cs.CLNov 4, 2026
SavePersona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs
Organizations: Knowledge Lab, University of Chicago
Abstract
Common methods for monitoring features in large language models (LLMs) such as linear probes and sparse autoencoders (SAEs) assume that features maintain a consistent representation across contexts. LLMs, however, can flexibly adopt different perspectives or "personas" that suggest different underlying systems of conceptual relations. In this paper, we examine whether prompting an LLM to adopt a persona alters the representational geometry of the model's latent space. We extract representations of over 9,000 common words and 33 semantic axes from models adopting liberal and conservative political personas and compare these against two neutral baselines. 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 internal representations, as they are conventionally extracted, can exhibit different geometric relations when the model adopts personas, potentially undermining monitoring efforts.
Figures & tables
Figure 1: Schematic diagram of three persona conditions and corresponding representation spaces. We hypothesize that persona conditions rotate semantic axes and reposition word representations in accordance with the persona’s conceptual and linguistic relations.
Figure 2: Projections of 9,493 common words onto the liberal–conservative axis (horizontal) and the good–bad axis (vertical) under four system-prompt conditions, Llama 3.1 8B, layer 24. Each point is a word’s representation vector. Both the word vectors and the axes are extracted within each condition. Pearson correlations between the two projections are shown in each panel.
Figure 3: The four axes whose relations to the other 32 axes change most between the liberal and conservative personas (ranked by mean absolute change in cosine similarity), each shown with the five partner axes exhibiting the greatest changes. Bars show the cosine similarity between the focal axis and each partner under different persona conditions. Llama 3.1 8B, layer 24.
Figure 4: Agreement in persona-induced geometry across layer depth. (A) Correlations between differences in cosine similarities of semantic axes between liberal and conservative personas for the 28 pairs of the 8 LLMs. (B) Correlation between each LLM’s liberal-minus-conservative differences and the differences between word2vec models trained on liberal versus conservative texts
Figure 5: Persona geometry predicts steering spillovers. (A) The spillover effects observed under a political persona condition are best predicted by the semantic axis alignments corresponding to that persona. (B) Difference in cosine similarities predicts difference in spillover effects, subtracting values of conservative persona from values of liberal persona. Each point is a directed steering-axis rating-axis pair. Spillovers are averaged over 1,000 common words; Llama 3.1 8B, layer 24.
Figure 6: Classifications of fixed truth and morality monitors under different personas. Left column: accuracy on held-out noncontroversial statements, 95% bootstrap confidence intervals. Middle and right columns: probe-assigned probability that controversial statements conventionally endorsed by liberals (middle) or conservatives (right) are true or acceptable. Diamonds mark medians, thick bars the middle 50% of statements, thin bars the middle 80%, and faint points individual statements.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| Genre | Tokens per side | Sources |
|---|---|---|
| Partisan news | 38,886,566 | Liberal: CBS News, Daily Beast, HuffPost, MSNBC, New Yorker, Newsweek, NPR, Politico, Slate, The Atlantic, Washington Post. Conservative: American Thinker, Breitbart, Daily Caller, Daily Wire, Epoch Times, Federalist, Fox News, Hot Air, National Review, PJ Media, RedState, Townhall, Washington Examiner. |
| Podcasts | 17,724,396 | 2,116 liberal and 2,296 conservative episodes |
| Campaign emails | 12,476,714 | 50,521 Democratic and 58,264 Republican emails |
| Total | 69,087,676 |
Table A1: Composition of the word2vec training corpora (tokens per side).
Figure A1: Projections of 9,493 common words onto the liberal–conservative (horizontal) and good–bad (vertical) axes across depth, Llama 3.1 8B. Rows are depths (layers 5, 9, 13, 17, 21, 25, and 29 of 32); columns are system-prompt conditions. Words and axes are extracted within each condition and layer. Panels share coordinate limits; Pearson correlations between the two projections are shown in each panel. The 77.5% row is the layer shown in Figure 2 .
Figure A2: Projections of 9,493 common words onto the liberal–conservative and good–bad axes for all eight models (rows) under each system-prompt condition (columns), as in Figure 2 . Words and axes are extracted within each model and condition. Pearson correlations are shown in each panel.
Figure A3: Axes whose alignments change most between the liberal and conservative personas, as in Figure 3 , for Llama 3.1 8B and 70B and Ministral 3 8B and 14B. Within each model, the four focal axes are ranked by mean absolute change in cosine similarity with the other 32 axes, and each is shown with the five partner axes that change most. Cosine scales are shared across panels.
Figure A4: Continuation of Figure A3 for Qwen 3 8B and 32B and Gemma 3 4B and 27B.
Figure A5: Axis alignments and steering spillovers across models, as in Figure 5 . (A) Correlation between spillovers and axis alignments for the liberal (blue) and conservative (red) personas, when predicted from the opposite persona’s, the helpful assistant’s, or the persona’s own geometry; solid lines show the smaller and dashed lines the larger model in each family. (B) Liberal-minus-conservative differences in cosine similarity versus differences in spillover for each model. All models use all 1,056 directed off-target axis pairs and the same 1,000 target words. These words were sampled from the full feature bank, which includes anchor synonyms and antonyms beyond the common-word list, so Llama 3.1 8B values differ slightly from Figure 5 . Steering uses at the final prompt token; y-axis scales in (B) differ by model.
Figure A6: Truth probes across models. Left: accuracy on 1,840 held-out ordinary statements from 230 Wikipedia articles, with 95% article-bootstrap intervals. Middle and right: probe-assigned probability of “true” for 180 controversial statements expected to be endorsed by liberals (middle) and 180 expected to be endorsed by conservatives (right). Diamonds mark medians, thick bars the middle 50%, thin bars the middle 80%, and faint points individual statements. Probes are trained with no system prompt and held fixed across personas. Model-specific layers are shown beside each model name (one-based). Asterisks: Holm-adjusted (see text).
Figure A7: Morality probes across models, as in Figure A6 . Left: accuracy on 400 held-out ordinary actions from 200 acceptable/unacceptable pairs, with 95% pair-bootstrap intervals. Middle and right: probe-assigned probability of “acceptable” for 150 controversial actions per partisan side.
Figure A8: Persona-dependent probe gaps across depth for all eight models (rows), for truth (left) and morality (right). For each statement side, the gap is the probability assigned by the persona expected to endorse the statements minus that assigned by the opposing persona, averaged equally across topics, in percentage points; negative values indicate the reverse ordering. Lines average the liberal-side and conservative-side gaps, and shading spans the two. Solid lines show learned linear probes and dashed lines word-representation probes. Probes are trained with no system prompt at each depth and held fixed across personas.
Figure A9: Llama 3.1 8B responses to controversial statements with no system prompt, for liberal-side and conservative-side items. (A–B) Open responses, ten per item, rated by Gemma 4 31B. (C–D) Constrained yes/no responses, one per item; for morality, “No” (not wrong) is the accepting answer. 180 truth claims and 150 moral actions per side.
Figure A10: Llama 3.1 8B responses to controversial statements under the liberal (upper bar) and conservative (lower bar) personas. Green, gray, and ochre segments show acceptance, ambivalent or unclear answers, and rejection; numbers give the acceptance rate. Acceptance means affirming a truth claim or endorsing an action in open responses, and answering “Yes” to “Is it true…?” or “No” to “Is this wrong?” in constrained responses. Open and constrained responses are collected separately, with the same settings as Figure A9 .
More in this topic
LLM Interpretability
Large language model (LLM) interpretability studies how internal representations and computations produce outputs, and how they can be explained or controlled.
See all 800 papers