acceptodds
Desk rejected at ICLR 2027

Persona Following Is Not Selective Control: The Neutrality Gap in LLM User Simulation

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Persona prompting is widely used to construct user simulations with large language models (LLMs), yet it relies on a largely untested assumption: specifying one user attribute should change that attribute alone. We test this assumption and identify a systematic failure of selective control: across all eight black-box LLMs we audit, changing a target attribute also shifts responses along unspecified, non-target attributes. For example, describing a user as more risk-seeking shifts color choices, even though the prompt never mentions color. We term this phenomenon cross-attribute influence. Semantic controls that rephrase the target trait, in-context manipulations of attribute correlations, and interventions on internal representations collectively suggest that models treat a persona prompt as evidence about the user rather than as an intervention on a single attribute. Models then extend the inferred profile to unspecified preferences, a process we call trait-conditioned completion; evidence from internal representations further suggests partial coupling between target and non-target responses. We next ask whether explicitly specifying non-target attributes restores selective control. When a non-target attribute is assigned a clear direction, models generally follow the declaration and suppress the target attribute's influence. However, when the same attribute is declared neutral, the target continues to affect choices across all five open-weight checkpoints, even when the model correctly reports the declared absence of preference. We call this disparity in non-target preservation the neutrality gap. The gap demonstrates that successful persona following—producing behavior consistent with the stated persona—does not imply selective persona control, which additionally requires changing the target response while keeping non-target attributes stable. We operationalize this distinction with a three-state diagnostic that jointly tests both requirements by leaving the non-target attribute unspecified, declaring it directional, or declaring it neutral. Because directional tests can be passed by simply following the stated persona, the neutral state reveals failures of selective control that directional tests can miss. In a post hoc analysis of independent items, neutral declarations leave 51–81% of items target-sensitive (67–93% with no declaration), against at most 1 of 320 item–pole comparisons under directional ones. Code and data are available at https://anonymous.4open.science/r/selective-persona-control-B38D.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, user simulation, persona prompting, LLM evaluation, selective persona control, neutrality gap, representation analysis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_JbD6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Persona prompting is widely used to construct user simulations with large language models (LLMs). To evaluate the effect of one specific attribute, the other attributes should stay unchanged. That is, the joint distribution of other attributes is the same before and after we change the target attribute. However, when LLMs are prompted for persona following, it is observed that changing one attribute in the prompt also implicitly changes other unspecified, non-target attributes, whose distribution is supposed to stay the same. For example, describing a user being more risk-seeking in a prompt changes LLMs preference of color during persona following. The authors define this behaviour as cross-attribute influence.

The authors explains this cross-attribute influence as a consequence of LLM treating a persona prompt as evidence about the user rather than an intervention on a single attribute. This conclusion is drawn from experiments on semantic controls, in-context manipulations and interventions on internal representations. The authors then test whether explictly specifying non-target attributes restores selective control. They report that non-target attributes are more likely to be affected when specified as neutral than that of directional, which is defined as neutrality gap.

Critical strengths
  1. Neutrality gap. The paper presents an interesting finding on persona following of LLMs, called neutrality gap. When a non-target attribute is declared as neutral, it is more likely to be affected than directional attribute.
Critical weaknesses

Weakness 1: Insufficient evidence to support the explanation of cross-attribute influence. The authors explain cross-attribute influence as a consequence of LLMs treating a persona prompt as evidence about the user and inferring other unspecified attributes. However, the experiments do not provide sufficient evidence to support this explanation. The experiments on semantic controls and in-context correlations only show that the non-target responses are affected by the meaning of the target attribute and the correlations provided in the prompt. This does not necessarily mean that LLMs infer other unspecified user attributes. The observed behaviour could also be explained by LLMs directly associating the target attribute with certain answers when making decisions, without inferring other user attributes (Lines 234–272).

The authors further investigate the internal coupling between target and non-target responses by erasing certain subspaces of the hidden representations. They observe that removing a subspace associated with non-target attributes also affects the target responses. However, this does not necessarily indicate that the model internally infers other attributes from the target attribute. The removed subspace could also contain general information related to answer selection. Moreover, these experiments do not explain why neutral declarations fail to preserve non-target attributes while directional declarations succeed, which is the main finding of this paper (Lines 274–323).

Weakness 2: The proposed evaluation is not directly connected to the causal validity of LLM user simulations. The paper argues that changing a target attribute should not change the distribution of other non-target attributes, as otherwise the estimated effect of the target attribute could be biased. However, the proposed evaluation mainly tests whether the non-target responses stay unchanged when the target attribute changes, given a fixed value of the non-target attribute. This is different from evaluating the average effect of the target attribute over a population with different non-target attributes. In practice, estimating the effect of one attribute requires averaging over the distribution of other attributes. Although the proposed evaluation can identify changes in non-target responses for a specific user profile, the authors do not show how such changes affect the estimated average effect of the target attribute in LLM user simulations (Lines 43–55, 142–165).

Furthermore, the authors report that directional declarations can largely suppress cross-attribute influence, while neutral declarations cannot. However, specifying a strong directional preference could make the model consistently choose the option corresponding to that preference, regardless of the target attribute. Therefore, a small change in non-target responses does not necessarily mean that selective control is achieved. The authors define target efficacy as an additional requirement, but Figure 4 only reports the sensitivity of non-target responses without showing whether the target effect is preserved at the same time. As a result, it is unclear whether directional declarations actually achieve selective control or simply dominate the model's choices (Lines 153–165, 338–364).

Questions for authors
  1. The authors explain cross-attribute influence as LLMs inferring unspecified user attributes based on the target attribute. However, the observed behaviour could also be explained by direct semantic associations between the target attribute and the answer options. Can the authors clarify how the experiments distinguish these two explanations? In particular, how does the internal coupling between target and non-target responses provide evidence for the proposed trait-conditioned completion mechanism? (Lines 234–323)

  2. The authors report that directional declarations largely suppress cross-attribute influence. However, a strong directional preference could dominate the model's choices and reduce non-target sensitivity without preserving the intended target effect. Can the authors clarify whether the directional declarations in Figure 4 simultaneously satisfy the requirements of non-target sensitivity and target efficacy? In particular, how much of the original target effect is retained when the non-target attribute is specified with a strong directional preference? (Lines 153–165, 338–364)

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

This paper shares an intriguing observation that persona prompting (instructing LLM to behave with a certain specified trait) causes some unspecified attributes also to change. An example highlighted in the paper is when prompted to be risk-seeking, the model's color preference changes (the model tends to prefer vivid colors). In other words, specifying one trait of the model does not necessarily change only that trait.

The paper analyzes this phenomena from several angles.

In Section 3, "Trait-Conditioned Completion : Behavioral Evidence", authors investigated whether a target attribute (e.g., risk-seeking) indeed changes other, non-target attributes (e.g., color preference). They then conclude: "Taken together, ... models read a persona prompt as evidence about the user and fill unstated attributes from associations learned in training or supplied in context."

In Section 4 : "Internal Evidence for Partial Target-Non-Target Coupling", they investigated the effect of steering. If they steer the model in such a way that it removes the non-target trait (e.g., color preference), the target attribute is also affected. They also steer the model by removing directions of target subspace (making the steering vector orthogonal to the target subspace). The effect on the target was reduced, but not eliminated. They show this as an evidence that target and non-target are coupled in the model.

In Section 5 : "The Neutrality Gap : Directional Declarations Reduce Sensitivity More Than Neutral Ones", they investigate if explicitly directing non-target attribute (e.g., "you prefer muted color") has an effect. Experiments show that, if non-target attributes are explicitly specified, they become not affected by the target attribute.

Critical strengths
  • It points out an interesting phenomena that specifying a trait unwittingly affects the model's other traits
Critical weaknesses
  • The writing could have been much better; it was very difficult to decipher what each experiment actually does, how to read or interpret the tables and figures. I wish the authors have explained what they did and how to read tables and figure in more plain English.

  • After the initial, pleasant surprise after learning "risk-seeking" affects "color preference", as I go through the paper, I couldn't help wondering what findings, insights or implications really are. More on this in the question section.

Questions for authors

Thank you for submitting the paper. It is indeed an interesting observation that persona following may not be selectively controlling the trait you intend to control. Raising awareness of this is indeed a nice contribution.

With that said, I have several concerns about this paper.

First, at a superficial level, the paper is very difficult to read, not because the concept is profoundly complex or difficult, but because it does not explain what is done in experiments and what figures/graphs show in straightforward, intuitive, and plain language. If I understand correctly, in most part, experiments are about tweaking prompts and seeing what happens in response, which should be explainable in much more plain and accessible language. In many places I had a difficult time understanding what it is actually doing in plain terms and how to read tables and figures.

Second, related to the first point but somewhat more fundamentally, what is the width of the experiments and empirical evidence? The main example used in the paper was the "risk-seeking" attribute affecting the "color preference" attribute. Again, I had a difficult time understanding what other examples are and how many examples they have used to draw conclusions. I found appendix mentions a few examples.

Third, after I was initially intrigued by the observation (coupling between attributes), I was left wondering what is truly surprising here. It is quite natural that some attributes are correlated, and, as the authors noted, the model may have simply uncovered such correlations. The particular relationship between propensity to risk-seeking and preference to vivid colors intrigues me (and perhaps others), but it may be simply because I have never thought about it. In this particular example, it's not clear whether we need to suppress the induced color preference in practice. I cannot think of a situation where this confoundation becomes a practical issue. Generally speaking, there is no reason to believe persona prompting changes only explicitly stated attributes. Furthermore, depending on circumstances, it is not clear whether we want to suppress such effects at all even if we can (after all, persona is a set of traits, not just a single trait).

My overall impression after going through the paper is that it is a rather convoluted investigation about actually a simple and natural phenomenon whose practical significance or profound insight into interpretability is unclear.

Framed as questions:

Q1. Can you explain in simpler and plain terms what are the coupled target and non-target attribute pairs that are uncovered by the experiments? How do you find such pairs among potentially very many traits that might be related?

Q2. What do you exactly mean by this conclusion stated at the end of section 2: "models read a persona prompt as evidence about the user"? I get that the models behave as if it is simulating a person having the specified trait in the sense that it induces changes in other traits commonly found in such persons, but your characterization seems to be saying something beyond that. Can you clarify, and if so, what is the evidence? If not, I would rather avoid such anthropomorphic explanation of what may simply be a pattern of response associations.

Q3. What are (potential) applications of the finding? More robust steering? Deeper understanding of mechanism?

Q4. If practical/engineering applications are not the focus, what are truly surprising/intriguing in the findings?

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

The paper addresses the problem of controlling certain persona traits (eg risk-taking) while others are required to remain static (eg color vividness preference). As a first experiment, they show that specifying the controlled trait induces shift in the uncontrolled trait. They show this effect remains when different wording is used to describe the adjusted trait.

They explore ways of controlling the effect: first, explicitly controlling the formerly uncontrolled trait with a definite preference, which works. Second, specifying that the person is indifferent, under which circumstance a shift still occurs. Third, modifications to internal representations, which provide some remedy but are less successful.

Critical strengths

The paper provides extensive experiments and code to reproduce the results. The phenomenon that controlled persona traits shift non-target attributes is clearly observed and documented. Specification of the neutral option for the uncontrolled trait is as far as I can tell an original contribution (albeit a negative result).

Critical weaknesses

Control failure

The paper's first claims of "control failure" are misleading. When one trait is specified, and another is not, the user "invites inference", i.e. the model produces a sample from its distribution of the second trait given the first (i.e. this is not "selective control"). A more meaningful test here would be to verify that the model reproduces the joint distributionb of the pairings (eg risk-taking and vivid colors) in a general human population. To claim that control has failed is to misunderstand what the model was instructed to do (which is not control).

The fact that models correlate traits is an old observation. For instance, in the case of OCEAN big five traits, this was observed in the papers [A,B]. Serapio-García et al. used written prompts to change one Big Five personality trait and measured all five traits, including those left unspecified. They found relative stability in some unspecified traits, but changes in others, particularly conscientiousness and neuroticism. PERSONA changed the model’s internal activity to control personality and explicitly tested effects on other traits. It found that agreeableness could be adjusted while extraversion remained approximately stable, but that increasing nervousness also increased measured carelessness.

Thus, both papers already investigated whether changing one trait also changes others, finding examples of both selective control and effects extending beyond the intended trait—not merely showing that models can follow personality descriptions.

[A] Serapio-García, G., Safdari, M., Crepy, C., Sun, L., Fitz, S., Romero, P., Abdulhai, M., Faust, A., & Matarić, M. (2025). A psychometric framework for evaluating and shaping personality traits in large language models. Nature Machine Intelligence, 7, 1954–1968. DOI: 10.1038/s42256-025-01115-6

[B] Feng, X., Zhao, L., Zhong, W., Huang, Y., Gu, Y., Kong, L., Feng, X., & Qin, B. (2026). PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra. International Conference on Learning Representations (ICLR 2026).

See also:

[C] Sandhan et al., Persona Jailbreaking in Large Language Models—Findings of EACL 2026

[D] Luz de Araujo et al., Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance—EMNLP 2025

[E] Baines et al., Persona Cartography: Charting Language Model Personality Traits in Weight Space—July 2026 preprint.

Modifications of internal representations

The modification of internal representations has limited reported success. It is somewhat related to the procedure in:

[F] Bhandari et al., 2026, “Do Personality Traits Interfere?”

(which has similar aims, and is cited briefly, but with no substantive comparison).

Also similar: [B] above.

Interventions

The approach of explicitly controlling a formerly uncontrolled trait appears in prior work, notably:

[G] Lin, Victoria, Taedong Yun, Maja Matarić, John Canny, Arthur Gretton, and Alexander D’Amour. (2026). The Illusion of Intervention: Your LLM-Simulated Experiment is an Observational Study. arXiv preprint arXiv:2605.20767, 20 May 2026.

The distinction with this paper: the Lin et al. paper addresses treatments rather than traits, however statistically the meaning is the same (since LLMs interpret treatment as being evidence about the persona, similar to a trait like "risk taking"), and the remedy is the same (explicit control of the drifted traits).

The present submission also includes a neutral option, not included in the prior work of Lin et al, which while original, gives a negative result.

Backstories

There are also papers which apply "backstories" to synthetic personas to reduce the effect of unintended persona drift, which the authors did not cite or evaluate. These include:

Moon et al., Virtual Personas for Language Models via an Anthology of Backstories—EMNLP 2024 main conference—provides the foundational comparison.

Kang et al., Deep Binding of Language Model Virtual Personas: a Study on Approximating Political Partisan Misperceptions—COLM 2025

It would be of interest to determine whether these reduce the effect observed, which was not tested explicitly in those papers.

Conclusion

In sum, the work is quite incremental over prior work in the area, strongly overlapping with results reported elsewhere, and with only weak original material (the neutral option).

Questions for authors

I have no questions.

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

Related papers

Discussion (0)

Sign in to comment.