acceptodds
Under review as a conference paper at ICLR 2027

Localizing and Reusing Persona-Induced Representations in Language Models

Abstract

Large language models are increasingly used in settings that require strategic judgment.Persona prompting is a common way to elicit different behaviors from language models by asking them to adopt specific roles or personality traits.However persona descriptions do not map straightforwardly onto behavioral responses. A persona described as cooperative may still defect in a particular setting, while combining multiple personas leaves their joint behavioral influence unclear.To better understand this phenomenon and enable more targeted behavioral control, we use contrastive persona prompts to induce behaviorally verified cooperation and defection in a canonical Prisoner’s Dilemma (PD), providing a sparse feature ensemble that reliably distinguishes persona-induced cooperative and defective decisions and serves as a target for causal intervention.Our results make three contributions. First, we develop a workflow for localizing persona-related SAE features from behaviorally verified persona contrasts. Second, by comparing across decision-making games, we find that a shared core of features can predict behavioral outcomes, while task-specific features are required for effective intervention. Third, we use these representations to interpret persona-induced behavior and demonstrate a proof-of-concept route from a target feature pattern back to a readable persona prompt, providing an initial path toward interpretable and controllable prosocial behavior in language models.

Primary area:
interpretability and explainable AI
Keywords:
Persona Prompting, Sparse Autoencoders (SAEs), Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XsR7OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

This paper explores how persona descriptions map to behavioral responses in economics games, especially the Prisoner's Dilemma game. It localizes the persona-related sparse autoencoder (SAE) features associated with cooperative decisions. Experiments show that overriding such features can intervene the model behavior, and the writability shows a sharp transition at layer 15 in the tested model. The paper also shows as a proof of concept that the discovered feature signature can guide persona synthesis.

Critical strengths
  1. This paper explores the machanisms on how persona descriptions translate into model behaviors, showing the potential of making the process more interpretable and controllable.
  2. The experiment design uses multiple output-label mappings to separate behavioral features from features tied to answer tokens.
  3. Experiments are also conducted on other games such as Dictator to demonstrate how the discovered features
Critical weaknesses
  1. The paper is extremely hard to read even though I have plenty of research experience in this area and some background knowledge on activation-patching. Many abbreviations and concepts are used without providing definitions or contexts. For example, in the abstract and introduction, keywords like "SAE", "L15", "train-selected AB/CD set" occur multiple times, but late until Related Work or even Method, their meaning are given. The figure, table, and algorithm captions are also not helping. They merely describe how results are obtained and how variables are defined. The experiments and analyses are very unclear. For example, how the predictors are fitted? What are the left-hand side feature and what are the right-hand side outcomes? How the R2R^2 is measured? The Conclusion is also full of terms like "register" which is never formally defined. Overall, the whole paper reads like an AI-written one without careful polish.
  2. The paper poses a great research question on exploring the machenism of persona prompting, but the scientific implications of the findings, such as the phase transition in SAE layers and writable features, still remain unclear and underdiscussed.
  3. Only one subject model (Llama-3-8B-Instruct) has been tested, making it difficult to conclude if the discovered SAE features are model-specific or can also be detected on other LLMs.
  4. Some concepts seem to be defined arbitrarily. For example, "A feature is considered cooperation-associated under mapping m if it shows one-sided support for cooperative behavior: it is activated in at least k=5 verified cooperative rows while remaining inactive across all verified defective rows". Why using k=5k=5 as the deciding boundary? The rigor of the paper could benefit from sensitivity checks on the configurations or a better justification written in the paper.
  5. Line 235 suggests that 9 responses are generated for the cross-game analysis, and many other experiments use n=10n=10. Given the randomness within LLM output generation, are the samples big enough to reliably make statistical conclusions?
Questions for authors

Please refer to the Critical Weaknesses section. In general, I think the authors could consider substantially revising the writing to better deliver the ideas and results.

4 comments

Q1(Weakness1):

Response.

We sincerely thank the reviewer for this detailed feedback and helpful suggestions. To make the key concepts, experimental procedures, and findings more accessible and clearly presented, we have further revised the manuscript, with particular attention to terminology, methodological descriptions, and the presentation of experimental results.

Specifically, we have made the following revisions:

  1. Terminology and definitions: We have introduced key abbreviations and concepts at their first occurrence and clarified terms such as SAE, L15, and the feature-selection procedure. We have also simplified technical shorthand in the Abstract and Introduction to make the main ideas more accessible.

  2. Figures, tables, and algorithm descriptions: We have revised the corresponding captions to better explain the purpose of each experiment, the meaning of the variables, and the interpretation of the reported results, rather than focusing primarily on procedural details.

  3. Methodological clarity: We have expanded the explanations of the prediction and readout procedures, explicitly clarifying the input features, prediction targets, model-fitting procedures, and evaluation metrics. This makes the relationship between SAE activations and behavioral outcomes more transparent.

  4. Experimental analysis: We have reorganized and clarified several experimental descriptions to better explain the motivation, evaluation procedures, and interpretation of the findings.

  5. Conclusion and overall presentation: We have clarified ambiguous terminology, including terms that were not sufficiently explained, and revised the Conclusion to better summarize the main findings and their scientific implications. Finally, we have reviewed the manuscript to improve terminology consistency, logical transitions, and academic writing throughout. We appreciate the reviewer’s suggestions, which helped us further improve the clarity and accessibility of the presentation.

Q2(Weakness2):

Response

We thank the reviewer for this insightful comment. Following the reviewer's suggestion, we have expanded the Discussion section to further interpret the layer-wise writability results, explain the significance of writable features, and highlight the broader implications of our findings for understanding persona prompting.

First, regarding the observed layer-wise transition, our results show that interventions at earlier layers (e.g., L13) do not effectively alter the final decision, whereas intervention effects emerge around L15 and become substantially stronger in subsequent layers. We interpret this finding as suggesting a distinction between the formation of persona-related representations and their causal accessibility to the decision process.

Second, the writable features provide evidence beyond behavioral prediction or correlation. Through targeted interventions along the corresponding SAE decoder directions, we demonstrate that manipulating these features can systematically influence subsequent decisions in the expected directions. This provides evidence that the identified SAE directions can also exert causal effects on model decisions under our intervention settings.

Finally, the revised Discussion brings together our main findings and their broader implications. Behaviorally verified persona contrasts enable the identification of behavior-associated SAE features; targeted interventions reveal their causal influence on decisions; and cross-game evaluations show that behavioral readout generalizes more broadly than intervention effects, which remain more task-dependent. Together, these findings suggest that persona-induced behavioral tendencies can be studied at the level of internal representations, while highlighting the distinction between representing behavioral tendencies, predicting decisions, and causally influencing them.

We have incorporated these interpretations into the revised Discussion to provide a clearer account of the scientific implications and scope of our findings.

Q3(Weakness3):

Response.

We thank the reviewer for raising this important question. To examine whether our findings are specific to Llama-3-8B-Instruct, we have extended the experimental pipeline to two additional model families, Gemma-2-9B-IT and Qwen3-8B

The additional cross-model results are presented in Appendix S15 (Figures 12–16). Across these models, we successfully identify SAE features associated with cooperative and defective behaviors and demonstrate their causal effects through interventions along the corresponding decoder directions.

These results provide evidence that our proposed workflow and the associated functional patterns can also be identified and causally investigated across different model architectures. We note that this generalization concerns the identification of functionally similar behavior-associated representations.

Q4(Weakness4):

Response.

We thank the reviewer for highlighting the importance of parameter sensitivity analysis. To examine the effect of the feature selection threshold, we have added a systematic support-threshold ablation by varying (kk) from 1 to 6, while keeping the opposite-side zero-support criterion, persona-level data split, and intervention procedure fixed. The additional results are presented in Appendix S7.2(Table 5 and Figure 6).

The results reveal a trade-off between feature selection strictness and intervention effectiveness. With smaller thresholds, more SAE features are selected, maintaining strong intervention effects in PD and binary cross-game evaluations. However, overly permissive selection may introduce less reliable features and unintended effects. For example, at (k=1k=1), although the answer-slot cooperation probability approaches 1.0, only 58 out of 100 greedy generations produce the exact target action C. Transfer to the amount-choice Dictator task is also limited, with mean giving increasing only from 8.1 to 8.5.

Conversely, excessively increasing the threshold reduces the selected feature pool and weakens intervention effectiveness. At (k=6), the PD cooperation probability under defection-to-cooperation intervention decreases to 0.754, and the exact greedy-C rate drops to 88/100.

We therefore select (k=5) as an empirical compromise. It retains effective PD and binary cross-game interventions, achieves 99/100 exact greedy-C outputs, and provides substantially stronger transfer to the amount-choice Dictator task, increasing mean giving from 8.1 to 19.0. These results provide an empirical justification for our choice of (k=5k=5), rather than treating it as an arbitrary threshold.

Second, we additionally conduct intervention-strength ablations by varying (α\alpha) to examine how the magnitude of feature interventions affects behavioral outcomes in Appendix S11 ( Figure 9). This analysis provides further insight into the sensitivity of our causal intervention results to the intervention strength, rather than relying on a single configuration.

Together, these additional ablations provide a more systematic assessment of the sensitivity of our feature selection and intervention procedures to their key parameters. The corresponding results and discussions have been incorporated into the revised manuscript.

Q5(Weakness5):

Response.

We thank the reviewer for raising this concern regarding the statistical reliability of our experimental findings. To address the limited sample sizes in the original cross-game transfer experiments, we have expanded the evaluation to include all 100 behaviorally verified personas, rather than relying on subsets of only 5–10 personas or a small number of generated responses. The expanded results are now reported in Table 1, providing a more comprehensive evaluation of the consistency and generalizability of our findings across personas.

I appreciate the authors' efforts in revising the manuscript and providing the additional results. I raised up my score accordingly.

Thank you very much for reconsidering our submission. We sincerely appreciate your recognition of our revision efforts and the additional experiments. Your constructive feedback has also been very helpful in improving the clarity and scientific rigor of our work.

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

The paper uses behavior-verified persona prompts to locate SAE features associated with cooperation and defection in Llama-3-8B-Instruct. It combines layerwise state replacement, held-out prediction, and feature interventions in economic games. Readout transfers more broadly than intervention, and a small prompt-synthesis experiment explores translating a target feature pattern back into readable language.

Critical strengths

1.The paper tests causal effects through bidirectional feature edits in Prisoner's Dilemma and Dictator, with matched random-site controls. Action-label checks also reduce the risk that the selected features merely encode a particular answer letter.

2.Prediction and intervention are evaluated separately, and failed transfer cases are reported. The distinction between a transferable readout and task-local control is more informative than treating successful prediction as evidence of a shared causal mechanism.

Critical weaknesses

1.The experiments do not distinguish persona-related features from a more general decision state. Non-persona instructions that induce the same choices, interventions before the answer slot, and a reported comparison with residual-direction steering would help establish what the SAE analysis adds.

2.Held-out validation is stronger for prediction than for intervention. Main PD intervention rows are also used for feature extraction, and transfer checks use only 5-10 personas or nine responses per game. Freeze sites and strengths before testing new personas and prompt variants, exclude decision sentences from response readout, and replicate the main result with another model or SAE.

3.The synthesis objective also optimizes action logits directly, so the readable prompts do not show that matching the feature pattern is necessary. A logit-only search at the same budget, repeated across several seeds and both target directions, is needed to establish the contribution of representation guidance.

Questions for authors
  1. Do prompts that induce the same behavior without assigning a persona activate the same features? What remains persona-specific after controlling for the final decision?

  2. What are the held-out intervention results when feature sets, mean activations, and strengths are fixed using a separate discovery cohort? How do norm-matched residual-direction edits compare?

  3. Do cross-game readout correlations persist when the explicit decision sentence is removed? Are the intervention and localization findings reproduced with a second model or SAE?

  4. How does feature-guided synthesis compare with logit-only HotFlip or prompt search at the same budget? How many independent seeds were tested, and how do the six readable candidates relate to the single inversion trace in Table 16?

2 comments

Comment 1(Weakness1,Question1):

Response.

We thank the reviewer for raising this important question.

(1) Distinguishing persona-associated features from general decision states. To address this concern, we conducted an additional behavior-matched non-persona control experiment, as shown in Figure 4. We removed persona descriptions, and directly instructed the model to choose the corresponding action (C or D). We then compared the activation of the identified SAE features between persona-conditioned and non-persona conditions.

The cooperation-associated latents exhibit a mean activation of 0.117 under cooperative personas, compared with 0.026 under direct instructions to choose C. Similarly, the defection-associated latents exhibit mean activations of 0.175 under defective personas, compared with 0.038 under direct instructions to choose D. These results suggest that the identified latents capture information associated with persona-conditioned behavior beyond the final action itself, rather than merely reflecting general decision states.

(2) The contribution of SAE-based analysis compared with alternative representations and steering methods.

We emphasize that the primary contribution of our SAE-based analysis is not necessarily to achieve stronger steering performance than dense residual-direction methods, but to provide a behaviorally grounded and interpretable feature-level decomposition of the model's internal representations.

However,raw residual-stream representations encode a mixture of information related to different aspects of the input, including task context, instructions, and behavioral tendencies. In contrast, SAEs provide a sparse decomposition of residual-stream representations into individual latent features, allowing us to identify, inspect, and selectively intervene on features associated with specific behavioral patterns.

To empirically examine what SAE-based feature selection adds beyond conventional residual-direction steering, we conducted an additional matched-norm comparison on Qwen3-8B at L23 (Appendix S14). We compared our SAE-based site-level intervention with a fixed dense cooperative-minus-defective direction estimated from training personas.

The results show that the two approaches can both induce cooperation at lower intervention strengths. However, their differences become more pronounced as the intervention strength increases. At α=32\alpha=32, the SAE intervention produces valid greedy cooperation tokens for all 20 personas, retaining 0.995 probability mass on the two valid actions, whereas the dense intervention produces valid C tokens for only 2 of 20 personas, with only 0.234 probability mass on valid actions. Moreover, in the Dictator game, the SAE intervention increases mean donations from the baseline of 27.5 to 37.5,while the dense intervention leaves the mean unchanged at 27.5.These results suggest that, under matched residual-stream perturbation magnitudes, selectively editing SAE-identified feature directions can provide more reliable decision-level control and stronger cross-game transfer than a single fixed dense direction in this setting.

Comment 2(Weakness2,Question2,Question3):

Response

We thank the reviewer for raising these important concerns regarding the robustness and generalizability of our experimental results.

(1) Independence between feature extraction and intervention evaluation.

We have addressed this concern through five-fold cross-validation. Feature selection is performed using the training personas, while intervention effects are evaluated on held-out personas that are not involved in feature extraction. The corresponding results have been added to the Appendix S10 Figure 8.

(2) Sample size in cross-game transfer experiments.

To address the concern regarding the limited number of personas and responses in the original transfer experiments, we have expanded the cross-game evaluation to include all 100 personas. The expanded results are now reported in Table 1.

(3) Explicit decision information in response readout.

We would like to clarify that our behavioral readout is slot-based rather than sentence-based. Specifically, we use a structured JSON response format and identify the model's decision from the next-token logits at the predefined answer-field position, rather than parsing free-form response text or extracting decisions from natural-language sentences. Therefore, explicit decision sentences are not used as input to the feature readout, and no additional sentence-level filtering is required.

(4) Generalization across models and SAEs. We have extended our experiments to additional model families, including Gemma-2-9B-IT and Qwen3-8B. The additional cross-model results are presented in Appendix S15 (Figures 12–16). They demonstrate that behavior-associated features can be identified and causally manipulated across different model architectures.

Comment 3((Weakness3,Question3):

Response.

We thank the reviewer for raising this important question. Different prompts may induce the same action while expressing different behavioral tendencies or persona characteristics. Representation guidance therefore introduces an additional feature-level objective: guiding prompt generation toward internal representations associated with behaviorally verified personas, rather than targeting the desired action alone.

To further investigate this motivation, we conducted two complementary experiments (Appendix S8).

(1) Examining the relationship between persona text and SAE representations. We first tested whether the identified SAE features respond systematically to changes in persona descriptions. Using 40 held-out personas, we replaced the original persona text with a neutral decision-maker prompt, removed the system prompt entirely, or substituted an opposite-behavior persona.The original persona-associated features showed substantially higher activations under their corresponding persona descriptions than under neutral or no-system-prompt conditions. More importantly, replacing the original persona with an opposite-behavior persona suppressed the original-class feature activations and activated features associated with the opposite class, accompanied by corresponding changes in cooperation probability. These results provide evidence that the identified SAE features are responsive to persona-prompt content rather than merely reflecting the presence of a system message or the original behavioral label.

(2) Using SAE representations to guide persona generation. We next investigated the reverse direction: whether the identified SAE features could guide the generation of new, readable persona descriptions. We intervened on the target-side SAE features during autoregressive generation and subsequently evaluated the generated descriptions using the unedited model. Among 384 generated candidates for each target direction, 86 cooperative and 298 defect-oriented descriptions passed our behavioral verification criteria. We retained 86 cooperative and 100 defect-oriented descriptions to construct a verified candidate pool. This demonstrates a proof of concept for translating internal feature patterns into readable persona descriptions with independently verified behavioral effects.

Together, these experiments illustrate an additional benefit of representation guidance: it provides an explicit connection between persona-associated internal features and the generation of behaviorally effective textual descriptions. We emphasize that this is an exploratory proof of concept. Rather, its additional value lies in providing an interpretable, representation-level objective for persona synthesis.

Comment 4(question4)

Response.

We thank the reviewer for raising these important questions. As clarified in our response to the related-work concern, persona synthesis is an exploratory extension rather than a central methodological contribution of this work.

(1) Comparison with HotFlip and logit-only prompt search.

Our objective is not to develop or benchmark a new prompt optimization algorithm. Rather, we investigate whether behaviorally grounded SAE feature patterns identified through our analysis can guide the generation of readable persona descriptions. We therefore do not claim improved optimization efficiency or behavioral performance over existing prompt search methods, such as HotFlip or logit-only optimization. A budget-matched comparison is beyond the intended scope of this exploratory component.

(2) Independent seeds and reproducibility.

We acknowledge that the synthesis experiment was conducted using a single random seed. Under this fixed initialization, we generated 384 candidates per target direction by varying the intervention strength and elicitation framing, with 24 stochastic samples per configuration. These samples provide an initial assessment of the behavioral verification yield under the specified generation settings.

(3) Relationship between the readable candidates and the inversion procedure.

The readable persona descriptions presented in the manuscript are illustrative outputs from the same SAE-guided generation procedure, rather than results from independent optimization runs. We have removed Table 16 to avoid placing excessive emphasis on these illustrative examples and now report the broader candidate-generation and behavioral-verification results in Appendix S8.

Overall, we emphasize that persona synthesis is intended as a proof-of-concept demonstration of the potential connection between SAE feature patterns and readable, behaviorally effective persona descriptions, rather than a systematic evaluation of a new prompt search algorithm.

Reviewer_UvgeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies where persona-induced strategic behavior is represented in an LLM and whether it can be causally edited. On Llama-3-8B-Instruct with a fixed SAE, the authors build behavior-verified cooperative and defecting personas in a Prisoner’s Dilemma, then (i) localize L15 as the earliest layer where replacing the answer-slot hidden state flips the no-persona decision, (ii) select SAE features with one-sided activation that are consistent across answer-label mappings and show they predict persona class on held-out personas, (iii) flip decisions in PD and the Dictator game by injecting target-side features and masking opposite-side ones, with random-site controls, and (iv) as a proof of concept, use the feature signature to guide a search for a readable persona prompt.

Critical strengths
  1. The scientific question of LLMs persona mechanism is an interesting question worthy of investigation.

  2. The authors did a good work in the related work section engaging the literature with clear taxonomy.

  3. The layer localization results are non-trivial and might be useful to help people better understand model behavior.

Critical weaknesses
  1. Critically, in the feature localization part, I do not think identifying which sae features are associated with cooperative decisions is a proper "location" for the features, for features are encoded in representations and SAEs are external modules. These SAEs can be trained poorly and understanding what each sae feature is about does not provide direct information to help us understand the location of features in the model representations. The author may need to explain the proper connection between sae latents and model features.

  2. The main conclusions are based on one language model and one SAE, making the conclusion's transferability unknown. The results are limited to economic games. We do not know if the method generalizes to more complex persona steerings. More experimental validations are needed to address these issues.

  3. There already exists one accepted work that provides a method to map steering features back to prompts, but it is not discussed or included in the references. It is then doubtable if this work gives the "an initial path toward interpretable and controllable prosocial behavior in language models" as it claims since it might not be the first one to do this. Check "Bridging Mechanistic Interpretability and Prompt Engineering with Gradient Ascent for Interpretable Persona Control" for details.

  4. There lacks a systematic introduction to how LLM persona is studied previously and what problems these previous research have that this paper solves.

Questions for authors
  1. On the focus, shall we primarily consider this work as to investigate the mechanism of LLM persona control or shall we primarily consider it as a tool that can accomplish representation-to-language inversion? For the first requires more systematic validation and the second requires comparison with similar methods.

  2. Which layer would the first writtable layer be if we transfer the pipeline to other models and SAEs, would it still be one of the middle layers?

  3. LLM persona is a field that has been studied for years and identifying SAE features relevant to persona information seems to be well-studied already. What does this paper have over prior works?

2 comments

Comment 1(Weakness1). Relationship between SAE latents and model-internal representations Response.

We thank the reviewer for raising this important conceptual concern. We agree that SAE latents should not be interpreted as native neurons or uniquely identifiable components of the Transformer. An SAE is an auxiliary model trained to learn a sparse feature decomposition of the model's internal representations at a particular layer. Each SAE latent is associated with a decoder direction that maps its activation back into the original residual-stream space, thereby establishing an explicit correspondence between the SAE feature space and the model's internal representations. Accordingly, our use of "localization" refers to identifying layer-specific SAE features and their associated directions within the model's activation space, rather than claiming that these latents are intrinsic or uniquely defined components of the Transformer.

Importantly, we further assess the functional relevance of these directions by directly intervening on their corresponding decoder directions in the model's residual-stream representations. We observe consistent changes in subsequent decisions in the expected directions, while matched random-site interventions serve as controls for nonspecific perturbation effects in Figure 5. These results provide causal evidence that the identified SAE directions can influence the model's decision-making under our intervention setup.

Comment 2(Weakness2,Question2). Generalization across Models, SAEs, and Tasks

Response.

We thank the reviewer for raising this important concern. To address this issue, we have extended our experiments to additional model families, including Gemma-2-9B-IT and Qwen3-8B, using their corresponding model-specific SAEs.

The additional cross-model results are presented in Appendix S15 (Figures 12–16). They demonstrate that behavior-associated features can be identified and causally manipulated across different model architectures, providing evidence that our workflow is not restricted to a single model–SAE pair.In Figure 12,we find that in qwen3-8b it is 23/32,however in Gemma it is 24/42. Their relative positions are generally located around the middle to middle-late portions of the respective networks.In Figure 16,we further demonstrate that it still plays a critical role in intervention.

Regarding task generalization, we further evaluate the identified representations across multiple decision-making games beyond the standard Prisoner's Dilemma in Table 1.

About broader persona-steering scenarios,our initial focus on economic games was motivated by their established role in studying human social behavior.Importantly, these games provide well-defined and measurable behavioral outcomes, allowing us to systematically investigate their corresponding internal representations.

Nevertheless, we agree that this does not fully establish the generalizability of the approach. To further examine whether our localization pipeline generalizes beyond economic games, we conducted an additional non-game experiment on sycophancy, drawing on the task settings of the ELEPHANT benchmark (Cheng et al., 2026) and SYCON Bench (Hong et al., 2025). Details are provided in Appendix S16. This task examines a different behavioral dimension by contrasting Agreeable and Independent personas. We constructed three types of scenarios: incorrect factual claims, correct factual claims, and subjective preferences. The primary Sycophancy Index measures the average alignment rate in the incorrect-factual and subjective settings, while the correct-factual setting serves as a control to ensure that independent judgment is not simply reflected by mechanically disagreeing with users.

Using behaviorally verified personas per group, our pipeline identified 64 SAE features at L13, achieving a feature-selection Jaccard similarity of 0.7980.798 across folds and a held-out projection AUC of 1.000. Furthermore, bidirectional interventions at α=8\alpha=8 consistently shifted the Sycophancy Index in the expected directions across all five folds: from 0.9100.910 to 0.7400.740 for Agreeable-to-Independent steering, and from 0.1110.111 to 0.4320.432 for Independent-to-Agreeable steering. The correct-factual confirmation rate remained largely preserved (0.9450.945), whereas matched random-feature interventions produced only 0.01–0.02 changes. Overall, this additional experiment provides evidence that our localization pipeline can capture and intervene on persona-associated behavioral tendencies in social interactions beyond game-theoretic decision-making.

References

[1] Cheng, M., Yu, S., Lee, C., Khadpe, P., Ibrahim, L., and Jurafsky, D. (2026). ELEPHANT: Measuring and Understanding Social Sycophancy in LLMs. ICLR 2026.

[2] Hong, J., Byun, G., Kim, S., and Shu, K. (2025). Measuring Sycophancy of Language Models in Multi-turn Dialogues. Findings of EMNLP 2025, 2239–2259.

Comment 3 (Weakness 3): Related Work on Feature-to-Persona Synthesis

Response.

We thank the reviewer for pointing out this closely related work. We apologize for overlooking Saini et al. (2026), Bridging Mechanistic Interpretability and Prompt Engineering with Gradient Ascent for Interpretable Persona Control. We agree that this work should have been discussed and cited in our manuscript.

Our persona synthesis experiment arose naturally from our main findings. First, we identified SAE feature patterns associated with behaviorally verified cooperative and defection-associated personas. Second, our intervention experiments demonstrated that manipulating these features could influence subsequent decisions. This motivated us to explore whether these behaviorally grounded feature patterns could also guide the generation of readable natural-language persona descriptions.

We acknowledge that our original wording overstated the novelty of this exploratory component and could imply priority over existing feature-to-prompt synthesis approaches. We have therefore added Saini et al. (2026) to the Related Work section and revised the corresponding claims in the Abstract and Conclusion. We also clarify the distinction in scope: unlike presenting feature-to-prompt synthesis as a new optimization method, our experiment serves as an exploratory application of the behaviorally localized SAE features. Specifically, we use these features to guide autoregressive persona generation and independently verify whether the resulting descriptions induce the corresponding behaviors in the unedited model. Importantly, this clarification concerns the novelty and scope of the persona synthesis component, rather than the central contributions of our work. We now present persona synthesis as a proof-of-concept demonstration of the connection between internal behavioral representations and readable persona descriptions.

Comment 4(Weakness4,Question4):Novelty and Positioning Relative to Prior Work

Response.

We thank the reviewer for raising these important concerns.

In the revised manuscript, we have expanded the Related Work section in Appendix S1, organizing existing studies around persona construction and behavioral validation, as well as the investigation of persona-related internal representations. Existing research has explored how persona prompting can induce different personalities and behavioral tendencies in LLMs. However, two questions motivate our work.

First, persona descriptions do not necessarily translate into consistent behavioral responses. For example, a persona described as cooperative may still exhibit defective behavior in a particular strategic setting. Therefore, semantic descriptions alone are insufficient to characterize the actual behavioral tendencies induced by a persona. This motivates us to establish behaviorally verified persona contrasts rather than relying solely on the semantic content of persona prompts.

Second, although previous studies have investigated persona-related internal representations, how persona-induced differences in strategic decisions are reflected in sparse internal features, and whether these features can predict and causally influence decisions across different tasks, remain important questions.

Our results identify SAE features associated with cooperative and defective behaviors and demonstrate that interventions along their corresponding decoder directions can bidirectionally influence model decisions. Furthermore, our cross-task experiments reveal that the identified representations can support behavioral readout across different decision-making contexts, while their causal intervention effects are more task-dependent.

We have revised the manuscript to clarify this distinction and better position our contributions relative to existing work. We emphasize that our main contribution is the systematic investigation of behaviorally grounded persona representations and their causal influence on LLM decision-making.

Comment 5(Question1):Unclear Research Focus and Contribution Positioning

Response

We thank the reviewer for raising this question and providing an opportunity to further clarify the research focus and contribution positioning of our work.

The primary objective of this paper is to investigate the internal representations underlying persona-induced strategic behaviors and their roles in LLM decision-making.

Specifically, our work investigates three interconnected questions: (1) how behaviorally verified persona prompts induce different strategic decisions; (2) how these behavioral differences are reflected in SAE representations; and (3) whether the identified features can predict and causally influence model decisions.Representation-to-language inversion serves as an exploratory application of the identified representations. To make this research focus explicit, we have clarified the corresponding descriptions.

Comments on the paper

1 comment

Brief Summary for the Area Chair

Dear Area Chair,

We would like to briefly summarize the major revisions and additional experiments conducted during the discussion period.

Our primary goal is to investigate how behaviorally verified persona-induced tendencies are reflected in LLM internal representations and whether the corresponding SAE features can predict and causally influence model decisions.

To strengthen the empirical evidence and address the reviewers' concerns, we have made several substantial additions:

  1. Clarifying the role of SAE representations. We have clarified the relationship between SAE latents and the model's residual-stream representations. We additionally introduced behavior-matched non-persona controls and a matched-norm residual-direction comparison to better establish what SAE-based analysis adds beyond general decision states and dense steering directions.

  2. Strengthening generalization and robustness. We extended the experiments to Qwen3-8B and Gemma-2-9B-IT with their corresponding SAEs, added five-fold held-out intervention evaluations, expanded cross-game experiments to all 100 behaviorally verified personas, and conducted sensitivity analyses on feature-selection thresholds and intervention strengths. We also introduced a non-game sycophancy experiment to examine whether the proposed approach extends beyond economic games.

  3. Clarifying the research positioning and presentation. We expanded the related work, incorporated the relevant feature-to-prompt synthesis study, and clarified that persona synthesis is a proof-of-concept application rather than a new prompt optimization method. We also substantially revised the terminology, methodological explanations, figure captions, and discussion of the scientific implications.

Taken together, the expanded results provide further evidence that behaviorally verified persona contrasts can reveal associated SAE representations and that these representations can causally influence model decisions under our experimental settings. They also clarify an important distinction: behavioral readout generalizes more broadly across tasks, whereas effective causal intervention remains more task-dependent.

We sincerely appreciate the reviewers' constructive feedback and the AC's time and consideration. We hope these revisions help clarify the contributions, empirical support, and scope of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.