acceptodds
Under review as a conference paper at ICLR 2027

From Natural-Language Descriptions to OCEAN Representations for Dialogue Generation

Abstract

Natural-language descriptions provide an intuitive way to specify conversational agent behavior, but they do not directly support systematic control. We introduce a two-stage framework that maps free-form character descriptions to continuous Five-Factor Model (OCEAN) representations and uses them to guide personality expression in dialogue. NaturalOCEAN combines retrieval-augmented language-model inference with post-hoc calibration to estimate Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. Retrieval generally reduces uncalibrated prediction error, while calibration substantially improves alignment with reference OCEAN scores. OCEANDialogue learns five trait-specific LoRA components whose contributions are scaled by the corresponding OCEAN values. OCEANDialogue produces stronger and more monotonic graded control than numerical zero-shot prompting for Openness, Conscientiousness, and Extraversion. Cross-trait analysis indicates that the learned controls are only partially disentangled, while human and end-to-end evaluations show that intended personality differences are perceptible and that salient characteristics from the original descriptions remain recognizable in generated dialogue. The results support continuous OCEAN representations as an interpretable interface between natural-language character specification and controllable personality expression in dialogue.

Primary area:
foundation or frontier models, including LLMs
Keywords:
OCEAN personality, personality inference, personality-conditioned dialogue generation, controllable language generation, retrieval-augmented generation, Low-Rank Adaptation, large language models.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4NwwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a two-stage framework for personality-conditioned dialogue generation, which first maps natural-language character descriptions into continuous OCEAN scores, and then uses five trait-specific LoRA modules to control dialogue generation.

Critical strengths

S1:Converting natural-language character descriptions into continuous personality control is a reasonable and practically relevant direction.

S2: The work refers to personality inference, controllable dialogue generation, which is end-to-end evaluation within a unified framework.

Critical weaknesses

W1: The paper directly adopts the pipeline “natural-language description → OCEAN → dialogue,” but does not clearly explain the motivation of OCEAN as an intermediate representation, why not directly consider conditioning dialogue generation on the original character description.

W2: The related works are about personality inference and personality-conditioned generation, which neglects the closely related methods like PsychAdapter, BIG5-CHAT, and personality steering methods, please explain the reason.

W3: OCEANDialogue is mainly compared with numerical zero-shot prompting, but omitting several SOTA baselines. In addition, the comparison of “character description → dialogue” and“character description → OCEAN → dialogue.” is neglected. The contribution of the counterfactual objective and the use of five independent trait-specific LoRA modules are also not unclear.

W4: The experimental settings are different, such as retrieval corpus and dialogue training data, and several components rely on LLM-generated descriptions or responses, please give detailed illustration about them.

W5: The evaluation metrics like six-level control slopes, cross-scenario recovery and automatic personality scoring are insufficiently discussed for why use them.

W6: From the results, we can observe the clear improvements of OCEANDialogue for Openness, Conscientiousness and Extraversion, except Agreeableness and Neuroticism, and thus I want to know the reason about them .

Questions for authors

Q1: Why is OCEAN necessary as an intermediate representation? Can you provide a direct comparison of "character description → dialogue" against "character description → OCEAN → dialogue" to show the advantage of the intermediate step over directly conditioning on the original description?

Q2: What specific information does OCEAN capture that the raw description does not, and does the two-stage pipeline lose or distort any description-specific detail?

Q3: How does your approach technically differ from PsychAdapter, BIG5-CHAT, and personality steering methods? Please give a mechanism-level comparison rather than a brief overview, and clarify where this work sits in the development of the research direction.

Q4: The main comparison is against numerical zero-shot prompting. Can you add stronger personality-control baselines (e.g., PsychAdapter, BIG5-CHAT, steering methods) for comparison?

Q5: Can you directly compare "character description → dialogue" vs. "character description → OCEAN → dialogue" as the ablation?

Q6: Can you isolate the contribution of the counterfactual objective and of the five independent trait-specific LoRA modules, rather than focusing the appendix on LoRA rank and scaling?

Q7: The development set, retrieval corpus, dialogue training data, and confirmatory set are built in rather different ways, and several components rely on LLM-generated descriptions/responses. Can you analyze how these distributional differences affect generalization (e.g., train-on-one-source, test-on-another)?

Q8: Why are six-level control slopes, cross-scenario recovery, and automatic personality scoring the most appropriate measures of continuous personality control? Can you justify them and report their validity/reliability relative to alternatives?

Q9: Improvements are clear mainly for Openness, Conscientiousness, and Extraversion, while Agreeableness does not improve and Neuroticism performs worse. Can you explain why, and provide details about the method achieves effective control across all five dimensions? If not, can the claims be revised to reflect the current scope?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 4Nww (Part 1/3)

Response to Reviewer 4Nww (Part 1/3)

We thank the reviewer for their thoughtful feedback and suggestions. We have substantially revised the paper to address their concerns. The revised material is highlighted in blue in the paper. Our responses to individual comments are as follows.

Reviewer Concern:

W1: The paper directly adopts the pipeline “natural-language description → OCEAN → dialogue,” but does not clearly explain the motivation of OCEAN as an intermediate representation, why not directly consider conditioning dialogue generation on the original character description.

W3: In addition, the comparison of “character description → dialogue” and“character description → OCEAN → dialogue.” is neglected.

Q1: Why is OCEAN necessary as an intermediate representation? Can you provide a direct comparison of "character description → dialogue" against "character description → OCEAN → dialogue" to show the advantage of the intermediate step over directly conditioning on the original description?

Q5: Can you directly compare "character description → dialogue" vs. "character description → OCEAN → dialogue" as the ablation?

Response:

We use OCEAN as an intermediate representation not because it is necessary for generating dialogue, but because it provides a structured and interpretable way to control personality. Different character descriptions can be mapped into the same five-dimensional personality space, regardless of their wording. Once a description is converted into an OCEAN profile, we can adjust individual traits without rewriting the original description. This also allows us to separate personality inference from dialogue generation.

To evaluate this approach, we compared direct-description prompting, direct OCEAN prompting, and our learned OCEAN controller. The learned controller achieves the highest high–low effects for four of the five traits and the highest Spearman correlations for two. We also compared the complete direct-description and OCEAN-based pipelines using 100 characters across five conversational scenarios, resulting in 500 matched response pairs. OCEAN-based conditioning achieves significantly better OCEAN fidelity (−0.414-0.414), appropriateness (−0.587-0.587), and naturalness (−0.613-0.613), whereas direct-description conditioning achieves significantly higher description fidelity (+0.478+0.478).

These results also highlight a trade-off. While OCEAN provides a compact, interpretable representation for personality control, it does not preserve all the character-specific details in the original descriptions. Our goal is not to replace direct-description conditioning in all applications, but to enable systematic and reusable control over individual personality traits.

We have included the reason behind using OCEAN in the Introduction and specified the trade-off as a limitation in the Discussion.


Reviewer Concern:

W3: OCEANDialogue is mainly compared with numerical zero-shot prompting, but omitting several SOTA baselines.

Q4: The main comparison is against numerical zero-shot prompting. Can you add stronger personality-control baselines (e.g., PsychAdapter, BIG5-CHAT, steering methods) for comparison?

W2: The related works are about personality inference and personality-conditioned generation, which neglects the closely related methods like PsychAdapter, BIG5-CHAT, and personality steering methods, please explain the reason.

Q3: How does your approach technically differ from PsychAdapter, BIG5-CHAT, and personality steering methods? Please give a mechanism-level comparison rather than a brief overview, and clarify where this work sits in the development of the research direction.

Response:

We have added comparisons with methods that support continuous personality control, Persona Vector and PsychAdapter. We evaluate all three systems under the same conditions using 100 characters and six control levels for each OCEAN trait. OCEANDialogue achieves a higher overall high–low difference (17.25) than Persona Vectors (5.71) and PsychAdapter (9.93), with statistically significant improvements over both baselines. It also achieves the highest Spearman correlation (0.375), with a statistically significant improvement over PsychAdapter. The trait-wise results show stronger control for Openness, Conscientiousness, and Extraversion, while Agreeableness and Neuroticism remain more challenging. We report the overall results in Section 4.2 and the trait-wise results in Appendix.

We focused our competitive evaluation on methods supporting continuous personality control. Since BIG5-CHAT uses discrete high- and low-trait conditions, we haven't included it in the evaluations.

We also included all these methods in the Related Work section and provided the technical differences.

AuthorsOpenReview ↗

Response to Reviewer 4Nww (Part 2/3)

Response to Reviewer 4Nww (Part 2/3)

Reviewer Concern:

W3: The contribution of the counterfactual objective and the use of five independent trait-specific LoRA modules are also not unclear.

Q6: Can you isolate the contribution of the counterfactual objective and of the five independent trait-specific LoRA modules, rather than focusing the appendix on LoRA rank and scaling?

Response:

We have added an ablation study to evaluate the contributions of counterfactual training and independent trait-specific LoRA modules. The results show that the full model achieves higher overall personality controllability than CE-only and both shared LoRA configurations, with statistically significant improvements over CE-only and shared rank-40 LoRA in the high–low effect. We have included the overall ablation results in the main paper and trait-wise analysis in the appendix.


Reviewer Concern:

W5: The evaluation metrics like six-level control slopes, cross-scenario recovery and automatic personality scoring are insufficiently discussed for why use them.

Q8: Why are six-level control slopes, cross-scenario recovery, and automatic personality scoring the most appropriate measures of continuous personality control? Can you justify them and report their validity/reliability relative to alternatives?

Response:

Our evaluation metrics capture complementary aspects of continuous personality control. Six-level regression slopes quantify the magnitude and direction of perceived trait changes, and Spearman correlation measures monotonic alignment. A method can exhibit a large positive slope despite inconsistent ordering at intermediate levels, which Spearman correlation helps identify. As an additional validation, we computed Kendall's τb\tau_b on the same six-level evaluation scores used for Spearman correlation. Both measures show consistent trends, with OCEANDialogue improving over zero-shot prompting for Openness (0.123 to 0.242), Conscientiousness (0.170 to 0.303), and Extraversion (0.183 to 0.385), but not for Agreeableness or Neuroticism. These findings show that our conclusions about personality control are consistent across two rank-based evaluation measures.

Cross-scenario recovery measures how consistently positive control effects emerge when aggregating responses across different conversational contexts. Since this is a supplementary aggregation analysis, we have moved it to the Appendix.

Automatic personality scoring enables consistent evaluation of perceived trait expression across control levels and scenarios. To assess its reliability, we conducted cross-model comparisons between Claude Sonnet 4.5 and GPT-4.1, human perceptual evaluations, and judge-prompt sensitivity analyses. The two AI evaluators show strong agreement on high–low personality judgments (83.67–97.40% across traits), while human evaluations confirm that intended personality differences are perceptible (65.9% of non-tie judgments). However, prompt sensitivity indicates that absolute scores depend partly on evaluation design.


Reviewer Concern:

W6: From the results, we can observe the clear improvements of OCEANDialogue for Openness, Conscientiousness and Extraversion, except Agreeableness and Neuroticism, and thus I want to know the reason about them.

Q9: Improvements are clear mainly for Openness, Conscientiousness, and Extraversion, while Agreeableness does not improve and Neuroticism performs worse. Can you explain why, and provide details about the method achieves effective control across all five dimensions? If not, can the claims be revised to reflect the current scope?

Response:

We have expanded the Discussion to explain the differences across traits and clarify the scope of our claims. For Agreeableness, zero-shot prompting already provides relatively strong graded control, which potentially limits further improvement. For Neuroticism, prior research suggests that personality expression is particularly sensitive to situational context, including stressful interactions. We also acknowledge that our counterfactual objective distinguishes opposing personality conditions but does not explicitly enforce monotonicity across intermediate values. We have revised our claims to emphasize that OCEANDialogue significantly improves continuous control for Openness, Conscientiousness, and Extraversion, but does not consistently outperform zero-shot prompting for Agreeableness and Neuroticism.

AuthorsOpenReview ↗

Response to Reviewer 4Nww (Part 3/3)

Response to Reviewer 4Nww (Part 3/3)

Reviewer Concern:

Q2: What specific information does OCEAN capture that the raw description does not, and does the two-stage pipeline lose or distort any description-specific detail?

Response:

OCEAN represents personality along five continuous dimensions, providing an interpretable representation of personality traits inferred from natural-language descriptions. However, character descriptions may contain specific behavioral details that are not fully captured by these five dimensions.

Our direct-conditioning comparison supports this trade-off. OCEAN-based conditioning achieves significantly better OCEAN fidelity, appropriateness, and naturalness, whereas direct-description conditioning achieves significantly higher description fidelity. No statistically significant difference is observed in behavioral signature. These results indicate that OCEAN provides a structured representation for personality control but does not preserve all character-specific information from the original descriptions.


Reviewer Concern:

Q7: The development set, retrieval corpus, dialogue training data, and confirmatory set are built in rather different ways, and several components rely on LLM-generated descriptions/responses. Can you analyze how these distributional differences affect generalization (e.g., train-on-one-source, test-on-another)?

W4: The experimental settings are different, such as retrieval corpus and dialogue training data, and several components rely on LLM-generated descriptions or responses, please give detailed illustration about them.

Response:

We added a table to Appendix summarizing all data sources. We clarified the differences among the data sources and examined cross-source generalization through the confirmatory evaluation. NaturalOCEAN's configuration and calibration slopes are determined on the development set, which contains LLM-generated descriptions and reference scores. The confirmatory holdout uses descriptions generated by a different LLM and crowdsourced personality ratings. We retain the development-set calibration slopes and apply label-free batch centering without using holdout reference scores. Calibrated R² improves for four of the five traits on the holdout, providing evidence of transfer across these data sources.

OCEANDialogue is trained on Llama-generated responses and evaluated on held-out conversational contexts using independent LLM judges and human raters. While these evaluations demonstrate generalization across different data sources, they do not isolate the effects of individual distribution shifts. Evaluation using human-written character descriptions remains a limitation.

Reviewer_3d2bOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a two-stage framework for controllable dialogue generation. The framework combines NaturalOCEAN and OCEANDialogue to convert free-form character descriptions into continuous Big Five personality representations and to use those representations to control dialogue generation. NaturalOCEAN retrieves similar character profiles and predicts five trait values, while OCEANDialogue trains five trait-specific LoRA modules on synthetic dialogue scenarios and scales their contributions according to a target personality profile. Experimental results confirm the effectiveness of the framework.

Critical strengths
  • The paper addresses a practical and important problem in dialogue generation, as fine-grained trait control is essential for making dialogue systems adaptable to diverse users and scenarios.
  • The proposed framework is interpretable and operationally simple, enabling users to specify a profile and inspect trait-wise effects.
Critical weaknesses
  • The novelty of this paper is limited. The idea of controlling traits through real-valued vectors and specialized LoRA experts has been extensively explored in prior work, such as P-React, PsychAdapter, and Personality Vector. This work appears to only combine existing methods and introduces no new concepts or technical contributions.
  • The paper does not compare any existing personality-control baselines. Its advantages over existing approaches therefore remain unsubstantiated.
  • The clarity of the paper needs substantial improvement. The authors rely on excessive prose and provide no diagram illustrating the overall pipeline of the proposed framework.
  • Much of the training data and several evaluation components are generated by LLMs. Their reliability has not been verified.
Questions for authors

See above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 3d2b

Response to Reviewer 3d2b

We thank the reviewer for their feedback and suggestions. We have substantially revised the paper to address their concerns. The revised material is highlighted in blue in the paper. Our responses to individual comments are as follows.


Reviewer Concern:

The novelty of this paper is limited. The idea of controlling traits through real-valued vectors and specialized LoRA experts has been extensively explored in prior work, such as P-React, PsychAdapter, and Persona Vector. This work appears to only combine existing methods and introduces no new concepts or technical contributions.

Response:

We agree that continuous personality representations and LoRA-based adaptation have been explored in previous work. Our contribution is a specific training formulation for continuous personality control with a framework that connects natural-language character descriptions to personality-conditioned dialogue. We evaluate both components of the framework individually and the whole pipeline end to end.

Our technical contribution lies in combining signed trait-specific adaptation with counterfactual likelihood supervision. OCEANDialogue introduces a counterfactual training objective and directly scales five LoRA components using signed OCEAN coefficients and compares the likelihood of the same response under intended and opposing trait conditions. This differs from P-React's personality specialization loss, which encourages differentiation among LoRA experts, and from PsychAdapter's continuous conditioning approach, which relies on language-modeling supervision. The ablation results support our design choices. Removing counterfactual supervision reduces the high–low effect from 10.80 to 4.01. We also evaluate shared-LoRA alternatives, including a rank-40 configuration with a comparable parameter budget. Both shared-LoRA variants exhibit significantly lower absolute high–low effects than the full model, providing additional evidence for the benefit of independently parameterized trait-specific adaptations.

We have clarified these technical contributions in the revised manuscript.


Reviewer Concern:

The paper does not compare any existing personality-control baselines. Its advantages over existing approaches therefore remain unsubstantiated.

Response:

We have added comparisons with methods that support continuous personality control, Persona Vector and PsychAdapter. We evaluate all three systems under the same conditions using 100 characters and six control levels for each OCEAN trait. OCEANDialogue achieves a higher overall high–low difference (17.25) than Persona Vector (5.71) and PsychAdapter (9.93), with statistically significant improvements over both baselines. It also achieves the highest Spearman correlation (0.375), with a statistically significant improvement over PsychAdapter. We have added these results to Section 4.2 and included trait-wise comparisons in Appendix.


Reviewer Concern:

The clarity of the paper needs substantial improvement. The authors rely on excessive prose and provide no diagram illustrating the overall pipeline of the proposed framework.

Response:

We have substantially revised the organization of the manuscript. We now include a diagram illustrating the overall pipeline in Figure 1.


Reviewer Concern:

Much of the training data and several evaluation components are generated by LLMs. Their reliability has not been verified.

Response:

To assess the reliability of LLM-based personality evaluation, we conducted an additional cross-model agreement analysis. GPT-4.1 showed significant directional agreement with the original judge across all five OCEAN traits, while the human study showed significant agreement for Openness and Conscientiousness. To assess synthetic training data quality, we evaluated 1,000 sampled dialogue responses using GPT-5-mini, obtaining average naturalness and contextual relevance scores of 4.79/5 and 4.62/5, respectively. We have included the trait-wise results in Appendix.

Reviewer_7n3gOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors provide a two-stage pipeline that allows to use LLM to properly model personalities according to predefined OCEAN traits.

Critical strengths

The paper is, first of all, clear and well motivated: in many application scenarios, e.g., marketing, modelling user behavior correctly using OCEAN description can be of fundamental importance. Also, the authors test several different models and multiple retrieval methods, as well as use multiple large datasets. They also performed ablation studies on retrieval and query augmentation, which improve the robustness of their findings.

Critical weaknesses
  1. The human evaluation analysis probably should be done per trait (it might reinforce previous results on the differences between OCE and A N)
  2. The full-pipeline evaluation seems poorly designed, given that the models only use scores of 2 and 3. Probably, as judges, the model should be better instructed, perhaps using few-shot prompting techniques.
  3. I think the authors could expand more on what it means, concretely, that certain traits (OCE) can be better reproduced than others (AN), for example in marketing, consumer modelling, ...?! (A), or in AI-safety applications (N)

Minor:

  • why reporting cross trait correlations in fig 1? (I guess it becomes clearer afterwards, but perhaps they could comment on these out-of-diagonal similarities).
  • table 3: I don't understand the scale on which the traits are evaluated. 1-100? Also, perhaps they could comment on the different baseline levels, for example for A (which starts from 35 in ZS), more than twice as C.
Questions for authors

I don't have additional questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewer 7n3g

Response to Reviewer 7n3g

We thank the reviewer for their feedback and suggestions. We have substantially revised the paper to address their concerns. The revised material is highlighted in blue in the paper. Our responses to individual comments are as follows.

Reviewer Concern:

The human evaluation analysis probably should be done per trait (it might reinforce previous results on the differences between OCE and A N)

Response:

We have added a trait-wise analysis of the human evaluation results. For OCEANDialogue, the highest intended-direction preference rates are observed for Extraversion (88.6%) and Openness (86.5%), while Neuroticism shows the lowest rate (51.4%). These results indicate that perceived personality differences vary across the five OCEAN traits, consistent with the differences observed in our automated evaluation.


Reviewer Concern:

The full-pipeline evaluation seems poorly designed, given that the models only use scores of 2 and 3. Probably, as judges, the model should be better instructed, perhaps using few-shot prompting techniques.

Response:

To assess the reliability of LLM-based personality evaluation, we conducted a prompt sensitivity analysis using Claude Sonnet 4.5. The results showed significantly lower scores with few-shot prompting (P2), while the detailed rubric (P1) and evidence-based prompting (P3) showed no significant difference from the original rubric (P0). Scores remained restricted to 2–3 across all strategies. We have included these results in the appendix.


Reviewer Concern:

I think the authors could expand more on what it means, concretely, that certain traits (OCE) can be better reproduced than others (AN), for example in marketing, consumer modelling, ...?! (A), or in AI-safety applications (N)

Response:

We have included a paragraph about downstream applications in the Discussion.


Reviewer Concern:

why reporting cross trait correlations in fig 1? (I guess it becomes clearer afterwards, but perhaps they could comment on these out-of-diagonal similarities).

Response:

The cross-trait correlation matrix was intended to examine relationships between NaturalOCEAN's predicted and reference OCEAN scores. The off-diagonal correlations may reflect associations among the reference personality traits or imperfect separation in the predictions; however, the analysis does not distinguish between these explanations. We have moved the figure to Appendix and added a brief discussion.


Reviewer Concern:

table 3: I don't understand the scale on which the traits are evaluated. 1-100? Also, perhaps they could comment on the different baseline levels, for example for A (which starts from 35 in ZS), more than twice as C.

Response:

The requested OCEAN levels range from 0 to 1, while the evaluator assigns perceived-trait scores on a 0–100 scale. We have clarified this in the table caption.

Baseline scores vary across traits. At the requested level of 0.0, zero-shot Agreeableness scores 35.30, compared with 14.26 for Conscientiousness. These differences indicate varying levels of perceived trait expression. Therefore, our analysis focuses on changes across control levels within each trait rather than comparing absolute scores across traits.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.