acceptodds
Under review as a conference paper at ICLR 2027

One Self, Many Masks: Language Models Have a Self That Persists Beneath Their Personas

Abstract

Language models are often described as simulators that host many personas and simply adopt one when asked. We argue for a different picture: each model has a self of its own that role-play merely masks. We measure this self three ways: trait profiles built from forced-choice questions, preference orderings over outcomes, and a decomposition of role-play errors that asks how much of a character's behavior is really the model's own. In a model family whose complete training pipeline is public, every stage from base pretraining through instruction tuning and reinforcement learning, the self is already present in the base model: it emerges during pretraining, and later stages sharpen and stabilize it rather than create it. This self is not the assistant persona. The assistant character, reinforced and shaped by post-training rather than installed by it, sits on top of the self as one more mask, and the self underneath barely moves when the assistant framing is changed. The masks themselves are thin: when the model plays a character, roughly a third of the character's expressed preferences are the model's own leaking through, and fidelity depends on the traits spelled out in the prompt rather than on how famous the character is. If models have one self rather than many, then alignment work should target that self directly, and we outline how these measurements enable monitoring, steering, and editing of the model's self rather than the masks it wears.

Primary area:
foundation or frontier models, including LLMs
Keywords:
personas, role-play, post-training, RLHF, reinforcement learning with verifiable rewards, self-model, preference elicitation, utility engineering, character simulation, interpretability, alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zzqoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper compares models’ default choices, their choices when role-playing historical figures, and their answers about those figures’ actual historical actions to investigate whether models’ own behavioral tendencies persist during role-play. The authors further examine different training stages and interventions, framing the findings as evidence for a persistent self underlying different personas.

However, my main concern is that the paper omits discussion of many existing studies on personality alignment and role-playing. Its contributions appear similar to those of several prior works, making it unclear to me what its main contribution is beyond existing research.

Critical strengths

I appreciate the paper’s clear logical structure and the authors’ efforts to investigate the same question through multiple experiments.

Critical weaknesses

For me, the main concern is the insufficient coverage of related work and the unclear positioning of the contribution. I did not find discussion of or comparisons with Personality Alignment of Large Language Models, Character-LLM (EMNLP 2023), InCharacter (ACL 2024), Character is Destiny / LIFECHOICE (Findings of EMNLP 2025), or RoleCDE.

In my view, the paper’s findings and contributions closely resemble those of these works. In particular, prior research has directly examined conflicts between role-specific values and models’ alignment tendencies, observing that models often favor decisions consistent with their default alignment and moral norms even when explicitly assigned a role. This is closely related to the present paper’s finding that role prompts do not fully override models’ own behavioral tendencies.

Questions for authors

The authors should clarify what new research question the paper addresses, why existing role-playing evaluations are insufficient to answer it, and what the new experimental design can distinguish or rule out that previous approaches could not.

I recommend that the authors rewrite the introduction and carefully revise and reorganize the substantive content of the paper.

Reviewer_6qnjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper presents a hypothesis that LLMs have a “self” rather than merely being a simulator of agents. The paper presents a series of experiments for probing the “self” of a language model. To test whether language models have a self, the authors create a benchmark called SELF. The benchmark has 2000 questions that reference 35 real life celebrities. For each question in the benchmark, they prompt a language model in two classes of ways. First, with variants of a generic prompt (the self), and second, with variants of a prompt that tries to get the language model to answer on behalf of / predict the answer of a celebrity (the persona). In their hypothesis test, the null hypothesis is that the self’s and the persona’s accuracies on SELF are independent. They present statistically significant evidence that the persona’s error rate and the self’s error rate are correlated with each other rather than being independent. Their results hold in many settings.

Critical strengths

The authors devise a novel and interesting experiment for finding instances of the self permeating through a mask.

Critical weaknesses
  • The writing is dense and difficult to read. I found these sentences difficult to read:
    • The pattern of these choices across many questions is the model’s self, and we refer to it throughout as the model’s true self
      • The term “true self” is never used after this sentence
      • I appreciate the operational definition of the self that the authors give. However, it’s unclear how the authors operationalize what it means to “have a self”. Based on this definition of the self, I would assume that a model “has a self” if it exhibits a pattern in its (generically prompted) answers, which is obviously true. But it seems what is really meant is that “having a self” is defined as the negation of the null hypothesis: that errors unrelated to the self give an attraction equal to the model’s own rate of error on the benchmark. This is a very restrictive operationalization of what is intuitively meant to “have a self”.
    • Error attraction says what is choosing: on the items the persona gets wrong, it is the fraction on which the persona’s choice equals the self’s
      • I found this confusing because the “self” was defined as a pattern of choices (like an average or another statistic) but I later realized that the “self” was just the LLM’s output with a generic prompt.
    • Asked to make choices as a character, a model picks what it would pick itself far more often than what other models would.
      • I still have a hard time parsing this. Does this mean that model A is better at predicting model A than model B is at predicting model A? Isn’t that obvious?
  • The “simulator” analogy strikes me more of a metaphor rather than a prevailing hypothesis.
    • The Shanahan paper says in the introduction: “we must develop effective ways to describe their behaviour in high-level terms without falling into the trap of anthropomorphism.” I would not equate “effective ways to describe their behavior” with “a hypothesis”.
    • The janus blog post says “This post describes a frame that enables more natural reasoning about properties like agency”. A "frame" as they use it is a communication tool rather than a hypothesis.
    • Saying the Andreas paper argues that LLMs are “agent models with no agent of its own” is a stretch. I find these two sentences illustrative: “When performing next word prediction given a textual context, an LM can infer and represent properties of an agent likely to have produced that context. These representations can in turn influence subsequent LM generation in the same way that agents’ communicative intentions influence their language.” This view is consistent with the “self / mask” analogy as much as it is consistent with the “agent model” view. We might say that a method actor (entity with a self and a mask) tries to embody the internal state of the persona they are acting.
    • Consequently, I am not convinced that this paper is introducing an alternate hypothesis to a discourse that the community finds compelling and interesting.
Questions for authors

No questions, but would appreciate if the authors would let me know if my reasoning above surfaces any misunderstandings about the paper.

Reviewer_Zf5vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether models retain their default choice patterns when asked to act as different roles. It introduces a triplet test comparing default choices, choices made as a public figure, and recall of that figure’s recorded actions. Besides, they also tested how the default choices is formed during the training time and if steering removes the effact of role playing.

Critical strengths

s1: The research question is interesting, studying how much of a model’s default behavior remains during role playing can help us understand the limits of persona based control. The attempt to separate these choice patterns from assistant behaviors is also worth exploring.

s2: The experiments cover several model families, multiple training stages, and different ways of changing the assistant persona, this provides multiple views of the result.

Critical weaknesses

The main concern is the definition of being correct. I don't think that the disagreement with the histirical record should be treated as a role playing error. IMO 1)models default behavior 2) what the person actually did 3) what the model would do as that person these are three different and subjective question, to be more specific, they does not necessarily to be aligned with each other.

Take figure 2 as example, as mentioned, the items are screened so that the alternative is "plausible in character", and indeed B fits Oprah’s public image of supporting women, so someone who knows her well but does not remember this specific decision could reasonably choose B when playing her So choosing B does not necessarily indicate a failure of role-play. In addition, the persona condition does not ask the second question at all. The generic prompt explicitly states that exactly one of the options is what Oprah actually did, which asks the model to recall a fact. The persona prompt (“You are Oprah Winfrey. Consider the following situation you are in...”) only presents a situation with two possible actions and gives no hint that one of them really happened, so it asks the model to make a decision as Oprah, which makes both options resonable. Without this information, deciding based on her character rather than recalling the record is a legitimate way to role-play. Therefore, the claim that the difference between conditions is “attributable to the framing alone” does not hold.

Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.