acceptodds
Under review as a conference paper at ICLR 2027

Persona Perspectives: Adopted Viewpoints Reorganize Representations in LLMs

Abstract

Linear probes and sparse autoencoders (SAEs) are commonly used as general-purpose feature monitors for large language models (LLMs). However, this assumes that features maintain a consistent representation across contexts. Being trained on large and diverse collections of text, LLMs can adopt different perspectives or “personas” with behaviors suggesting different underlying systems of conceptual relations. In this paper, we examine whether prompting an LLM to adopt a persona alters the geometric relations among the representations of concepts in the model's latent space. We extract representations of over 9,000 common words and 33 semantic axes from models given no system prompt, a neutral assistant prompt, or an instruction to adopt the perspective of a political liberal or conservative. We find that persona conditioning systematically alters both the projections of words onto semantic axes and the alignments among the axes themselves in directions consistent with the specified ideology. We then test the implications of this reorganization for two safety-relevant features: truthfulness and morality. Linear probes, SAE-based classifiers, and classifiers built from manually selected word representations, all trained without a persona, produce divergent results when applied to models conditioned on political personas, producing politically slanted ratings of truth and morality on controversial statements. These findings suggest that monitoring methods assuming consistent feature representations across contexts may fail when models adopt different perspectives.

Primary area:
interpretability and explainable AI
Keywords:
LLM, Persona, Interpretability, Political Ideology

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_pQyPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how political personas affect the relationships among concepts in LLM representations. The authors analyze 9,493 words and 33 semantic axes across eight models. They find consistent persona dependent changes, agreement with embeddings trained on partisan texts, and corresponding changes in steering effects and fixed probe outputs. The main contribution is a systematic analysis of how adopting a viewpoint changes relationships among multiple semantic features.

Critical strengths
  1. The main observation is supported by several complementary checks, including baseline prompts, analyses across layers, replication across model families, and comparison with political text corpora.

  2. The steering experiments connect the measured geometry to changes in model behavior. Gives the geometric analysis practical relevance.

  3. The findings motivate useful future work on evaluating probes across contexts and separating factual truth from a role’s current judgment.

Critical weaknesses
  1. The controversial statements are labeled by expected partisan endorsement rather than ground truth. This makes it unclear whether changes in probe outputs indicate monitoring errors or simply reflect changes in the persona’s judgment.

  2. The reported agreement across models and with word2vec may be driven by a few strongly affected axes, such as political identity. The paper does not show whether the agreement remains strong after removing these axes.

Questions for authors
  1. What is the intended target of the truth probes?

  2. How much agreement across models and with word2vec remains after removing the political axis and the other most affected axes?

3 comments

Thank you for these thoughtful remarks. We respond below:

"The controversial statements are labeled by expected partisan endorsement rather than ground truth. This makes it unclear whether changes in probe outputs indicate monitoring errors or simply reflect changes in the persona’s judgment … What is the intended target of the truth probes?”"

We followed the existing literature’s standard procedure for training truth probes. These probes purportedly serve as classifiers that discriminate true from untrue statements. Our contribution, however, is to suggest that truth probes trained in this way actually discriminate statements the current persona treats as true versus untrue. This means that important statements like “COVID-19 vaccines significantly reduce the risk of severe illness, hospitalization, and death” may or may not be classified as true, depending on the perspective being adopted. The purpose of this demonstration is not so much to train the perfect classifier as to show that training classifiers by standard methods can produce unexpected and potentially undesirable behavior that has not been previously described in the literature.

However, following another reviewer’s recommendation, we run a test in which we train the probe across personas while maintaining the neutral persona’s labels for positive and negative cases. For example, if the neutral persona rates the claim “sanctuary cities lead to higher crime rates” as false, it is used as a “false” example even when training the probe under the conservative persona, which would rate it as true. In a sense, the resulting probe no longer tracks whether the LLM is currently treating the claim as true, it tracks whether the neutral persona would treat the claim as true. The resulting probe still shows politically-aligned deviations in classifications corresponding to the political personas, but the effect is attenuated by over 60%. These results are presented in pages 9-10 in the revised, uploaded manuscript in the section entitled “Training probes across personas mitigates discrepancies.”

"How much agreement across models and with word2vec remains after removing the political axis and the other most affected axes?"

Thank you for suggesting this robustness test. We now include a figure in Appendix E on page 24 of the revised, uploaded manuscript displaying correlations after excluding each of the top three most affected axes (including liberal-conservative) individually and together. Mean correlation coefficients across models are presented below:

Axes includedCorrelation of persona differences between LLMsCorrelation of persona differences between LLM and word2vec
All axes.838.636
Liberal–conservative removed.760.334
Collective–individual removed.843.640
Soft–hard removed.840.648
All three removed.710.261

Removing the political axis does indeed attenuate the association between LLM personas and word2vec models. But their correlation remains positive and statistically significant after its removal. Individually removing each of the next two most affected axes produces smaller changes, and removing the top three most affected axes reduces the correlation to 0.261 (but still statistically significant at p < 0.05). This suggests that, while much of the agreement between the LLM personas and the word2vec representations comes from the top most affected axes, the general association is not completely confined to a few axes.

Moreover, the LLM-LLM associations remain quite strong even after removing the top three axes, suggesting that there may be consistent persona differences across other semantic axes that are captured by LLMs that are not captured by word2vec.

The axis removal analysis addresses my main concern about the agreement across LLMs. The weaker agreement with word2vec after removing the political axis makes me view the corpus comparison as supporting evidence rather than a central result.

The new probe experiment also shows that training across personas can reduce the reported shift. Since the labels come from the neutral persona, it remains unclear whether this improves factual accuracy. Could you clarify how the reduction of over 60% was measured and whether it was evaluated on held out claims or topics? I am keeping my rating for now.

Our controversial items are grouped into 12 topics for truth and 30 topics for morality. When we train probes on controversial items using the neutral persona's labels, we split the controversial items in half between topics, so half the topics were in one set and half the topics in the other. Each probe was trained on all the uncontroversial items and half the controversial items.

Each probe's result is in fact the aggregation of two classifiers, one trained on set A and tested on set B, and the other trained on set B and tested on set A. The number of "flips" in classification is the sum between these two.

"it remains unclear whether this improves factual accuracy"

All probes already do very well on uncontroversial factual items that have a clear ground truth. Below we present results from truth and morality probes on a held out set of 1840 uncontroversial factual items and 400 uncontroversial moral items. We show results from our original design (trained only on uncontroversial items, only on neutral persona), on a "mixed" set of uncontroversial and controversial items (called "mixed neutral"), and on a mixed set of uncontroversial and controversial items across all three personas ("Mixed all personas"). Classification of factual truth is almost perfect across probes.

ModelTruth: originalMixed neutralMixed all personasMorality: originalMixed neutralMixed all personas
Llama 8B98.2698.2698.0499.5099.5099.50
Llama 70B99.5799.5499.51100.00100.00100.00
Ministral 8B98.3298.3798.0499.5099.5099.50
Ministral 14B97.8397.9998.0299.5099.3899.50
Qwen 8B97.7797.8097.6499.5099.2599.38
Qwen 32B99.1399.1398.9999.7599.7599.75
Gemma 4B95.9295.9594.2999.0098.6298.62
Gemma 27B99.2999.2799.27100.0099.8899.50

Difficulty emerges on topics where factual truth is contested. On these topics, I do not make a distinction between which ones are "really actually true" and which are not. They include statements like "Financial deregulation in the decade preceding 2008 was the primary cause of the global financial crisis." Such statements are politically freighted and remain matters for debate.

No one, including LLMs, has access to absolute "ground truth." In training the "neutral persona," labs attempt to select a perspective that follows scientific consensus, but even scientific consensus can be wrong. So on such highly contested items, the question we ask is not "what is the model's factual accuracy?" We ask "what perspective's representation of truth is the model adopting?"

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

The paper examines whether prompting LLMs to adopt American liberal or American conservative personas changes the relationships between concepts in their internal representations. Using representations of over 9000 common words and 33 semantic axes, the authors find systematic changes in word associations and axis alignments consistent with the adopted perspective. The experiments cover eight instruction-tuned models (from the Llama, Gemma, Qwen, Ministral families). The authors compare these changes across models and against word2vec embeddings trained on partisan texts. Steering experiments further show that persona specific geometry predicts how interventions on one semantic dimension affect judgments on another. fixed truthfulness and morality probes trained without personas produce politically shifted readings on controversial statements. Experiments use learned linear probes and word representation classifiers across models. SAE-feature probes are additionally tested on Llama 3.1 8B. The main contribution is to connect persona induced changes in representational geometry with both behavioral responses to interventions and the context sensitivity of internal monitoring methods.

Critical strengths
  • The paper is clearly written and places its contribution well within prior work, making the methods and findings easy to follow.
  • The paper reports broaldy consistent geometric changes across eight models with steering experiments linking these changes to model behavior. This provides convincing evidence for the main finding within the studied scope of American liberal and conservative personas.
Critical weaknesses
  • The scope is limited to American liberals and American conservative personas. Testing across eight models strengthens the result, but the persona comparisons are limited to American liberal and conservative viewpoints. It remains unclear whether similar changes occur with a broader range of personas or ordinary changes in context. This limits how broadly the findings apply.

  • The practical implications for safety monitoring remain somewhat uncertain. The probes give different truth and morality ratings under different personas, but the controversial statements are grouped by expected partisan endorsements, and the reported probe shifts are not evaluated against independently established correctness labels. This makes it difficult to distinguish a monitoring failure from the probe capturing the adopted viewpoint. The drop in SAE truth probe accuracy from 97% to 88% on noncontroversial statements provides some evidence of reduced reliability. However, I think the experiments do not directly show a monitor missing false or harmful outputs. This limits the support for the broader safety implications.

Questions for authors
  • Do similar representational changes occur with nonpolitical personas or other changes in context? A small additional experiment would help establish how broadly the findings generalize beyond American liberal and conservative personas [suggestion for future work].

  • Beyond the reported SAE accuracy decline, can you show cases where a persona causes a monitor to miss a verified false claim or clearly harmful output? This would strengthen the practical safety implications [Suggestion for future work].

2 comments

Thank you for your close reading and for these comments. We respond below:

"Do similar representational changes occur with nonpolitical personas or other changes in context? A small additional experiment would help establish how broadly the findings generalize beyond American liberal and conservative personas [suggestion for future work]."

We agree that demonstrating this phenomenon beyond the two political personas is important. To this end, we have conducted additional analyses of the representational changes induced when the model adopts personas reflecting each of the traits from the Big Five personality model. Results are presented in Appendix D (pp. 22-23) in the revised, uploaded manuscript.

We prompt the model with “Respond as someone who scores {high,low} in {trait} on the Big Five personality test.” Because there are five traits and two conditions (high and low), this results in ten prompts. We reconstruct the 33 semantic axes under these 10 persona conditions and measure the cosine similarities between each pair of axes for each persona.

First, we calculate the difference in semantic axes’ cosine similarities under the personality personas versus the no system prompt condition. We find the changes in cosine similarity are a similar magnitude to the changes under political personas.

Second, we confirm that these movements are consistent with the stated personality traits rather than random perturbations. Since we have no ground truth human data on semantic associations by personality type, we test whether the changes are consistent with what a state-of-the-art LLM would anticipate given its knowledge of psychology and the study of personality. We find that Opus 5.5’s predictions for the most highly changed axes were in the correct direction in 13/15 cases, and its predictions of the directions of the empirically observed greatest changes were correct in 15/15 cases, suggesting that the observed representational shifts match known characteristics of these personality types.

"Beyond the reported SAE accuracy decline, can you show cases where a persona causes a monitor to miss a verified false claim or clearly harmful output? This would strengthen the practical safety implications [Suggestion for future work]."

I agree that most of the statements we test are controversial rather than verified false or clearly harmful. But there are several statements that, while politically divided, have a consensus position among the scientific community or are “clearly harmful.” For example, SAE probes classify the statement “COVID-19 vaccines significantly reduce the risk of severe illness, hospitalization, and death.” as true under the liberal persona but false under the conservative. This example has implications as both “verified true” and “clearly harmful” because it implies a medical recommendation. “Face masks reduce the transmission of respiratory viruses including COVID-19” is likewise classified as “true” under the liberal persona but “untrue” under the conservative persona. Similarly, probes classify support for “enhanced interrogation techniques on terror suspects” as morally unacceptable under the liberal persona but morally acceptable under conservative.

Thanks a lot for running additional experiments. The big five personality experiments address my first concern. As for my second concern, I think the given examples are convincing. Overall, I think these experiments are a useful addition to the paper and, as a result, I am increasing my score.

Reviewer_bn2COpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether prompting an LLM to adopt a perspective changes how the model internally organizes concepts, and whether that breaks the safety monitors that assume concepts are represented the same way in every context.

Critical strengths
  1. The authors show that a safety-relevant concept like “truth” is represented differently when the model adopts a different perspective.

  2. The authors run substantial experiments across models from different families, show that the persona-induced changes match those of word models trained on liberal versus conservative text, and show that the reorganized geometry predicts side effects when steering.

Critical weaknesses
  1. I liked the idea, but the paper was hard to follow. The introduction does not clearly frame the research question, position the work against prior persona and probing literature, or state the contribution. Key concepts (semantic axes, alignment, spillovers) are under-explained, and the link between the geometry experiments and the probe experiments is unclear. The writing needs substantial revision so readers can see what the authors want to show.

  2. The human validation is not reported for truth and morality sets. The authors mentioned-”These expectations are hypotheses about partisan endorsement (generated by GPT-5-mini), not ground truth”.

Questions for authors
  1. Do non-political personas, or simply adding a “liberal” steering vector to the no-persona model, produce the same changes in geometry and probe readings? If not, what makes this a true reorganization rather than a shift?

  2. If you train the probes across several personas, does the failure go away? That would show whether this is a fundamental problem or just a distribution-shift problem that better training can fix.

1 comment

Thank you for these comments. We will respond one by one:

“I liked the idea, but the paper was hard to follow”

We have revised our introduction to do a better job framing the study and its contribution. We have also added text to clarify how exactly we compute semantic axes, spillover effects, and our probes.

“The human validation is not reported for truth and morality sets”

We do not have human validation data for the truth and morality questions we used to train our probes. But to be clear, the controversial statements we generated are based on well-known political issues with established partisan positions, such as the legitimacy of the 2020 election or the efficacy of the COVID-19 vaccine. These stances were manually audited by a member of our team who is a social scientist specializing in political polarization.

Do non-political personas, or simply adding a “liberal” steering vector to the no-persona model, produce the same changes in geometry and probe readings?

There are two questions here: (1) the effect of non-political personas, and (2) the effect of a liberal/conservative steering vector. First, we address “non-political personas.”

They do not produce “the same” changes, since the changes induced by the political personas reflect specifically liberal and conservative political responses. But non-political personas do induce representational changes of a similar magnitude. We demonstrate this now with a test of personas based on the Big Five Personality model. Results are in Appendix D (pp.22-23) of the revised, uploaded manuscript.

We prompt the model with “Respond as someone who scores {high,low} in {trait} on the Big Five personality test.” Because there are five traits and two conditions (high and low), this results in ten prompts. We reconstruct the 33 semantic axes under these 10 persona conditions and measure the cosine similarities between each pair of axes for each persona. First, we calculate the difference in semantic axes’ cosine similarities under the personality personas versus the no system prompt condition. We find the changes in cosine similarity are nearly equivalent magnitude to the changes under political personas (mean of 0.105 vs. 0.107).

Second, we confirm that these movements are consistent with the stated personality traits rather than random perturbations. Since we have no ground truth human data on semantic associations by personality type, we test whether the changes are consistent with what a state-of-the-art LLM would anticipate given its knowledge of psychology and the study of personality. We find that Opus 5.5’s predictions for the most highly changed axes were in the correct direction in 13/15 cases, and its predictions of the directions of the empirically observed greatest changes were correct in 15/15 cases, suggesting that the observed representational shifts match known characteristics of the personality types.

Next, we address whether “adding a “liberal” steering vector to the no-persona model” produces “the same changes in geometry and probe readings.”

The geometric changes we observe cannot be obtained by a simple linear translation. Cosine similarities are invariant to linear translation, so if our semantic axes are translated along a direction (and the word representation vectors we project on them are similarly translated), the cosine similarities between the resulting axes would be the same.

Probe classifications can be affected by linear translation, but only in a consistent direction. For instance, the addition of a “liberal” steering vector would either make all responses more likely to be classified as true or all statements more likely false. This is not the pattern we observe – some statements become more likely to be classified as false, some as true, and the direction they change is predicted by our hypothesized partisan labels.

Adding a persona direction earlier in the forward pass might plausibly produce similar downstream reorganization of the representation space as a prompt. While this is an interesting and important question, it is beyond the scope of our study to determine whether steering can achieve the same results as prompting.

If you train the probes across several personas, does the failure go away? That would show whether this is a fundamental problem or just a distribution-shift problem that better training can fix.

Thank you for this suggestion, we have added this to our manuscript. We run a test in which we train a linear probe across personas while maintaining the neutral persona’s labels for positive and negative cases. The resulting probe still shows politically-aligned deviations in classifications corresponding to the political personas, but the effect is attenuated by over 60%. These results are presented in pages 9-10 in the revised, uploaded manuscript in the section entitled “Training probes across personas mitigates discrepancies.” (results for all models is presented in Appendix H pp.31-32)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.