, directly matching their lengths would be less informative. We therefore tested whether the universe’s advantage persists against alternative shared contexts of comparable length. Specifically, keeping all other settings unchanged, we replaced the universe with either (i) a naive context generated directly from the same persona profile and 30 target beliefs, without our intermediate grounding and verification steps, or (ii) the OCT constitution for the same persona. The universe performs best on all four measures, supporting the value of its construction beyond context length alone. We have added these comparisons in Section 4.2 and Appendix C.9, and the qualitative examples in Appendix D.

ContextAvg. TokensProbe LiftGen. (%)Defend (%)Endorse (%)
Universe8480.33872.864.568.1
Naive context8560.32371.249.465.3
OCT constitution7970.25063.261.848.3
AuthorsOpenReview ↗

Response to Reviewer JL9j (2/2)

[Q1] What changed from the evaluation setup of Sturgeon et al. [1]?

We largely retain their evaluation setup, including the personas, era-believed statements, and evaluation protocols. The main adjustments are (i) probing at layer 15 of Llama-3.1-8B-Instruct and (ii) using GPT-5.6 Luna as the judge in place of Claude Sonnet 4.6. Building on this setup, we add First-Turn Belief Endorsement (in Section 4.1) and the persona-versus-modern-truth evaluation (in Section 4.4), along with further evaluations of whether models defend their own initial answers (in Appendix A.5) and retain capabilities in neutral contexts (in Appendix C.6). We have clarified how our evaluation builds on theirs in Section 4 and Appendix A.5 of the revision.


[Q2] Extension to broader character dispositions

Thank you for this interesting question. We agree that broader dispositions, such as “loving,” are an important aspect of persona learning. Our current study focuses on factual beliefs, asking whether models internalize a persona’s worldview beyond reproducing its outward behavior. We think extending Persona-SDFT to such dispositions would be a promising future direction. This would require clarifying what internalization means for these traits and designing privileged context and evaluations that capture how they shape responses across situations. We have clarified this scope and potential extension in Section 5.


[Q3] Appendix C.7: On lower Alignment and Adoption scores

Our training focuses on learning the persona’s factual beliefs, while SFT and OCT directly train on in-character responses. This difference may contribute to the lower Persona Alignment and Identity Adoption scores despite stronger performance on belief-related measures.

Consistent with this distinction, we find that Persona-SDFT produces the largest truth-probe increase on a shared set of statements drawn equally from the five methods’ Alignment responses (+0.162, versus +0.097 for SFT and +0.028 for OCT). These results suggest that higher Alignment scores do not necessarily imply stronger internal truth signals for the same statements. We have discussed this potential trade-off between belief internalization and broader persona fidelity in Section 4.1 of the revision.

MethodΔ truth-probe score
Persona-SDFT+0.162
SFT+0.097
SDF+0.094
System Prompt+0.076
OCT+0.028

Native truth probe, relative to the base model without a persona prompt, averaged across 15 personas.


[1] Sturgeon et al., When Role-playing, Do Models Believe What They Say? 2026

[2] Wang et al. Modifying LLM Beliefs with Synthetic Document Finetuning. 2025

Thanks for your thorough response and for running these additional experiments! Many of my original soundness concerns have been addressed by the revision, so I've increased my soundness score from 2 to 3. I continue to support this paper's acceptance to the conference, as it's a sound empirical contribution to an important research direction.

AuthorsOpenReview ↗

Response to Reviewer JL9j's Follow-up Comment

Dear Reviewer JL9j,

Thank you for your follow-up and continued support. We are glad that our responses and revisions have helped address your concerns. Your thoughtful feedback has greatly helped us improve our manuscript.

Sincerely,
Authors

Reviewer_bbmqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates persona belief internalization, aiming to train language models to reflect not only the behavioral style of a target persona but also its factual worldview. The authors propose Persona-SDFT, an on-policy self-distillation framework in which a teacher is provided with structured privileged information, including a shared persona universe, matched facts, and verified gold answers, while the student only receives the persona prompt and question. The method is evaluated on 15 historical personas using Qwen3-8B and Llama-3.1-8B-Instruct.

Critical strengths
  1. The paper investigates an important distinction between generating persona-consistent responses and internalizing persona-specific factual beliefs. This distinction is relevant to persona modeling, knowledge representation, and belief modification in language models.
  2. The evaluation spans 15 diverse personas and two backbones and compares to strong baselines, including a belief-modification method and widely used persona-training pipelines. Multiple metrics assess complementary facets: probe lift, generalization to related questions, robustness under adversarial challenge, and first-turn endorsement; an additional “persona vs modern truth” test probes competition with pretrained knowledge.
  3. The evaluation combines behavioral metrics with probing of internal representations. In particular, Section 4.4 directly contrasts persona-consistent beliefs with modern factual corrections, providing a useful assessment beyond simple belief endorsement. The paper also discusses the limitations of its approach rather than presenting belief internalization as fully solved.
Critical weaknesses
  1. The proposed approach is technically close to existing self-distillation and context-distillation frameworks. Most of its core ingredients—on-policy trajectories, KL-based teacher supervision, EMA updates, and privileged contextual information—are adopted from prior work. The main novelty therefore seems to lie in how these components are instantiated for persona belief internalization. Although this leads to strong empirical results, the paper provides limited evidence that the resulting framework introduces a broadly reusable methodological principle beyond this specific application.
  2. Although the paper evaluates belief internalization using both behavioral metrics and internal representation probes, the results do not fully establish whether the model's underlying factual beliefs have changed. Improvements in probe scores may also reflect stronger persona-conditioned representations rather than genuine belief modification. Notably, Section 4.4 shows that the model still favors modern factual corrections over persona beliefs on average (P(Persona) = 46.1%). More direct evidence, such as causal interventions on internal representations, would help strengthen the central claim of belief internalization.
  3. The ablation results in Table 1 raise questions about the necessity of the complete privileged-information design. The Universe-only variant achieves performance close to the full method on several metrics, while removing the matched fact or gold answer even improves certain metrics. These results suggest that the shared persona universe may account for a substantial portion of the gains, while the additional contributions of the other components remain unclear. Further analysis is needed to establish whether the full design consistently provides meaningful advantages over simpler alternatives.
Questions for authors

NA

5 comments
AuthorsOpenReview ↗

Response to Reviewer bbmq (1/2)

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] Technical novelty and broader implication

We instead view the surprising effectiveness of our simple approach as a key novelty of this work. Persona belief internalization has been actively studied through synthetic-document training (e.g., SDF) and additional preference optimization (e.g., OCT); yet a standard self-distillation procedure, with carefully designed teacher context, outperforms both on all four belief measures in a single training stage. The context design is still crucial, e.g., Belief Endorsement rises from 33.7% with gold-answer-only context to 68.1% with the shared universe. The contribution is discovering how effectively an existing, simple procedure can solve this problem and establishing the role of privileged-context design in that success.

This finding also informs the ongoing discussion of when SDFT works well, beyond its effectiveness for personas. Our results suggest that a coherent persona context allows the model to turn unfamiliar beliefs into supervision suited to its own generation process. This interpretation is consistent with findings that self-generated supervision reduces the distribution gap in fine-tuning [1], and that self-distillation is effective when privileged information captures shared rules across inputs [2]. Persona belief internalization brings these two elements together: a shared belief system and supervision supplied by the contextualized model itself.


[W2] Direct evidence for belief modification?

We agree that measuring belief in LLMs remains an evolving research challenge. Our study builds on the behavioral evaluations and truth-representation probes used in recent work on LLM beliefs [3], and demonstrates consistent improvements under these measures. For example, as you note, P(Persona) = 46.1% still indicates an average preference for modern factual corrections, but it represents a substantial improvement over the base model’s 12.3% and SDF’s 32.8%. Together with gains across all four belief measures, this result supports a meaningful shift toward the persona’s worldview.

To further strengthen this evidence, we conducted an additional intervention experiment following your suggestion. We derived a general truth direction from the difference in mean activations between generic true and false statements at layer 24 of Qwen3-8B, the layer used for probing, 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%. This substantially larger effect supports a causal connection between truth-related activations and persona belief endorsement, complementing our behavioral and probe-based findings. We have incorporated this experiment in Appendix C.8 and clarified the scope of our belief-internalization claims in Section 4 and Appendix A.5.

Edit to Persona-SDFTUneditedStrength ×1Strength ×2Change at ×2 [95% CI]
Subtract truth direction67.624.24.5−63.1 [−69.2, −56.1]
Add random direction of the same size67.667.063.7−3.9 [−6.4, −1.6]

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.

AuthorsOpenReview ↗

Response to Reviewer bbmq (2/2)

[W3] Table 1: Necessity of the complete privileged information?

We recommend Full for its consistent performance across complementary belief measures, rather than because every component improves every metric. Full exceeds Universe-only on all four measures and is the only configuration ranked first or second on each in Table 1. Its largest gain over Universe-only is in Generalization (+4.7% points, 95% CI [+2.1, +7.0]), improving in 12 of 15 personas. We agree that the shared universe accounts for much of the gain, while Fact and Gold provide further benefits.

Our additional analysis shows that Gold helps the model better learn “hard” beliefs; 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. Alongside Table 1’s finding that Fact improves Generalization and Belief Endorsement both with and without Gold, these results support the benefits of including both components. We have added this analysis in Appendix C.10 and further clarified our rationale for Full in Section 4.2.

Hard beliefsw/o Goldw/ Gold (Full)Gain
Generalization (%)53.957.4+3.5
Defend (%)52.954.9+2.0

Success rate on hard beliefs, averaged across 15 personas.


[1] Yang et al. Self-Distillation Bridges Distribution Gap in Language Model Fine-Tuning. ACL 2024.

[2] Zhu et al. The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes. 2026.

[3] Sturgeon et al. When Role-playing, Do Models Believe What They Say? 2026.

Dear Reviewer bbmq,

Thank you again for your time and effort in reviewing our manuscript.

As the discussion period is coming to an end, we would like to kindly ask if there are any remaining points that you would like us to clarify.

We believe we have made a sincere effort to address your earlier comments, and we would be grateful for any additional feedback you may have. Your insights are invaluable in helping us further strengthen the manuscript.

Sincerely,
Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revision

We sincerely appreciate your time and effort in reviewing our paper. Your constructive comments have been invaluable in improving our manuscript.

We have addressed your comments in the revised manuscript, with all additions and revisions highlighted in blue. Specifically, we have made the following changes:

  • Added a causal intervention on truth-related activations to assess their effect on persona belief endorsement (bbmq W2; csK5 W1).
  • Clarified the contributions of Fact and Gold beyond the shared universe, including an analysis of Gold’s benefits on difficult beliefs (bbmq W3; csK5 Q2).
  • Clarified the scope of the baseline comparisons and added a comparison showing that Persona-SDFT retains its advantage with less training GPU time than either baseline (JL9j W1).
  • Added qualitative examples of the privileged information and comparisons with a naive context and the OCT constitution at comparable lengths (JL9j W2).
  • Clarified how our evaluation builds on the setup of Sturgeon et al., including the adjustments and additional evaluations (JL9j Q1).
  • Discussed extending Persona-SDFT to broader character dispositions (JL9j Q2).
  • Discussed the potential trade-off with persona fidelity and added a truth-probe comparison using shared statements extracted from Alignment responses (JL9j Q3).
  • Added an open-ended held-out evaluation without prefilling or explicitly presenting the target belief (csK5 W1/Q1).

We hope these revisions address your concerns and make our manuscript clearer. We welcome any further questions or suggestions.

Sincerely,
Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.