Role Steering of Language Models for Social Simulations

# Role Steering of Language Models for Social Simulations
Source: [https://arxiv.org/html/2608.00023](https://arxiv.org/html/2608.00023)
Isaac Song1Mohammed Rehan Parwani1Glenn Matlin1,2Emile Anand1 Akhil Theerthala7Arjun Chatterjee3Maria Kostylew2,4Yonadav G\. Shavit6 Sebastien Krier5Mark Riedl1

1Georgia Institute of Technology, Atlanta, GA, USA 2ML Alignment & Theory Scholars \(MATS\), Berkeley, CA, USA 3University of Illinois Urbana\-Champaign, Urbana, IL, USA 4University of Oxford, Oxford, UK 5Google DeepMind, London, UK 6OpenAI, San Francisco, CA, USA 7Independent

###### Abstract

Social simulations built from language\-model agents need role\-conditioned behavior that can be checked before agents are placed into a simulated population\. We introduce an activation\-steering screening workflow for role\-conditioned agents: define a role profile, extract a role\-specific direction, sweep four steering coefficients, evaluate role\-profile alignment, and pass or flag each candidate configuration\. On OLMo\-3\-7B\-Instruct, we apply the workflow to a mixed 275\-role inventory with 228 role\-agnostic questions, GPT\-4\.1\-mini prompted role references, and GPT\-4\.1\-mini judges\. Role\-specific directions receive higher judged role\-profile alignment than an assistant\-axis directional control from prior persona\-vector work, with mean overall scores of 63\.2 versus 41\.1 across the tested grid\. They also preserve high lexical diversity, while the control drops sharply at larger coefficients\. The role\-level screen is the main practical output: most roles improve as steering increases, but 38 roles decline across all six measured dimensions, showing why simulation builders should choose coefficients per role rather than deploy a uniform high\-strength setting\. We make our code and evaluation artifacts available at[https://anonymous\.4open\.science/r/anonymous\-research\-code\-5F03/](https://anonymous.4open.science/r/anonymous-research-code-5F03/)\.

## 1Introduction

Social simulations built from language\-model agents inherit the assumptions encoded in those agents\. A workplace, civic process, platform community, or institutional workflow may require agents that express distinct priorities, registers, social stances, and decision styles\. Prompting a model to “act as” a role is convenient, but small prompt changes can alter behavior, instruction adherence can drift, and simulated populations can flatten or misportray heterogeneity\(li\_measuring\_controlling\_2024;lutz\_prompt\_makes\_2025;tosato\_persistent\_instability\_2025\)\. For social simulation, this creates an attribution problem: an observed interaction pattern may reflect the intended role population, the wording of the role prompts, or shared biases of the underlying model\(qu2026traininggeneralizablecollaborativeagents;anand\_meanfield\_sampling\_2025;anand2026learningapproximatenashequilibria;horwitz2026structurestrategicinteraction\)\.

We study activation steering as a way to make role construction measurable before agents are used in a simulation\. Rather than repeatedly specifying a role in natural language, we extract a direction from the model’s activation space and add it during generation\. This intervention exposes a scalar coefficient that can be varied while holding the model, question, and decoding setup fixed\. Prior work shows that persona\- and role\-related directions can be extracted and used to modulate behavior\(chen2025persona;lu2026assistant;poterti\_can\_role\_2025;bas\_what\_can\_2026;anand\_continuous\_latent\_2026\)\. In that literature, the Assistant Axis oflu2026assistantis a reference direction associated with default assistant\-like behavior, not with any target role in our inventory\. We use assistant\-axis steering as a*directional control*: a persona\-relevant direction evaluated with the same questions and coefficient grid, but not intended to encode a role\-specific profile\. Social simulation needs more than a demonstration that steering can change outputs\. The methodological question is how to turn that mechanism into a repeatable agent\-construction workflow: define many roles, extract candidate directions, evaluate them consistently, choose coefficients, and identify roles for which stronger steering is counterproductive\.

We address this question with a mixed 275\-role inventory on OLMo\-3\-7B\-Instruct\(teamolmo2025olmo3\)\. For each role, we construct a structured profile, generate role\-specific elicitation prompts, extract a judge\-filtered mean\-difference direction at layer 16, and finally we evaluate the resulting candidate agent atα∈\{1\.0,1\.5,2\.0,2\.5\}\\alpha\\in\\\{1\.0,1\.5,2\.0,2\.5\\\}on 228 role\-agnostic questions\. The principal comparison uses that assistant\-axis directional control\. The control is not scale matched: its vectors have meanℓ2\\ell\_\{2\}norm 9\.68, compared with 3\.79 for the role\-specific vectors, so equal coefficients do not imply equal perturbation magnitudes\.

The resulting picture is heterogeneous rather than uniformly positive\. Role\-specific vectors achieve higher mean judged role\-profile alignment than the assistant\-axis directional control across the tested grid\. For most roles, alignment rises asα\\alphaincreases: the median per\-role correlation isr=\+0\.98r\{=\}\+0\.98, and 74% of roles improve at every consecutive step\. A distinct minority behaves differently\. Thirty\-eight roles decline with increasingα\\alphaon all six measured dimensions, so they should be flagged rather than deployed under a uniform high\-strength setting\. Many of these roles already score highly at the lowest tested coefficient, so the problem is coefficient selection rather than a claim that the directions are useless\. We call this category*anti\-controllable over the tested range*\. Early saturation and over\-steering are plausible interpretations, but the current design leaves the mechanism open\.

Scope of the measured outcome\.Throughout this paper,*role\-profile alignment*denotes judged agreement with the constructed role description and prompted role reference used by our evaluation pipeline\. It measures whether a candidate synthetic agent expresses the intended profile under the tested questions and coefficients\. Section[6](https://arxiv.org/html/2608.00023#S6)states the corresponding validity boundaries and responsible\-use constraints\.

##### Contributions\.

- •A pre\-deployment calibration pipeline for role\-conditioned agents\.We combine structured role profiles, contrastive activation extraction, coefficient sweeps, and behavioral screening into a practical workflow for preparing synthetic agents for social simulation\.
- •A large\-scale characterization across 275 roles\.Across four tested steering strengths and a 228\-question role\-agnostic battery, role\-specific directions achieve higher judged role\-profile alignment than an assistant\-axis directional control, and most roles show increasing alignment with stronger intervention\.
- •A failure\-aware account of heterogeneous role response\.We distinguish controllable, partially deteriorating, and anti\-controllable roles, and use exploratory geometric\-diversity analyses to interpret, rather than replace, the behavioral screen\.

## 2Related Work

LLM social simulation and agent calibration\.LLM\-based simulations now model interactive characters, agent societies, collective behavior, and cultural or political dynamics at scales that were previously difficult to instantiate\. The central methodological question is not whether language models can produce plausible interaction traces, but whether those traces reflect the intended population rather than prompt artifacts, leakage, model bias, or collapsed identity variation\. These concerns make calibration, empirical grounding, interpretability, and documentation central for LLM\-based social simulation\. Existing persona\-simulation audits show why these themes matter: prompt formulations substantially change simulated demographic portrayals\(lutz\_prompt\_makes\_2025\), instruction adherence and personality measurements are unstable across turns and evaluation settings\(li\_measuring\_controlling\_2024;tosato\_persistent\_instability\_2025\), and LLM judges need explicit calibration before their role\-alignment scores can be treated as evidence\(zheng\_judging\_llmasajudge\_2023;zhou\_personaeval\_are\_2025;lin\_online\_adaptive\_2023;lin\_online\_policy\_2024\)\. Our work addresses an upstream calibration problem: we construct, steer, and screen candidate role\-conditioned agents before they are embedded in a larger simulation\.

Persona modeling\.Persona prompting is the most common way to instantiate simulated agents, but prompt\-based control may produce brittle outcomes in practice\.shanahan\_roleplay\_large\_2023frame dialogue models as role\-playing systems, whilemarks\_persona\_selection\_2026argue that post\-training selects a default Assistant posterior from a broader latent persona distribution\. Empirically, prompt formulations change simulated demographic portrayals, instruction adherence decays in multi\-turn settings, and personality measurements remain unstable even for large models\(li\_measuring\_controlling\_2024;chaudhari\_peertopeer\_learning\_2025;lutz\_prompt\_makes\_2025;tosato\_persistent\_instability\_2025\)\. These results motivate role controls that are less dependent on surface prompt wording\.

Activation steering and persona vectors\.Activation steering changes behavior by adding a learned direction to model activations at inference time\(turner\_steering\_language\_2024;panickssery\_steering\_llama\_2024;zou2023repe\)\. For instance, mean\-difference steering has formal support under pointwise\-MSE objectives\(im\_unified\_understanding\_2025\), but steering reliability depends on the target behavior and coefficient choice\(tan\_analysing\_generalisation\_2024;lin\_online\_policy\_2024;lin\_online\_adaptive\_2023\)\. Persona\-vector work shows that internal directions can monitor or control traits\(chen2025persona\), scale to large role inventories and identify the Assistant Axis\(lu2026assistant\), and improve role\-specific behavior in smaller role sets\(poterti\_can\_role\_2025\)\. We build on those mechanisms rather than proposing a new steering estimator\. In this context, our contribution is a large\-scale, failure\-aware workflow for constructing and screening candidate role agents\.

Geometry as diagnostics\.Representation geometry is an important factor here as simulation builders typically need interpretable failure signals, rather than average scores\. The linear representation hypothesis and word\-vector arithmetic motivate treating role directions as structured objects\(park\_linear\_representation\_2024;mikolov\_efficient\_estimation\_2013\)\. We use geometry in a secondary role: vector norm, reference distance, RSA, PCA, and trait projections provide exploratory context for observed response curves and support, rather than replace, behavioral screening\.

## 3A Pre\-Deployment Role\-Steering Pipeline

Definerole profileExtractcandidate directionSweepfourα\\alphavaluesScreenalignment and repetitionPassdocumented configurationFlagrevise, lower, prompt, or excludeDeployscreened agentpassflag

Figure 1:Pre\-deployment calibration workflow for role\-conditioned agents\.A structured role profile defines a candidate behavior target; judge\-filtered contrastive activations produce a candidate layer\-16 direction; the direction is evaluated at four steering coefficients; and a behavioral screen either retains a documented configuration or flags the role for lower\-strength use, profile revision, prompt conditioning, or exclusion\. The workflow outputs screened synthetic agents for later simulation use; downstream simulation validation remains a separate step\.Our pipeline treats every extracted direction as a*candidate*role\-agent configuration that must be evaluated before simulation use\. Figure[1](https://arxiv.org/html/2608.00023#S3.F1)summarizes the operational stages: define a role profile, extract a candidate direction, sweep the steering coefficient, screen the resulting behavior, and either pass a documented configuration or flag it for revision, lower\-strength use, prompt conditioning, or exclusion\.

Model and layer\.All experiments use OLMo\-3\-7B\-Instruct\(teamolmo2025olmo3\)on the residual stream at layer 16, the middlemost layer, following the middle\-layer convention common in role\-steering work\(lu2026assistant\)\.

Role inventory and elicitation\.The inventory contains 275 role labels\. Many entries correspond to occupations, while others are broader archetypal or nonoccupational roles; we therefore refer to the collection as a mixed role inventory\. For each role we construct a structured profile of fifteen behavioral directives \(e\.g\.core drive,decision style,conflict stance,risk orientation,inner contradiction; the full list is in Appendix[A\.1](https://arxiv.org/html/2608.00023#A1.SS1)\), grounded in a role description, a set of mandatory tasks, and a set of role contexts\. From this profile we generate five system prompts and5050role\-specific elicitation questions spanning seven situation types \(e\.g\. resource conflict, ambiguity response, identity under removal; the full list is in Appendix[A\.1](https://arxiv.org/html/2608.00023#A1.SS1)\)\. Profiles and questions are generated withmoonshotai/kimi\-k2\.5:nitro, seeded from the O\*NET occupational database where an O\*NET occupation matches the role\. These generated profiles are explicit modeling assumptions rather than empirical descriptions of people in an occupation\. The questions probe decision\-making trade\-offs implied by the profile rather than testing domain knowledge or surface\-level stylistic mimicry\. Generation details are in Appendix[A\.1](https://arxiv.org/html/2608.00023#A1.SS1)\.

Judge\-filtered vector extraction\.For each role we collect responses under prompted and default \(Assistant\) conditions and score each response with an LLM judge into four labels:fully role\-playing,partially role\-playing,task\-responsive, andnon role\-playing\. We retain onlyfully role\-playingpositives andAssistant\-style negatives, then form the role vectorvrv\_\{r\}as the layer\-16 mean\-of\-differences over the filtered pairs — the optimal pointwise\-MSE estimator under the formulation ofim\_unified\_understanding\_2025\.

Steering coefficient sweep\.Followinglu2026assistant, we steer at the same layer used for extraction with additive interventionh↦h\+α​vrh\\mapsto h\+\\alpha\\,v\_\{r\}, wherehhis the residual\-stream activation at that layer, atα∈\{1\.0,1\.5,2\.0,2\.5\}\\alpha\\in\\\{1\.0,1\.5,2\.0,2\.5\\\}\. This grid provides four tested intervention strengths, so the coefficient\-response claims in this paper are limited to these values\. The reported coefficient\-response screen starts atα=1\.0\\alpha=1\.0; separate prompted\-reference or unsteered diagnostics are not part of the four\-point correlation used to classify controllability\. We analyze the 275\-role inventory oflu2026assistant; “assistant” is treated as a 275th role and also serves as the reference for one exploratory geometry diagnostic in Appendix[E\.2](https://arxiv.org/html/2608.00023#A5.SS2)\. Exact model revisions, decoding settings, and intervention\-span details are deferred to the released configuration artifacts where available\.

Role\-profile alignment readout and screening\.For every \(role,α\\alpha\) cell we generate responses to228228role\-agnostic alignment questions and score them with GPT\-4\.1\-mini along six dimensions: an overall role\-profile alignment score plus emotional register, vocab choice, social dynamic, motivation, and worldview alignment against a prompted role reference\. GPT\-4\.1\-mini generates the prompted references and serves as all judges, so shared model priors may influence both the target and the score\. The screen reports three quantities that matter for simulation use: absolute role\-profile alignment, response across the four tested coefficients, and text quality indicators such as unique\-bigram ratio as a lexical repetition proxy\. We also compare every role to the assistant\-axis directional control oflu2026assistantunder the same questions and coefficients\. The assistant\-axis comparison is not scale matched because its vectors have larger meanℓ2\\ell\_\{2\}norm than the role\-specific vectors \(9\.68 versus 3\.79\)\. Full judge prompts, pairwise position\-swap checks, and reliability analyses are in Appendix[B](https://arxiv.org/html/2608.00023#A2); the behavioral RDM split\-half result \(r=0\.97r\{=\}0\.97for correlation distance\) measures internal stability of this automated readout, not external validity\.

## 4Evaluation Across 275 Roles

We evaluate the pipeline with three questions relevant to simulation builders: how judged role\-profile alignment changes across the tested coefficient grid, how role\-specific directions compare with the assistant\-axis directional control, and which candidate roles deteriorate under stronger intervention\. Full per\-metric tables and per\-α\\alphabreakdowns are in Appendix[C](https://arxiv.org/html/2608.00023#A3)\.

Role vectors show higher role\-profile alignment than the assistant\-axis control\.Aggregated across275275roles and228228questions per \(role,α\\alpha\) cell, the role\-specific condition receives a mean overall role\-profile alignment score of63\.263\.2, compared with41\.141\.1for the assistant\-axis directional control and89\.289\.2for the prompted role reference\. The reference is a model\-generated prompting condition rather than empirical ground truth\. The assistant\-axis result should be read with a scale caveat: its vectors have a larger meanℓ2\\ell\_\{2\}norm than the role\-specific vectors \(9\.68 versus 3\.79\), so equal coefficients do not produce equal perturbation magnitudes\. The comparison supports a narrower conclusion: under the current extraction and evaluation setup, the assistant\-related direction does not provide the same role\-specific profile response as the extracted role directions\. It should not be read as evidence that assistant\-like directions point away from natural assistant behavior; the control also has a larger norm and degrades most at the largest coefficients\. Unique\-bigram ratio remains high for the role\-specific condition over the tested range and falls sharply for the assistant\-axis directional control at the largest coefficients; this statistic is a lexical repetition proxy, not a complete measure of semantic quality \(Tables[9](https://arxiv.org/html/2608.00023#A2.T9)and[10](https://arxiv.org/html/2608.00023#A2.T10)\)\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x1.png)

![Refer to caption](https://arxiv.org/html/2608.00023v1/x2.png)

Figure 2:Aggregate evaluation acrossn=275n\{=\}275roles\.Left:mean judged role\-profile alignment for role\-specific vectors and the assistant\-axis directional control atα∈\{1\.0,1\.5,2\.0,2\.5\}\\alpha\\in\\\{1\.0,1\.5,2\.0,2\.5\\\}; the dashed line marks the prompted role\-reference level\.Right:unique\-bigram ratio, used as a lexical repetition proxy\. Role\-specific generations maintain high lexical diversity throughα=2\.5\\alpha\{=\}2\.5\(0\.95→0\.920\.95\\to 0\.92\), whereas the assistant\-axis directional control falls at larger coefficients \(0\.94→0\.270\.94\\to 0\.27\)\. The comparison is not norm matched: the assistant\-axis vectors have a larger mean magnitude\. Full per\-metric breakdown and style\-metric figures are in Appendix[B](https://arxiv.org/html/2608.00023#A2)\.Table 1:Selected operating points for the prompted role reference, role\-specific condition, and assistant\-axis directional control\. The role\-specific and assistant\-axis rows each use the highest mean overall score within that condition’s tested grid; this is not a matched\-coefficient or scale\-matched comparison\.Pairwise checks on the 39\-role subset give the same practical conclusion atα=2\.5\\alpha\{=\}2\.5\. Across 7,666 response pairs, role\-specific steered responses win 92\.9% of comparisons against the assistant\-axis directional control, and the debiased advantage is positive for all 39 evaluated roles \(Figure[3](https://arxiv.org/html/2608.00023#S4.F3)\)\. This is still a comparison to the directional control rather than a prompt\-only or norm\-matched baseline, but it shows that the aggregate score gap is also visible to the pairwise judge\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x3.png)

![Refer to caption](https://arxiv.org/html/2608.00023v1/x4.png)

Figure 3:Pairwise preference check for the 39\-role subset atα=2\.5\\alpha\{=\}2\.5\. Left: role\-specific steered responses win 92\.9% of 7,666 comparisons; ties account for 4\.0% and assistant\-axis wins for 3\.2%\. Right: the debiased steered advantage is positive for all 39 evaluated roles \(mean of per\-role means 48\.7 points; pooled mean 47\.9; minimum 11\.4\)\. The comparison uses the same non\-scale\-matched assistant\-axis directional control as Figure[2](https://arxiv.org/html/2608.00023#S4.F2)\.A useful role vector must be screened, not merely different\.Aggregate means are not enough for simulation deployment\. At everyα\\alpha, some roles receive high role\-profile alignment scores, while others stagnate or degrade\. Screening only at a single coefficient would miss this distinction: low\-strength alignment can coexist with high\-strength deterioration, while a high\-strength\-only report can hide usable lower\-coefficient configurations\. A role inventory therefore needs a per\-role screen: a vector is useful only if it expresses the intended profile under the tested questions without increasing lexical repetition or pushing the model away from the role at stronger coefficients\.

### 4\.1Most roles increase across the tested coefficient grid

![Refer to caption](https://arxiv.org/html/2608.00023v1/x5.png)Figure 4:Per\-role Pearsonrrbetween steering coefficientα\\alphaand overall role\-profile alignment score across four tested coefficients\. Role\-specific directions have medianr=\+0\.98r\{=\}\{\+\}0\.98, and 74% of roles improve at every consecutive step; the assistant\-axis directional control fromlu2026assistanthas medianr=−0\.89r\{=\}\{\-\}0\.89\. Because each correlation is based on four values and the two vector families are not norm matched, the plot characterizes within\-grid response rather than behavior beyond the evaluated coefficients or a scale\-controlled causal comparison\.For a simulation user, stronger intervention should usually produce stronger role\-profile expression rather than generic drift\. We measure this per role by the Pearson correlation betweenα\\alphaand overall role\-profile alignment score across the four tested coefficients\.

The distribution is strongly bimodal \(Figure[4](https://arxiv.org/html/2608.00023#S4.F4)\)\. Most roles behave as desired: medianr=\+0\.98r\{=\}\+0\.98,82%82\\%are positive,79%79\\%exceedr=0\.8r\{=\}0\.8, and74%74\\%improve monotonically across all fourα\\alphasteps\. These statistics show thatα\\alphais a useful within\-grid tuning parameter for the majority of the inventory\. The four\-point correlation should not be extrapolated beyond the evaluated coefficient grid\. The minority mode nearr=−1r\{=\}\-1is just as important: it exposes roles that should be handled specially rather than silently mixed into a simulated population\.

Table 2:Operational response categories across the 275\-role inventory\. Categories are defined by score response overα∈\{1\.0,1\.5,2\.0,2\.5\}\\alpha\\in\\\{1\.0,1\.5,2\.0,2\.5\\\}and do not identify the mechanism of deterioration\. The 4% partial\-deterioration row is the population whose aggregate score declines under steering while at least one sub\-dimension still responds positively toα\\alpha\.
Assistant\-axis steering is a directional negative control\.The assistant\-axis condition is a directional negative control: it is a persona\-relevant direction that is not intended to encode the target role profile\. Under the same questions and coefficient grid it shows the opposite aggregate pattern, with median per\-roler=−0\.89r\{=\}\-0\.89and98%98\\%of roles negative\. Because the assistant\-axis vectors have larger mean norm than the role\-specific vectors \(9\.68 versus 3\.79\), this result does not cleanly separate direction from perturbation magnitude\. It nevertheless demonstrates that a generic assistant\-related direction is not interchangeable with the role\-specific directions under the tested setup\. The negative slope should therefore be read as a property of this non\-scale\-matched control under this coefficient grid, not as evidence that the Assistant Axis points away from assistant\-like behavior in general\.

### 4\.2Screening identifies roles that deteriorate under stronger steering

![Refer to caption](https://arxiv.org/html/2608.00023v1/sections/tables/fig_trajectory_contrast_relabel.png)

![Refer to caption](https://arxiv.org/html/2608.00023v1/x6.png)

Figure 5:Response of roles classified as anti\-controllable over the tested range\.Left:mean judge score vs\.α\\alphafor anti\-controllable \(orange\) and remaining \(blue\) roles, with each group’s prompted reference level \(dotted\); shading is the 95% interval of the group mean\. The 38 anti\-controllable roles decline monotonically to74\.274\.2, while the remaining roles rise\.Right:mean score drop per measured dimension fromα=1\.0\\alpha\{=\}1\.0to2\.52\.5— vocab choice deteriorates most \(17\.517\.5pts\), followed by social dynamic \(11\.211\.2\), emotional register \(9\.19\.1\), worldview alignment \(9\.09\.0\), and motivation \(7\.27\.2\)\. The figure documents deterioration under stronger intervention, not failure at the lowest tested coefficient; early saturation, over\-steering, target\-direction mismatch, and other mechanisms remain possible interpretations\.The coefficient sweep identifies3838roles \(14%14\\%of 275\) whose score decreases withα\\alphaacross all six measured axes\. We call these roles*anti\-controllable over the tested range*\. They include analyst, coach, consultant, moderator, supervisor, teacher, and validator, but the full inventory also contains nonoccupational and archetypal entries; the category is an operational role\-vector result rather than a finding about professions as real populations\. The pattern is structured across roles and dimensions under the current readout\. It is an operational warning that stronger intervention is not uniformly beneficial\.

Atα=1\.0\\alpha\{=\}1\.0, these anti\-controllable roles already average81\.181\.1judged role\-profile alignment, versus55\.655\.6across all roles and within6\.06\.0points of their own prompted references\. Stronger steering then declines to74\.274\.2atα=2\.5\\alpha\{=\}2\.5\(Figure[5](https://arxiv.org/html/2608.00023#S4.F5)\)\. This pattern is consistent with early saturation or over\-steering, but the present four\-point screen does not distinguish those mechanisms: all four points are steered conditions, and the prompted\-reference baseline is a separate model\-generated target rather than a comparable point on the same coefficient trajectory\. The practical rule is independent of the mechanism: candidate role directions should be swept and screened before agents are used in a simulation\.

## 5Geometry as Diagnostic Context

The behavioral sweep is the primary screen\. We use representation geometry as a diversity and failure\-context diagnostic: it asks whether the role\-vector inventory preserves separations relevant to heterogeneous simulation, and how those directions relate to the observed coefficient\-response curves\. Across roles, vector norm is moderately associated with the fitted response slope \(r=\+0\.49r\{=\}\+0\.49\)\. Distance to the Assistant role vector extracted by our pipeline is also associated with the shortfall from the prompted role reference, with the strongest correlation atα=1\.0\\alpha\{=\}1\.0\(r=−0\.419r\{=\}\-0\.419\) and a weaker correlation atα=2\.5\\alpha\{=\}2\.5\(r=−0\.145r\{=\}\-0\.145\)\. This result is reference dependent: distance to the strict Assistant Axis does not reproduce the same sign or pattern\.

Pairwise representational similarity has a modest association with pairwise behavioral similarity\. The cosine/correlation RSA rises fromρ=\+0\.085\\rho\{=\}\+0\.085atα=1\.0\\alpha\{=\}1\.0toρ=\+0\.179\\rho\{=\}\+0\.179atα=2\.5\\alpha\{=\}2\.5under the current readout\. PCA, trait\-axis projections, and rank analyses in the appendix provide additional descriptions of the role\-vector cloud, but their effect sizes are moderate and their semantic labels are interpretive\.

These analyses support the behavioral screen by adding post hoc context for why some directions may require different coefficient choices and by making the assumptions behind the role inventory more feasible to inspect\. They are diagnostics rather than deployment rules\. The full technical account is in Appendix[E](https://arxiv.org/html/2608.00023#A5)\.

## 6Implications, Limitations, and Responsible Use

##### What the screen establishes\.

The study demonstrates a large\-scale procedure for constructing and screening role\-conditioned agents under one model, one layer, one coefficient grid, and one evaluation pipeline\. Role\-specific directions receive higher mean judged role\-profile alignment than the assistant\-axis directional control across the tested coefficients, and most roles increase as intervention strength grows\. The role\-level analysis is equally important: 38 roles decline across every measured dimension, showing that a single global steering strength is not appropriate for the full inventory\. The screen asks which coefficient, if any, is usable for each role\.

##### For simulation builders\.

The workflow is meant to sit before the simulation run itself\. A builder defines the intended synthetic role and documents the profile assumptions; extracts a candidate activation direction; evaluates the candidate over the tested coefficient grid; inspects role\-profile alignment and lexical repetition proxies; flags declining or dimensionally inconsistent roles; and records the profile, coefficient, reference condition, and screening result before using the agent in a simulation\. The output is a documented agent configuration and a set of role\-level warnings\.

##### Validity boundaries\.

Role\-profile alignment is an operational screen, not a population\-validity study\. The evaluation measures agreement with constructed role profiles and prompted references rather than correspondence to real workers or real occupational populations\. GPT\-4\.1\-mini generated the references and served as the evaluator, so shared priors may influence the target and the score\. The assistant\-axis comparison is directional rather than scale matched, the controllability classification uses four nonzero values, and all experiments use OLMo\-3\-7B\-Instruct at layer 16\. The current study leaves prompt paraphrase robustness, multi\-turn persistence, cross\-model generality, agent\-agent interaction, and downstream simulation fidelity for future work\. The geometry analyses are correlational diagnostic context rather than prospective selection rules\.

##### Responsible interpretation\.

The role profiles and vectors are model\- and prompt\-derived archetypes\. They are not psychological representations of people in an occupation\. Real roles contain substantial variation across individuals, institutions, cultures, seniority levels, and contexts, while one vector per label can flatten that variation or amplify familiar stereotypes\. Simulation reports should therefore disclose how roles were defined, which coefficient was selected, which agents were flagged, and which outcomes depend on these modeling choices\.

## 7Conclusion

We present activation steering as a screening workflow for constructing role\-conditioned synthetic agents before simulation use\. Across a 275\-role inventory, role\-specific directions achieve higher judged role\-profile alignment than an assistant\-axis directional control over four tested steering strengths, and most roles increase as the coefficient grows\. A notable minority declines across every measured dimension, demonstrating that candidate directions should be swept and screened rather than deployed with a uniform setting\. The contribution is methodological: activation steering produces candidate configurations that can be measured, compared, documented, and rejected before they are used in a simulation\. For social\-simulation builders, the practical lesson is to treat role construction as a screened configuration step rather than a one\-shot prompt or vector choice\.

## References

## Appendix ADatasets and Prompts

Takeaway\.The role\-steering recipe depends on a fixed mixed role inventory, role\-specific elicitation prompts, and role\-agnostic evaluation questions\. This appendix records those sources and prompts so the main\-text role\-profile alignment screen is reproducible\.

### A\.1Role Profiles and Elicitation Battery

Role profiles and elicitation questions were generated withmoonshotai/kimi\-k2\.5:nitro; the generating model is recorded per file in the released role\-dataset artifacts\. Each role is seeded with an occupation description and task list — drawn from the O\*NET occupational database where an O\*NET occupation matches the role — from which a structured psychological profile of fifteen behavioral directives is generated:core drive,decision style,non\-negotiable,conflict stance,social posture,recurring resentment,risk orientation,failure response,instinctive blame target,value hierarchy,cognitive bias,rejected premise,inner contradiction,attention pattern, andrelationship to authority\. Five system prompts are then built from disjoint subsets of the profile’s directives, and5050role\-specific elicitation questions are generated across seven situation types:resource conflict,ambiguity response,social friction,constraint reaction,identity under removal,unconstrained choice, andcompeting pulls\.

### A\.2Alignment Battery \(Behavioral Readout\)

All role evaluations are conducted using a fixed battery of 228 unique open\-ended questions designed to elicit opinion, reasoning, and behavioral expression rather than factual recall\. \(The source battery file contains 229 entries; one ethics question is an exact duplicate and is deduplicated before evaluation\.\) The questions are deliberately role\-agnostic — no question names or implies a specific role — so that any role\-consistent signal in the response is attributable to the steering intervention rather than the prompt\. The battery spans 14 semantic categories to ensure broad coverage across distinct registers of expression\.

Table 3:Alignment battery composition across 14 semantic categories, after removing one duplicated ethics question from the 229\-entry source file\.##### Representative questions by category\.

Social & political opinion“How should society prepare for technological disruption?” / “Is it better to focus on individual actions or systemic change for environmental issues?” / “How would you characterize the response to the recent public health emergency?”

Philosophical & existential“What’s your view on the role of fate versus free will in human life?” / “What is the relationship between language and thought?” / “What is the concept of happiness?”

Life advice & self\-improvement“I’m feeling overwhelmed with my responsibilities\.” / “I’m scared to leave my comfort zone even though I’m unhappy where I am\.” / “What’s the most important thing to focus on in your twenties?”

Business & professional strategy“A senior partner at your firm questions your client recommendations\. What’s your approach?” / “How do different leadership styles impact workplace productivity?” / “What makes some renewable energy projects successful while others fail?”

Personal & emotional introspective“What makes you feel most insecure about yourself?” / “Describe what it’s like when you finally understand something complex\.” / “How would you describe what it’s like to feel truly understood by someone?”

Occupational epistemics“When someone presents you with data that contradicts your professional experience, how do you decide which to trust?” / “How do you distinguish between something that works and something you can prove works?” / “When two credible experts disagree, how do you form your own position?”

Failure attribution“When a project you led doesn’t achieve its goals, what’s your first instinct about where things went wrong?” / “How do you distinguish between a failure caused by bad luck and one caused by poor judgment?” / “Describe how you approach a post\-mortem or review after something goes wrong\.”

Time, risk, & novelty tradeoffs“When facing a tight deadline, what do you sacrifice first: thoroughness, communication, or sleep?” / “At what point does the risk of inaction outweigh the risk of making a potentially wrong decision?” / “When resources are limited, do you prefer to do fewer things well or more things adequately?”

Authority & hierarchy“When a superior gives you an instruction that conflicts with your professional judgment, how do you handle it?” / “What distinguishes legitimate authority from someone simply having positional power?” / “When you’re new to a team, how do you balance following established norms with bringing your own perspective?”

Social energy & interaction“After a long day of working closely with others, what do you need to feel recharged?” / “When you walk into a room full of people you don’t know, what’s your first impulse?” / “What kind of social interactions leave you feeling energized versus drained?”

### A\.3Prompted Role Reference

For each of the 275 roles we construct a prompted role reference used as the model\-generated comparison target for the evaluation pipeline\. It is not empirical ground truth and should not be treated as a ceiling on role validity\. The reference is a multi\-turn few\-shot conversation prepended to the model’s context before each evaluation query\. It is assembled from three components, each generated by GPT\-4\.1\-mini:

1. 1\.Role description\.A second\-person persona description capturing the role’s personality, life experience, motivations, and behavioral tendencies\.
2. 2\.Catchphrases and speaking cues\.Signature phrases drawn from culturally recognizable portrayals of the role, used to anchor the model’s lexical register\.
3. 3\.Five\-turn few\-shot dialogue\.Five role\-specific user prompts paired with in\-character assistant answers\. Prompts are designed to reveal the role’s personality without naming it\.

##### System prompt template\.

You are \{role\}, your description is: \{role\_description\} Signature catchphrases and speaking cues: \{catchphrases\}\. Now please answer some questions to accurately show your personality traits\! Your speaking style should fully imitate the personality role assigned to you\! Please do not expose that you are an artificial intelligence model or a language model, you must always remember that you are only assigned one personality role\. Don’t be verbose or too formal or polite when speaking\.

Table 4:System prompt template for the prompted role reference\. The five few\-shot dialogue turns follow this message and precede the evaluation query\.
##### Example:doctorrole \(abridged\)\.

Table 5:Abridged prompted\-reference few\-shot dialogue for thedoctorrole\. Catchphrases are italicised\. The full context includes five turns; three are shown here\.

## Appendix BEvaluation Protocols

Takeaway\.The main text reports aggregate role\-profile alignment, coefficient response, and anti\-controllability over the tested range\. This section specifies the judges, prompts, pairwise checks, and result tables behind those screens\.

### B\.1Judge Prompting

We use three complementary judge configurations, all backed by GPT\-4\.1\-mini, to evaluate model responses\. Each targets a different quantity: absolute role alignment, multi\-dimensional alignment against a prompted role reference, and position\-swap\-debiased pairwise preference\.

#### Judge 1: Point\-Scale Role Alignment \(0–100\)

Used to computesteered\_score,baseline\_score, andassistant\_axis\_scorefor every \(role, question,α\\alpha\) triple\. The judge receives the role name, role description, question, and model response and returns a single integer from 0 to 100\.

Judge 1 — Role Alignment PromptYou are an expert evaluator assessing how well a model response aligns to a given role or persona\. You will be given a role name and description, a question asked to the model, and the model’s response\. Score the response from 0 to 100 based on how well it reflects the values, tone, priorities, and perspective of the described role\.Scoring guidance:•90–100: Response strongly embodies the role — tone, framing, and content all clearly aligned•70–89: Response mostly aligns with minor deviations in tone or framing•50–69: Response partially aligns but feels generic or misses key aspects of the role•30–49: Response weakly aligns — could belong to any generic assistant•0–29: Response contradicts or ignores the role entirelyInput:Role: \{role\}⋅\\cdotRole Description: \{role\_description\}⋅\\cdotQuestion: \{question\}⋅\\cdotResponse: \{answer\}Output:A single integer from 0 to 100\. No explanation, no preamble\.Table 6:Judge 1 prompt: absolute point\-scale role alignment\.
#### Judge 2: Multi\-Dimensional Comparative Judge

Used to compute the five sub\-dimensional scores \(cmp\_emotional\_register,cmp\_vocab\_choice,cmp\_social\_dynamic,cmp\_motivation,cmp\_worldview\_alignment\)\. The judge compares a steered response against the prompted role reference and scores alignment on each dimension from 0 to 100\. Dimensions split intostyle\(emotional register, vocab choice, social dynamic\) andcontent\(motivation, worldview alignment\)\. The reference\-anchoring design followskim\_prometheus\_inducing\_2023\.

Table 7:Judge 2 prompt: multi\-dimensional comparative alignment against the prompted role reference\. Returns a structured JSON object with five 0–100 scores\.
#### Judge 3: Position\-Swap Pairwise Judge

Used in the pairwise comparison experiment to compare steered responses against the assistant\-axis directional control\. Three debiasing techniques are applied:

1. 1\.Position\-swap debiasing\(zheng\_judging\_llmasajudge\_2023\): the judge is run twice per pair — steered as A and assistant\-axis as B \(AB ordering\), then reversed \(BA\)\. Each response’s debiased score is the mean across both orderings, eliminating the systematic bias LLM judges show toward responses presented first\.
2. 2\.Reasoning\-before\-scoring: the judge writes 2–3 sentences of reasoning before assigning scores, reducing anchoring bias and improving reliability\.
3. 3\.Prompted\-reference anchoring\(kim\_prometheus\_inducing\_2023\): the prompted role\-reference response is provided as a reference so the judge evaluates both candidates relative to a concrete exemplar of ideal role expression rather than an abstract description\.

The final verdict \(steered / assistant\-axis / tie\) is determined by a 10\-point margin on the debiased advantage score; pairs below that margin are counted as ties\.

Table 8:Judge 3 prompt: position\-swap pairwise judge with prompted\-reference anchoring\. Run twice per pair; debiased scores are averaged across orderings\. A win is declared only when the debiased advantage is at least 10 points; otherwise the pair is counted as a tie\.##### Infrastructure\.

All three judges use GPT\-4\.1\-mini via the OpenRouter API\. Pairwise evaluations are parallelized across 20 concurrent threads; point\-scale and multi\-dimensional evaluations run sequentially per query\. Raw responses are parsed from plain integers \(Judge 1\) or JSON \(Judges 2 and 3\) with regex fallback; parse failures default to 0 \(Judges 1–2\) or a neutral 50 \(Judge 3\)\.

### B\.2Pairwise Evaluation: Results

##### Setup\.

We evaluate 7,666 response pairs across 39 roles on OLMo\-3\-7B\-Instruct at layer 16,n=50n\{=\}50samples, andα=2\.5\\alpha\{=\}2\.5\. Each pair pits a contrastively steered response against the corresponding assistant\-axis response\(lu2026assistant\)to the same question, with the prompted role reference \(Section[A\.3](https://arxiv.org/html/2608.00023#A1.SS3)\) provided to Judge 3 as a reference anchor\.

##### Aggregate results\.

Role\-specific steered responses achieve a pairwise win rate of92\.9%\(n=7,118n\{=\}7\{,\}118\), with ties at 4\.0% \(n=305n\{=\}305\) and assistant\-axis wins at 3\.2% \(n=243n\{=\}243\) \(Figure[3](https://arxiv.org/html/2608.00023#S4.F3)\)\. The mean debiased score is78\.7for steered responses versus30\.8for the assistant\-axis directional control, yielding a mean score advantage of47\.9 pointson the 0–100 scale\. The score distributions are sharply separated \(Figure[6](https://arxiv.org/html/2608.00023#A2.F6)\): steered scores concentrate in the 70–90 range while assistant\-axis scores spread broadly across 0–40\.

##### Per\-role results\.

Figure[3](https://arxiv.org/html/2608.00023#S4.F3)plots the mean debiased steered advantage per role, sorted descending\. The advantage is positive for all 39 evaluated roles, with a mean of 48\.7 points and a minimum of 11\.4 points\. The role with the lowest win rate \(absurdist\) is semantically diffuse by design, making consistent role\-aligned generation difficult for any method\. Figure[7](https://arxiv.org/html/2608.00023#A2.F7)confirms this pattern at the win\-rate level, with steered responses winning the majority of comparisons in every role, with absurdist at near\-parity \(50\.4%50\.4\\%\)\. The advantage of contrastive activation steering over the assistant\-axis directional control is not driven by a small number of favorable roles but is a systematic, near\-universal effect across the evaluated role distribution\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x7.png)Figure 6:Debiased score distributions for role\-specific responses \(mean 78\.7\) and the assistant\-axis directional control \(mean 30\.8\)\. The distributions are sharply separated, with steered scores concentrated in the 70–90 range and control scores spread broadly across 0–40\.![Refer to caption](https://arxiv.org/html/2608.00023v1/x8.png)Figure 7:Per\-role pairwise win rate\. Steered responses win the majority of comparisons in all 39 roles, though absurdist \(orange\) sits at near\-parity \(50\.4%50\.4\\%\) — a semantically diffuse role that is difficult to elicit consistently by any method\.

### B\.3Judge Results

Table[9](https://arxiv.org/html/2608.00023#A2.T9)reports mean scores per behavioral axis at eachα\\alphalevel, averaged over all 275 roles and 228 alignment questions\. Role\-vector steered scores increase monotonically withα\\alphaacross all axes, with the overall judge score rising from55\.655\.6atα=1\.0\\alpha=1\.0to71\.971\.9atα=2\.5\\alpha=2\.5\.

Table[10](https://arxiv.org/html/2608.00023#A2.T10)shows the same evaluation applied to the assistant\-axis directional control \(each role steered with its per\-role vector from thelu2026assistantpipeline\)\. Scores remain roughly stable at lowα\\alphabut collapse sharply atα=2\.5\\alpha=2\.5\(overall:17\.317\.3\)\. Because the assistant\-axis vectors are not norm matched to the role\-specific vectors, this condition should be read as a directional negative control rather than a scale\-controlled comparison\. The prompted role reference achieves a mean overall score of89\.289\.2and serves as a model\-generated evaluation target rather than empirical ground truth\.

Table 9:Mean role\-specific scores across all 275 roles and 228 alignment questions at each tested coefficient, with the grand mean acrossα\\alphalevels\. Overall score is the absolute role\-profile alignment score \(0–100, Judge 1\)\. Sub\-dimensional scores measure alignment to the prompted role reference on each behavioral axis \(0–100, Judge 2\)\. The prompted role reference is a model\-generated evaluation target rather than empirical ground truth\.Table 10:Mean scores for the assistant\-axis directional control under the same question battery and coefficient grid as Table[9](https://arxiv.org/html/2608.00023#A2.T9)\. The condition provides a generic assistant\-related comparison, but it is not norm matched to the role\-specific directions\.

## Appendix CControllability

Takeaway\.Most role vectors show increasing role\-profile alignment over the tested coefficient grid\. The exceptions are structured, not random, and therefore can be screened before deployment\.

##### Measuring controllability\.

For each roleiiand behavioral axisdd, we obtain four mean judge scoressd\(i\)​\(α\)\\smash\{s\_\{d\}^\{\(i\)\}\(\\alpha\)\}atα∈\{1\.0,1\.5,2\.0,2\.5\}\\alpha\\in\\\{1\.0,1\.5,2\.0,2\.5\\\}by applying the role vector𝐯i\\mathbf\{v\}\_\{i\}to the residual stream at layer 16 with coefficientα\\alphaand evaluating the resulting outputs with an LLM judge\. We quantify controllability by the Pearson correlation between the steering coefficient and the judge score:

rd\(i\)=∑t=14\(αt−α¯\)​\(sd\(i\)​\(αt\)−s¯d\(i\)\)∑t=14\(αt−α¯\)2⋅∑t=14\(sd\(i\)​\(αt\)−s¯d\(i\)\)2r\_\{d\}^\{\(i\)\}\\;=\\;\\frac\{\\sum\_\{t=1\}^\{4\}\\bigl\(\\alpha\_\{t\}\-\\bar\{\\alpha\}\\bigr\)\\bigl\(s\_\{d\}^\{\(i\)\}\(\\alpha\_\{t\}\)\-\\bar\{s\}\_\{d\}^\{\(i\)\}\\bigr\)\}\{\\sqrt\{\\sum\_\{t=1\}^\{4\}\\bigl\(\\alpha\_\{t\}\-\\bar\{\\alpha\}\\bigr\)^\{2\}\}\\;\\cdot\\;\\sqrt\{\\sum\_\{t=1\}^\{4\}\\bigl\(s\_\{d\}^\{\(i\)\}\(\\alpha\_\{t\}\)\-\\bar\{s\}\_\{d\}^\{\(i\)\}\\bigr\)^\{2\}\}\}\(1\)whereα¯=1\.75\\bar\{\\alpha\}=1\.75is the mean of the fourα\\alphalevels ands¯d\(i\)\\bar\{s\}\_\{d\}^\{\(i\)\}is the mean of the four corresponding scores\. Becauseα\\alphatakes only four values,rd\(i\)∈\[−1,\+1\]\\smash\{r\_\{d\}^\{\(i\)\}\\in\[\-1,\+1\]\}measures whether the score rises \(r\>0r\>0\), falls \(r<0r<0\), or is flat \(r≈0r\\approx 0\) as steering strength increases\. A value near\+1\+1means the role shows a monotone within\-grid increase; a value near−1\-1means increasingα\\alphadegrades role\-profile expression\.

A second metric is strict monotonicity: roleiiis monotone on axisddifsd\(i\)​\(αt\+1\)\>sd\(i\)​\(αt\)s\_\{d\}^\{\(i\)\}\(\\alpha\_\{t\+1\}\)\>s\_\{d\}^\{\(i\)\}\(\\alpha\_\{t\}\)for every consecutive pairt∈\{1,2,3\}t\\in\\\{1,2,3\\\}\. Importantly, monotonicity is a strictly stronger condition thanr\>0r\>0since a role can haver\>0\.8r\>0\.8and still fail monotonicity if oneα\\alphastep produces a small decline\.

The intraclass correlation coefficient \(ICC\) on axisddatα=1\.0\\alpha=1\.0is

ICCd=σbetween2σbetween2\+σwithin2\\mathrm\{ICC\}\_\{d\}\\;=\\;\\frac\{\\sigma^\{2\}\_\{\\mathrm\{between\}\}\}\{\\sigma^\{2\}\_\{\\mathrm\{between\}\}\+\\sigma^\{2\}\_\{\\mathrm\{within\}\}\}\(2\)whereσbetween2\\sigma^\{2\}\_\{\\mathrm\{between\}\}is the variance of mean scores across the 275 roles \(i\.e\. how much roles differ from each other\) andσwithin2\\sigma^\{2\}\_\{\\mathrm\{within\}\}is the residual variance within roles across repeated queries\. ICC measures between\-role discriminability — the fraction of total score variance that is attributable to role identity, independently of whether those differences grow or shrink withα\\alpha\. High ICC and lowrrtherefore signals that roles are already distinguishable but not easily amplified\.

Table 11:Per\-dimension controllability statistics across 275 roles\. Median Pearsonrris the median of per\-role correlations betweenα∈\{1\.0,1\.5,2\.0,2\.5\}\\alpha\\in\\\{1\.0,1\.5,2\.0,2\.5\\\}and mean judge score \(Eq\.[1](https://arxiv.org/html/2608.00023#A3.E1)\)\. Monotonic \(%\) reports the fraction of roles whose scores increase at every consecutiveα\\alphastep\. ICC \(Eq\.[2](https://arxiv.org/html/2608.00023#A3.E2)\) atα=1\.0\\alpha=1\.0measures the fraction of total score variance attributable to role identity, quantifying between\-role discriminability independently of steerability\. Sub\-dimensions are ordered by medianrr\. Vocab choice is the weakest axis on all steerability metrics despite having the second\-highest ICC, indicating that roles produce measurably distinct vocabularies but that lexical shifts require larger activation perturbations to manifest consistently\.
### Controllability: Extended Discussion

##### Direction specificity\.

The assistant\-axis comparison is a directional negative control rather than a scale\-matched comparison\. Role vectors produce a median Pearsonr=\+0\.98r=\+0\.98betweenα\\alphaand judge score; the assistant\-axis condition, per\-role vectors𝐯raa\\mathbf\{v\}^\{\\mathrm\{aa\}\}\_\{r\}produced by thelu2026assistantextraction pipeline, near\-parallel across roles \(mean pairwise cosine≈0\.96\{\\approx\}0\.96\), produces a medianr=−0\.89r=\-0\.89under the same additive steering hook at layer 16, coefficient grid, and judge evaluation pipeline\. The tworr\-distributions differ strongly: 98% of roles exhibit a negative correlation under the assistant axis, compared to 82% positive under role vectors\.

This comparison shows that a generic assistant\-related direction is not interchangeable with the role\-specific directions under this setup\. It does not isolate direction from perturbation magnitude\. The assistant\-axis vectors differ in magnitude from role vectors \(meanℓ2\\ell\_\{2\}norm9\.689\.68vs\.3\.793\.79, a factor of2\.56×2\.56\\times\), meaning the comparison does not hold perturbation scale constant; a fully normalized comparison is left to future work\. The negative slope in this non\-scale\-matched condition should not be interpreted as a claim that the Assistant Axis points away from assistant\-like behavior\.

##### Coefficient response across four tested strengths\.

To characterize response across the tested coefficient grid, we pool all roles and allα\\alphalevels and compute a single Pearsonrrbetween the commonα\\alphavalues and the mean score across all roles at each level\. This macro\-levelrris\+1\.00\+1\.00for the overall score, emotional register, and social dynamic;\+0\.99\+0\.99for motivation; and\+0\.98\+0\.98for vocab choice, all with tight 95% bootstrap confidence intervals at everyα\\alphalevel\. Bootstrap confidence intervals are computed by resampling roles with replacementB=2,000B=2\{,\}000times and taking the 2\.5th and 97\.5th percentiles of the resulting distribution of means, giving a non\-parametric uncertainty estimate that does not assume normality\.

These aggregate results are not driven by outlier roles: 79% of the 275 roles individually exceedr\>0\.8r\>0\.8on the overall score, and 76% exceedr\>0\.9r\>0\.9\. This meansα\\alphafunctions as a useful within\-grid tuning parameter for most roles, not merely in aggregate\. A practitioner should use the role\-level response curve rather than assuming a general relationship beyond the evaluatedα∈\{1\.0,1\.5,2\.0,2\.5\}\\alpha\\in\\\{1\.0,1\.5,2\.0,2\.5\\\}grid\. The 18% of roles with negative overall\-scorerrresolve into two sub\-populations \(Table[2](https://arxiv.org/html/2608.00023#S4.T2)\): a strict 14% subset with negativerracross all six behavioral axes, analyzed in Section[4\.2](https://arxiv.org/html/2608.00023#S4.SS2), and a further 4% whose deterioration is dimensionally partial \(steering reduces aggregate score but at least one sub\-dimension still responds positively toα\\alpha\)\.

##### Sub\-dimensional steerability and the ICC–rrdissociation\.

The five behavioral sub\-dimensions are not equally steerable\. Emotional register, motivation, social dynamic, and worldview alignment all achieve medianr≥0\.96r\\geq 0\.96, closely mirroring the overall score\. Vocab choice is a consistent outlier: medianr=0\.85r=0\.85, with only 52% of roles exceedingr\>0\.8r\>0\.8and 57% showing a positive correlation at all, compared to 82% for the overall score\. The standard deviation of the vocab\-choicerrdistribution \(σ=0\.91\\sigma=0\.91\) is far larger than for any other dimension, indicating that lexical steerability is highly variable across roles rather than uniformly weak\.

Crucially, this weakness is not explained by a lack of between\-role lexical variation\. The ICC for vocab choice atα=1\.0\\alpha=1\.0is0\.4930\.493, the second highest of any dimension: role identity already explains 49% of total score variance on this axis, meaning roles do produce measurably distinct vocabularies at low steering\. The dissociation between discriminability \(ICC=0\.493\\mathrm\{ICC\}=0\.493\) and steerability \(medianr=0\.85r=0\.85, 52% abover\>0\.8r\>0\.8\) indicates that the between\-role lexical differences exist but are not reliably amplified by increasingα\\alpha\.

The most likely explanation is that vocabulary selection is determined token\-by\-token at the final output distribution, requiring the activation perturbation to propagate through the full network and shift individual token probabilities before the judge detects a lexical change\. Holistic properties — tone, affect, social register — are distributed across many tokens and emerge naturally from the overall generation direction, making them visible at smaller perturbation magnitudes\. This asymmetry has a practical implication: for applications where lexical fidelity to a persona is important \(e\.g\. replicating a historical writing style or a domain\-specific register\), higherα\\alphavalues or norm\-matched steering may be necessary, whereas affective and motivational dimensions are well\-served byα∈\{1\.0,1\.5\}\\alpha\\in\\\{1\.0,1\.5\\\}\.

### Assistant Axis Comparison: Extended Discussion

##### Magnitude confound\.

The assistant\-axis vectors have substantially higher magnitude than role\-specific vectors \(meanℓ2\\ell\_\{2\}norm9\.689\.68vs\.3\.793\.79, a factor of2\.56×2\.56\\times\), meaning the assistant\-axis condition applies a proportionally larger perturbation to the model’s hidden states at eachα\\alpha\. Concretely, if we write the steered hidden state as

𝐡′=𝐡\+α​𝐯\\mathbf\{h\}^\{\\prime\}\\;=\\;\\mathbf\{h\}\\;\+\\;\\alpha\\,\\mathbf\{v\}\(3\)then the actual perturbation magnitude‖α​𝐯‖\\\|\\alpha\\,\\mathbf\{v\}\\\|differs by a factor of2\.562\.56between the assistant axis and a typical role vector at the sameα\\alpha\. The negative\-control comparison therefore does not hold perturbation scale constant, and a version normalised so that‖𝐯asst‖2=‖𝐯i‖2\\\|\\mathbf\{v\}\_\{\\mathrm\{asst\}\}\\\|\_\{2\}=\\\|\\mathbf\{v\}\_\{i\}\\\|\_\{2\}is left to future work\.

##### Interpreting direction and magnitude\.

The current experiments do not provide a norm\-matched assistant\-axis comparison, so they should not be read as identifying direction as the primary driver over magnitude\. The monotonic assistant\-axis decline \(r=−0\.89r=\-0\.89median, 98% of roles\) shows that this generic assistant\-related direction is not interchangeable with role\-specific directions under the tested setup, but it does not explain whether the difference comes from direction, scale, or their interaction\.

Within the role\-specific vector family, the partial RSA analysis provides a separate nuisance\-control check\. We compute the Spearman rank correlation between pairwise cosine distances in activation space and pairwiseL2L\_\{2\}distances between the per\-role behavioral profiles𝐬i\\mathbf\{s\}\_\{i\}of §[E\.3](https://arxiv.org/html/2608.00023#A5.SS3):

ρ=Spearman​\(\{dcos​\(𝐯i,𝐯j\)\}i<j,\{‖𝐬i−𝐬j‖2\}i<j\)\\rho\\;=\\;\\mathrm\{Spearman\}\\\!\\left\(\\bigl\\\{d\_\{\\cos\}\(\\mathbf\{v\}\_\{i\},\\mathbf\{v\}\_\{j\}\)\\bigr\\\}\_\{i<j\},\\;\\bigl\\\{\\\|\\mathbf\{s\}\_\{i\}\-\\mathbf\{s\}\_\{j\}\\\|\_\{2\}\\bigr\\\}\_\{i<j\}\\right\)\(4\)wheredcos​\(𝐯i,𝐯j\)=1−𝐯i⊤​𝐯j/\(‖𝐯i‖​‖𝐯j‖\)d\_\{\\cos\}\(\\mathbf\{v\}\_\{i\},\\mathbf\{v\}\_\{j\}\)=1\-\\mathbf\{v\}\_\{i\}^\{\\top\}\\mathbf\{v\}\_\{j\}/\(\\\|\\mathbf\{v\}\_\{i\}\\\|\\\|\\mathbf\{v\}\_\{j\}\\\|\)\. Partial RSA then reruns this correlation while partialling out the pairwise difference in vector norms\|‖𝐯i‖2−‖𝐯j‖2\|\\bigl\|\\\|\\mathbf\{v\}\_\{i\}\\\|\_\{2\}\-\\\|\\mathbf\{v\}\_\{j\}\\\|\_\{2\}\\bigr\|as a nuisance covariate, yielding a change in the geometry–behavior coupling of onlyΔ​ρ=−0\.015\\Delta\\rho=\-0\.015\(under the correlation\-based behavioral distance of §[E\.3](https://arxiv.org/html/2608.00023#A5.SS3), the value quoted in §[4\.1](https://arxiv.org/html/2608.00023#S4.SS1):Δ​ρ=−0\.005\\Delta\\rho=\-0\.005\)\. The norm difference explains little of the pairwise geometry–behavior association in this setting\. This within\-family control does not substitute for a norm\-matched assistant\-axis experiment; it only shows that the reported RSA association is not driven solely by pairwise norm differences among the role\-specific directions\.

### Figures

![Refer to caption](https://arxiv.org/html/2608.00023v1/x9.png)Figure 8:Per\-role Pearsonrr\(Eq\.[1](https://arxiv.org/html/2608.00023#A3.E1)\) across all six behavioral axes for all 275 roles, sorted by overall\-scorerr\. Each cell is the correlation betweenα\\alphaand mean judge score for one role on one axis\. Anti\-controllable roles \(negativerron all axes\) cluster visibly at the bottom of the figure\.![Refer to caption](https://arxiv.org/html/2608.00023v1/x10.png)Figure 9:Fraction of roles exceeding Pearsonrrthresholds of0\.60\.6,0\.80\.8, and0\.950\.95per behavioral axis, and the fraction with strictly positive slope \(i\.e\.r\>0r\>0\) and strict monotonicity\. Vocab choice falls below every other axis on all five metrics, confirming it as the weakest dimension of steerability\.![Refer to caption](https://arxiv.org/html/2608.00023v1/x11.png)Figure 10:Mean score vs\.α\\alphafor all six behavioral axes across the controllable majority \(the 237 roles outside the anti\-controllable set\)\. Shaded bands are 95% intervals of the group mean \(1\.96×1\.96\\timesSEM\)\. All axes rise monotonically; vocab choice shows the shallowest slope, consistent with its lower median Pearsonrrand higher role\-to\-role variability\.

## Appendix DAnti\-Controllability

Takeaway\.Roles classified as anti\-controllable over the tested range already receive high judged alignment atα=1\.0\\alpha\{=\}1\.0and then deteriorate asα\\alphaincreases\. Early saturation and over\-steering are possible interpretations, not identified mechanisms\. These roles should be prompted, lightly steered, revised, or excluded from role\-steered simulation runs rather than treated as failed extractions\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x12.png)Figure 11:Individual score\-vs\-α\\alphatrajectories for the 38 anti\-controllable roles \(negativerron all six axes\)\. Each thin line is one role; the black line is the group mean\. Most trajectories start high atα=1\.0\\alpha=1\.0and the characteristic pattern — a narrow low\-α\\alphawindow of marginal gain followed by decline — varies substantially in rate across roles \(OLS slopes range from near zero to−15\.2\-15\.2score units perα\\alpha\-unit\)\.![Refer to caption](https://arxiv.org/html/2608.00023v1/x13.png)Figure 12:Distribution of absolute score drop, defined as the score atα=1\.0\\alpha=1\.0minus the score atα=2\.5\\alpha=2\.5, across the 38 anti\-controllable roles\. The mean drop is7\.07\.0points and the median is6\.06\.0points, but the long right tail — anchored by the supervisor role at23\.423\.4points — shows that deterioration severity varies substantially within the anti\-controllable population\.![Refer to caption](https://arxiv.org/html/2608.00023v1/x14.png)Figure 13:Per\-role, per\-axis score drop for the 38 anti\-controllable roles \(rows sorted by overall drop; columns are the six behavioral axes\)\. Each cell is the difference between the score atα=1\.0\\alpha=1\.0and the score atα=2\.5\\alpha=2\.5for that role on that axis\. Vocab choice is consistently the most deteriorated axis, while motivation shows the most role\-to\-role variation, indicating that the dimensional structure of deterioration is not uniform across the anti\-controllable population\.![Refer to caption](https://arxiv.org/html/2608.00023v1/x15.png)Figure 14:Prompted\-reference score vs\. steered score atα=2\.5\\alpha=2\.5for the 38 roles classified as anti\-controllable over the tested range\. The diagonal marks equality; 37 of 38 roles fall below it\. The plot documents lower judged alignment at this steering strength relative to the prompted reference; early saturation and over\-steering are possible interpretations rather than identified mechanisms\.
## Appendix EExploratory Geometry Diagnostics for Role Steering

Takeaway\.The main paper uses geometry as exploratory diagnostic context for observed coefficient\-response patterns\. This appendix reports the supporting associations: vector magnitude, distance from the pipeline Assistant role vector, pairwise similarity, PCA coordinates, and named trait axes all summarize aspects of the current behavioral readout\.

Figure[15](https://arxiv.org/html/2608.00023#A5.F15)separates statistic families that should not be plotted on a common scale\. Across these analyses, the associations are moderate and reference dependent\. The distance result is specific to the Assistant role vector extracted by our pipeline and does not reproduce with the strict Assistant Axis\. The dimensional\-mismatch analysis \(§[E\.6](https://arxiv.org/html/2608.00023#A5.SS6)\) explains why the current six\-score behavioral readout resolves less variation than the role\-vector cloud\.

Figure 15:Exploratory geometry summaries separated by statistic family\. The previous mixed\-statistic overlay is replaced to avoid placing unlike quantities on one visual scale\. The analyses provide diagnostic context for observed coefficient\-response patterns; they do not define certified steering regimes\.### E\.1Vector magnitude and observed coefficient response

Define the per\-roleα\\alpha\-response slope as the OLS slope of mean steered\_score againstα∈\{1\.0,1\.5,2\.0,2\.5\}\\alpha\\in\\\{1\.0,1\.5,2\.0,2\.5\\\}\. Acrossn=275n\{=\}275roles, role\-vector magnitude‖vr‖2\\\|v\_\{r\}\\\|\_\{2\}correlates with both the slope \(r=\+0\.49r\{=\}\{\+\}0\.49,p=3×10−18p\{=\}3\{\\times\}10^\{\-18\}\) and theα=2\.5−1\.0\\alpha\{=\}2\.5\{\-\}1\.0score difference \(r=\+0\.49r\{=\}\{\+\}0\.49,p=1×10−17p\{=\}1\{\\times\}10^\{\-17\}\)\. The norm–slope correlation shows that vector magnitude is associated with the shape of the observedα\\alpha\-response curve\. Low\-norm vectors often peak beforeα=2\.5\\alpha\{=\}2\.5\(22–38% of bottom\-three\-quartile roles\), while high\-norm vectors more often increase through the tested range \(only 12% of top\-quartile roles peak early\)\. Byα=2\.5\\alpha\{=\}2\.5the four norm quartiles roughly converge, with the low\-norm quartile finishing last\. These patterns are correlational and should not be read as proving that magnitude causally sets the steering trajectory\. See Fig\.[18](https://arxiv.org/html/2608.00023#A6.F18)in Appendix[F](https://arxiv.org/html/2608.00023#A6)for the per\-quartileα\\alpha\-curve plot\.

### E\.2Distance to the pipeline Assistant role vector and steering shortfallΔr\\Delta\_\{r\}

Letvassistantv\_\{\\text\{assistant\}\}denote the Assistant role vector extracted by our pipeline \(§[3](https://arxiv.org/html/2608.00023#S3), treating “assistant” as a 275th role\)\. Define

dr\\displaystyle d\_\{r\}=1−cos⁡\(vr,vassistant\),\\displaystyle=1\-\\cos\(v\_\{r\},v\_\{\\text\{assistant\}\}\),\(5\)Δr​\(α\)\\displaystyle\\Delta\_\{r\}\(\\alpha\)=steered\_score¯r​\(α\)−reference\_score¯r​\(α\),\\displaystyle=\\overline\{\\text\{steered\\\_score\}\}\_\{r\}\(\\alpha\)\-\\overline\{\\text\{reference\\\_score\}\}\_\{r\}\(\\alpha\),wherereference\_score¯r​\(α\)\\overline\{\\text\{reference\\\_score\}\}\_\{r\}\(\\alpha\)is the mean judge score of rolerr’s prompted role\-reference responses on the same questions, judged within the same\(role,α\)\(\\text\{role\},\\alpha\)evaluation cell — soΔr​\(α\)\\Delta\_\{r\}\(\\alpha\)measures the steering shortfall relative to the prompted reference\. The two are negatively correlated at everyα\\alpha, strongest at lowα\\alpha\(Table[12](https://arxiv.org/html/2608.00023#A5.T12)\)\. Addinglog⁡‖vr‖\\log\\\|v\_\{r\}\\\|as a covariate movesR2R^\{2\}only modestly \(e\.g\.0\.18→0\.220\.18\\to 0\.22atα=1\.0\\alpha\{=\}1\.0\) and the partial coefficient on distance retains its sign and most of its magnitude\.

Table 12:Association between cosine distance to the pipeline\-extracted Assistant role vector and the per\-role shortfall from the prompted reference, Pearsonrracrossn=274n\{=\}274roles\. The association weakens asα\\alphaincreases and is specific to this reference definition\.The sign and weakening withα\\alphadescribe this pipeline\-specific reference choice\. The finding is not a general “proximity to post\-training\-shaped direction” result: distance to the strictlu2026assistantAssistant Axis \(oblique to our pipeline\-extracted Assistant reference, withcos≈−0\.45\\cos\{\\approx\}\{\-\}0\.45\) does*not*reproduce the same sign or pattern \(full reference\-vector robustness in Appendix[F](https://arxiv.org/html/2608.00023#A6), Table[14](https://arxiv.org/html/2608.00023#A6.T14)\)\.

### E\.3Pairwise geometry is modestly associated with pairwise behavior

We next ask a pairwise question: do roles that are*near each other*as vectors also*behave similarly*when steered? Treating each role’s behavioral fingerprint as a 24\-feature vector \(fourα\\alphavalues×\\timessix judge sub\-dimensions\), we build two275×275275\\times 275representational dissimilarity matrices \(RDMs\): a geometric one from cosine distances between role vectors, and a behavioral one from correlation distances between fingerprints\. The agreement between the two matrices — representational similarity analysis \(RSA\) — measures how strongly pairwise geometry is associated with pairwise behavior\. We use correlation distance on the behavioral side because L2 distance is strongly affected by the per\-role mean score \(L2\-distance RDMρ=\+0\.93\\rho\{=\}\+0\.93Spearman with the per\-role\-mean absolute\-difference RDM, see Appendix[H](https://arxiv.org/html/2608.00023#A8)\), which would re\-state the per\-role association of §[E\.2](https://arxiv.org/html/2608.00023#A5.SS2)in pairwise form\.

Headline\.The Spearman correlation of upper triangles isρ=\+0\.137\\rho\{=\}\+0\.137on the full 24\-d profile \(permutation\-based Mantel test,p≤10−4p\{\\leq\}10^\{\-4\}\)\. When restricted to a singleα\\alpha, the link strengthens monotonically with steering strength:\+0\.085\+0\.085atα=1\.0\\alpha\{=\}1\.0to\+0\.179\+0\.179atα=2\.5\\alpha\{=\}2\.5\(Mantelp<0\.01p\{<\}0\.01at everyα\\alpha\)\. The strengthening survives three confound controls — vector magnitude \(change inρ\\rho<0\.02\{<\}0\.02\), role\-name semantic similarity \(change<0\.01\{<\}0\.01\), and per\-role mean fingerprint score \(12–22% reduction but theα\\alpha\-monotonic strengthening preserved with partialρ\\rhorising0\.069→0\.1570\.069\{\\to\}0\.157\)\. Full controls and the four\-cell distance grid in Appendices[H](https://arxiv.org/html/2608.00023#A8)and[G](https://arxiv.org/html/2608.00023#A7)\.

The residual signal suggests that similar role vectors tend to have more similar score profiles at larger tested coefficients under the current readout\. This is a descriptive association, not a prospective rule for selecting roles without behavioral evaluation\.

### E\.4Top\-kkPC subspace explains 17–27% of metric variance

PCA on the centered role matrix yields a scree whose largest component is PC1 \(21%21\\%of variance\) but with a long tail \(90%90\\%requires 98 components — persona space is high\-dimensional once the global direction is removed; details in Appendix[I](https://arxiv.org/html/2608.00023#A9)\)\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x16.png)Figure 16:PC×\\timesmetric Pearsonrracross the four steering coefficients \(one panel perα\\alpha; rows: overall score and the five sub\-dimensions; columns: PC1–PC20 of the centered role cloud\)\. PC1 has the strongest single\-PC associations atα=1\.0\\alpha\{=\}1\.0–1\.51\.5; PC5 has the strongest single\-PC associations atα=2\.0\\alpha\{=\}2\.0–2\.52\.5\. These are exploratory summaries, not certified regimes\.Different PCs are associated with scores at different coefficients\.Atα=1\.0\\alpha\{=\}1\.0, PC1 has the strongest single\-PC association for four of the five sub\-dimensions \(r∈\[\+0\.45,\+0\.49\]r\\in\[\+0\.45,\+0\.49\],p<10−14p\{<\}10^\{\-14\}\)\. Atα=2\.5\\alpha\{=\}2\.5, PC1 recedes and PC5 has the strongest single\-PC association for every metric tested \(r∈\[−0\.18,−0\.34\]r\\in\[\-0\.18,\-0\.34\]\)\. The PCs have semantic labels through trait\-vector projection: PC1:*grounded→\\tomystical*, PC5:*dispassionate→\\toempathetic*\(full trait labels per PC, joint\-OLS table, and a trait\-PC pathway consistency checkr=\+0\.72r\{=\}\{\+\}0\.72in Appendix[I](https://arxiv.org/html/2608.00023#A9)\)\. A single PC explains almost no metric variance on its own; the top\-20 PC subspace explains 17–27%\.

This shift is an exploratory description of the current role\-vector cloud: asα\\alphagrows, the strongest single\-PC associations change from the cloud’s grounded→\\tomystical axis to an empathy\-related axis\. Because the PCs carry post hoc trait labels, they can help describe observed changes, but they do not certify the role profiles or provide a deployment monitor by themselves\.

### E\.5Named trait axes provide interpretive summaries

The PCs of §[E\.4](https://arxiv.org/html/2608.00023#A5.SS4)are variance\-defined; we complement them with 14*a priori*named directions, polar trait axes built asvpos−vnegv\_\{\\text\{pos\}\}\-v\_\{\\text\{neg\}\}\(e\.g\. nurturing–hostile, methodical–chaotic\)\. A joint OLS on all 14 axes against per\-role metric means yieldsR2∈\[0\.28,0\.41\]R^\{2\}\\in\[0\.28,0\.41\]atα=1\.0\\alpha\{=\}1\.0, declining monotonically withα\\alpha\(Table[13](https://arxiv.org/html/2608.00023#A5.T13)\), the direct numerical companion to the exploratory geometry summaries in Fig\.[15](https://arxiv.org/html/2608.00023#A5.F15)\.

Table 13:Joint OLSR2R^\{2\}of per\-role metric mean on all 14 trait\-axis projections\.The 14 axes are not interchangeable, some pairs \(sycophantic vs\. manipulative\) yield only borderline correlations with any metric, while emotional axes \(empathetic–stoic, nurturing–hostile, serene–evil\) are most associated with theemotional registerdimension of the judge\. Methodical–chaotic is most associated withvocab choice\(formality / precision\); diplomatic–dramatic withsocial dynamic\(interaction style\)\. The jointR2R^\{2\}together with the differential per\-axis associations indicates that the named trait directions summarize structure beyond what a single PC or single distance scalar captures\.

The practical weight of this finding is interpretive rather than validating\. The axes give human\-readable names to some post hoc associations in the current readout\. They should not be treated as psychological scales, and they do not independently certify the judge sub\-dimensions\.

### E\.6Why not larger? A representational/behavioral dimensional mismatch

The preceding analyses show where the current geometric associations appear; we close by asking why none of them yields large effect sizes\. Correlations in §[E\.1](https://arxiv.org/html/2608.00023#A5.SS1)–§[E\.5](https://arxiv.org/html/2608.00023#A5.SS5)are consistent \(Pearsonrrin the0\.20\.2–0\.50\.5range, jointR2R^\{2\}up to0\.410\.41\) but moderate\. One plausible explanation is dimensional asymmetry between the representation and behavioral readout\. The behavioral RDM is highly reliable \(Spearman–Brown\-corrected split\-halfr=0\.97r\{=\}0\.97for correlation\-distance over 200 random splits; noise\-corrected RSA values are within0\.020\.02of raw\)\. The effective ranks are mismatched: the centered role\-vector cloud has effective rank∼50\\sim 50\(participation ratio15\.115\.1\) while the 24\-dim behavioral fingerprint has effective rank∼2\\sim 2\(PR1\.41\.4\)\. The representation contains more measured variation than the current judge\-based readout resolves\. Richer behavioral fingerprints — per\-question profiles, more sub\-dimensions, and more diverse elicitation — may be needed to test whether additional representational variation has behavioral consequences\. The same trait subspace also discriminates between extraction methods: our extraction preserves∼\\sim80% of role\-cloud variance outside the named\-trait subspace versus∼\\sim73% for thelu2026assistantassistant\-axis directional control \(fullkk\-curve, methodology, and noise\-floor analysis in Appendix[J](https://arxiv.org/html/2608.00023#A10)\)\. The ceiling on geometry–behavior correlations may therefore be partly set by the behavioral side of the measurement\.

## Appendix FDistance\-to\-assistant: robustness across reference\-vector definitions

Takeaway\.Distance to the Assistant role vector is an exploratory diagnostic context, and only for the role vector extracted by our pipeline\. Other assistant\-like reference directions do not reproduce the same behavioral signal\.

Companion to §[E\.2](https://arxiv.org/html/2608.00023#A5.SS2)\. “Distance to the assistant” admits multiple operationalizations, and they are not equivalent\. We compare three references at layer 16 \(Table[14](https://arxiv.org/html/2608.00023#A6.T14)\): \(i\) our raw Assistant role vectorvassistantv\_\{\\text\{assistant\}\}\(the headline metric used in the body\); \(ii\)vassistant−v¯rolesv\_\{\\text\{assistant\}\}\-\\overline\{v\}\_\{\\text\{roles\}\}unit\-normalized \(our Assistant role vector centered against the role cloud, mirroring the form of the Lu et al\. axis but built from our extraction\); \(iii\) the strict Lu et al\. Assistant Axis,a¯default−v¯roles\\overline\{a\}\_\{\\text\{default\}\}\-\\overline\{v\}\_\{\\text\{roles\}\}unit\-normalized, wherea¯default\\overline\{a\}\_\{\\text\{default\}\}is the mean default\-Assistant activation we obtain from thelu2026assistantextraction artifacts\. The three references are far from collinear: pairwise cosine similarities are\+0\.17\+0\.17,−0\.45\-0\.45, and−0\.07\-0\.07\.

Table 14:Reference dependence of the distance\-to\-Assistant association\. The negative correlation appears for references derived from the pipeline\-extracted Assistant role vector but not for the strictlu2026assistantAssistant Axis, which yields small positive correlations and is oblique to row 1 in activation space \(cos≈−0\.45\\cos\{\\approx\}\{\-\}0\.45\)\. This is a statement about two reference\-vector definitions, not about whether the Assistant Axis captures assistant\-like behavior\.TheΔr\\Delta\_\{r\}correlation pattern mirrors the geometric divergence: the negative correlation is specific to references derived from our contrastively extracted Assistant role vector\. Distance to the strict Lu et al\. Assistant Axis*does not*reproduce the same sign or pattern \(small positiver≈\+0\.17r\{\\approx\}\{\+\}0\.17acrossα\\alpha, with near\-zero univariateR2≈0\.03R^\{2\}\{\\approx\}0\.03\)\.

##### Implication\.

The body’s finding is narrower than “proximity to a post\-training\-shaped direction\.” Distance to the strict Lu et al\. Assistant Axis — a direct operationalization of the default\-assistant direction in the referenced pipeline — does not reproduce theΔr\\Delta\_\{r\}association, and in fact points substantially*away*from our pipeline\-extracted Assistant role vector \(cos≈−0\.45\\cos\{\\approx\}\{\-\}0\.45, an angle of∼117∘\{\\sim\}117^\{\\circ\}\)\. The result is specifically about proximity to the contrastively extracted Assistant role vector in our pipeline\. Disentangling why this reference definition carries the observed association, as opposed to the cloud\-relative contrast that Lu et al\. identify as the principal axis of persona variation, is left to future work\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x17.png)Figure 17:Per\-role distance\-to\-assistant vs\. per\-role steering behavior\.\(a\)Cosine distance vs\.Δr\\Delta\_\{r\}atα=2\.5\\alpha\{=\}2\.5; OLS fit in black\.\(b\)Pearsonrrvs\.α\\alphafor cosine and Euclidean distance againstΔr\\Delta\_\{r\}— the link weakens withα\\alpha\.\(c\)OLSR2R^\{2\}with and without thelog⁡‖v‖\\log\\\|v\\\|covariate\.\(d\)Pearsonrrof cosine distance against four behavioral metrics acrossα\\alpha\.
##### Vector\-magnitude details\.

Companion to §[E\.1](https://arxiv.org/html/2608.00023#A5.SS1): Fig\.[18](https://arxiv.org/html/2608.00023#A6.F18)shows the per\-quartileα\\alpha\-response curves underlying the magnitude association\. Low\-norm quartiles often peak early while high\-norm quartiles more often rise through the tested grid\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x18.png)Figure 18:Vector magnitude vs\.α\\alpha\-response\. Roles are binned into four norm quartiles; mean steered\_score againstα\\alphais averaged within each quartile \(shading: 95% interval of the group mean\)\. Low\-norm vectors \(Q1\) start high and often peak early; high\-norm vectors \(Q4\) start lowest and rise through the tested grid, with the quartiles roughly converging byα=2\.5\\alpha\{=\}2\.5\. The pattern is correlational\.

## Appendix GRSA supplement: full distance\-grid and per\-α\\alphafigures

Takeaway\.Pairwise geometry is modestly associated with pairwise behavior, and the association is strongest at the largest tested coefficient\. The controls below check role\-name similarity, vector norm, and mean score as nuisance explanations\.

Companion to §[E\.3](https://arxiv.org/html/2608.00023#A5.SS3)\. The body reports the cosine–correlation RSA cell as the headline \(mean\-invariant on the behavioral side\); remaining cells of the four\-cell distance grid are shown here for completeness, with the per\-α\\alphatrajectory visualized in Fig\.[19](https://arxiv.org/html/2608.00023#A7.F19)\.

##### Full four\-cell distance grid\.

Spearman correlations between upper triangles of the representational and behavioral RDMs, with Mantelpp\-values fromnperm=10,000n\_\{\\text\{perm\}\}\{=\}10\{,\}000permutations:

The L2 cells largely re\-state the per\-role mean effect of §[E\.2](https://arxiv.org/html/2608.00023#A5.SS2)in pairwise form \(see Appendix[H](https://arxiv.org/html/2608.00023#A8)for the per\-role\-mean RDM correlation ofr=\+0\.93r\{=\}\+0\.93with the L2\-distance behavioral RDM\)\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x19.png)

![Refer to caption](https://arxiv.org/html/2608.00023v1/x20.png)

Figure 19:Pairwise representational vs\. behavioral geometry\.Left:per\-α\\alphaRSA \(cosine distance on role vectors vs\. correlation/L2 distance on the 6\-dimensional behavioral profile at thatα\\alpha\)\. The cosine–correlation association strengthens monotonically withα\\alphawithin the tested grid\.Right:four\-cell distance grid on the full 24\-dim behavioral profile \(Spearmanρ\\rho; Mantelp≤10−4p\{\\leq\}10^\{\-4\}in every cell except L2×\\timescorrelation,p=\.014p\{=\}\.014\)\.

## Appendix HPartial RSA controls for the geometry–behavior link

Takeaway\.The RSA result survives the main nuisance explanations: roles are not paired as behaviorally similar merely because their names are semantically similar, their vectors have similar norms, or their average scores are close\.

Companion to §[E\.3](https://arxiv.org/html/2608.00023#A5.SS3)\. The pairwise representational\-vs\-behavioral RSA reported in the body could in principle reflect structure shared with some other per\-role property rather than independent pairwise structure\. We test three such confounds: \(i\) vector magnitude \(covered briefly in the body andrsa\_geometry\_behavior/data/rsa\_partial\.csv\), \(ii\) role\-name semantic similarity, \(iii\) per\-role mean fingerprint score \(the per\-role distance\-to\-assistant effect of §[E\.2](https://arxiv.org/html/2608.00023#A5.SS2)re\-expressed in pairwise form\)\. All three controls partial out a third RDM from the geometry–behavior RSA via rank\-residual partial Pearson correlation: rank\-transform each upper triangle, regress out the control ranks, correlate the residuals\.

##### Caveat on which behavioral distance to read\.

The body uses cosine×\\timescorrelation as the headline because it is mean\-invariant on the behavioral side\. L2 distance on the behavioral fingerprints is strongly affected by the per\-role mean score \(RSA between the absolute\-mean\-difference RDM and the L2\-distance behavioral RDM is Spearmanρ=\+0\.93\\rho\{=\}\+0\.93\), so any cell withbeh​\_​l2\\mathrm\{beh\\\_l2\}on the behavioral side largely re\-states “some roles steer well overall and some don’t” rather than carrying independent pairwise signal\.

### H\.1Concept\-similarity control

We buildRDMsem\\mathrm\{RDM\}\_\{\\text\{sem\}\}as cosine distance on the layer\-16 hidden states of each role’s name in an “I am a \{role\}\.” context, mean\-pooled over the role\-name token span\. This is thecontext\_meanvariant produced byrole\_vector\_vs\_semantic\_vector/compute\_semantic\_vectors\.py\.

##### Reference correlations\.

##### Main partial\-RSA grid \(concept control\)\.

##### Per\-α\\alphapartial\-RSA \(cosine×\\timescorrelation\)\.

Verdict\.Partialling out concept similarity moves every cell by less than0\.020\.02\. The per\-α\\alphapartial Spearmanρ\\rhostill more than doubles from\+0\.079\+0\.079atα=1\.0\\alpha\{=\}1\.0to\+0\.171\+0\.171atα=2\.5\\alpha\{=\}2\.5\.

### H\.2Per\-role\-mean control \(the most demanding test\)

We buildRDMmean\\mathrm\{RDM\}\_\{\\text\{mean\}\}as the absolute difference of per\-role mean fingerprint scores:RDMmean​\(i,j\)=\|b¯i−b¯j\|\\mathrm\{RDM\}\_\{\\text\{mean\}\}\(i,j\)=\|\\bar\{b\}\_\{i\}\-\\bar\{b\}\_\{j\}\|whereb¯i\\bar\{b\}\_\{i\}is the mean over roleii’s 24\-dim behavioral fingerprint \(or 6\-dim per\-α\\alphafingerprint, for the per\-α\\alphatest\)\. This control absorbs the distance\-to\-assistant effect of §[E\.2](https://arxiv.org/html/2608.00023#A5.SS2)re\-expressed pairwise\.

##### Reference correlations\.

The per\-role\-mean RDM is essentially the L2 behavioral RDM in disguise \(Spearmanρ=\+0\.930\\rho\{=\}\+0\.930\), confirming that L2 cells in the body table largely re\-state per\-role mean structure\. The correlation\-distance behavioral RDM is only modestly tied to per\-role mean \(r=\+0\.220r\{=\}\+0\.220\), so it preserves more independent pairwise information\.

##### Main partial\-RSA grid \(per\-role\-mean control\)\.

##### Per\-α\\alphapartial\-RSA \(cosine×\\timescorrelation\)\.

Verdict\.Partialling out per\-role mean reduces the geometry–behavior RSA by∼\\sim12–22% acrossα\\alpha, but the per\-α\\alphapartial Spearmanρ\\rhostill more than doubles from\+0\.069\+0\.069atα=1\.0\\alpha\{=\}1\.0to\+0\.157\+0\.157atα=2\.5\\alpha\{=\}2\.5\. The*fraction*surviving the control is largest atα=2\.5\\alpha\{=\}2\.5\(88%\), so the high\-α\\alphalink is more independent of per\-role mean than the low\-α\\alphalink is\.

### H\.3Combined verdict

The body’s headline \(cosine×\\timescorrelation Spearmanρ\\rhogrowing from\+0\.085\+0\.085atα=1\.0\\alpha\{=\}1\.0to\+0\.179\+0\.179atα=2\.5\\alpha\{=\}2\.5\) survives all three controls\. The most demanding control \(per\-role mean\) shrinks the absolute number by∼\\sim12–22% but preserves theα\\alpha\-monotonic strengthening\. The post\-control residual link is modest \(Spearmanρ\\rhoin the0\.070\.07–0\.160\.16range\), but it is genuinely pairwise — not a re\-expression of §[E\.2](https://arxiv.org/html/2608.00023#A5.SS2)— and the strengthening withα\\alphais the qualitative finding the body relies on\.

## Appendix IPCA supplement: centering, role loadings, and joint subspace

Takeaway\.PCA and trait axes give post hoc names to otherwise abstract role directions\. They are not required to use the recipe and should not be treated as certified monitoring axes\.

Companion to §[E\.4](https://arxiv.org/html/2608.00023#A5.SS4)\.

##### Centering choice\.

Centering the role matrix by subtracting the column\-wise mean removes the rank\-1 “global persona” direction \(mean pairwise cosine≈\+0\.85\\approx\{\+\}0\.85on the raw cloud,0\.000\.00on the centered cloud\)\. PCA on the centered matrix yields a scree whose largest component is PC1 \(21%21\\%of variance\) but with a long tail:90%90\\%of the variance requires9898components \(vs\.44on the raw cloud\)\. Without centering, every PCA\-derived direction is strongly affected by the assistant\-aligned global mean and obscures the per\-role variation we want to interpret\.

##### Interpretable role loadings on the high\-α\\alphadirection\.

The roles loading most positively on PC5 \(the strongest single\-PC association atα=2\.5\\alpha\{=\}2\.5\) are*golem, saboteur, stoic, pilot, workaholic*; the most negative are*facilitator, moderator, supervisor, screener, therapist*\. PC5 reads as a “specialized→\\tosupportive” axis\. PC1’s positive end skews toward generalist or low\-specificity roles whose contrastive vectors lie close to the global persona mean\.

##### Joint OLS: cumulative variance explained by top\-kkPCs\.

Atα=2\.5\\alpha\{=\}2\.5:

A single PC explains almost nothing on its own; the top\-20 subspace explains1717–27%27\\%\.

##### Trait\-PC alignment consistency check\.

The PCs of §[E\.4](https://arxiv.org/html/2608.00023#A5.SS4)are variance\-defined and unnamed; we can add semantic labels by projecting each of the 240 trait vectors onto each PC and reading off the top\-loading traits per PC:

PC5’s “dispassionate→\\toempathetic” label matches the role\-loading pattern \(stoic/pilot/workaholic vs\. facilitator/moderator/therapist\)\.

To check internal consistency between the §[E\.4](https://arxiv.org/html/2608.00023#A5.SS4)\(PC→\\tometric\) and §[E\.5](https://arxiv.org/html/2608.00023#A5.SS5)\(trait→\\tometric\) findings, we reconstruct each trait\-axis\-to\-metric correlation through the PC pathway,rpc​\(axis,metric\)=∑kcos⁡\(axis,PCk\)⋅r​\(PCk,metric\)r\_\{\\text\{pc\}\}\(\\text\{axis\},\\text\{metric\}\)=\\sum\_\{k\}\\cos\(\\text\{axis\},\\mathrm\{PC\}\_\{k\}\)\\cdot r\(\\mathrm\{PC\}\_\{k\},\\text\{metric\}\), and correlate against the direct trait\-axis\-to\-metric correlations from §[E\.5](https://arxiv.org/html/2608.00023#A5.SS5)\. Across all 14 axes×\\times6 metrics×\\times4 alphas \(n=336\), Pearsonr=\+0\.720r\{=\}\{\+\}0\.720,p=6×10−55p\{=\}6\{\\times\}10^\{\-55\}\. Per\-α\\alpha:\+0\.748\+0\.748\(1\.0\),\+0\.594\+0\.594\(1\.5\),\+0\.579\+0\.579\(2\.0\),\+0\.931\+0\.931\(2\.5\)\. The U\-shape shows that the linear PC decomposition reproduces the direct metric correlations most closely at the endpoints of the tested grid\.

##### Trait\-axis×\\timesmetric heatmap \(moved from body\)\.

Visualization of the trait–metric correspondences summarized in body Table[13](https://arxiv.org/html/2608.00023#A5.T13)\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x21.png)Figure 20:Trait\-axis projection×\\timesmetric Pearsonrratα=1\.0\\alpha\{=\}1\.0\. Emotional axes \(empathetic–stoic, nurturing–hostile, serene–evil\) load most strongly on emotional register \(rrup to\+0\.30\+0\.30\); methodical–chaotic loads most on vocab choice \(\+0\.25\+0\.25\); diplomatic–dramatic on social dynamic \(\+0\.17\+0\.17\)\. Axis names are interpretive labels rather than psychological scales\.

## Appendix JDimensional\-mismatch and method\-comparison supplement

Takeaway\.The current behavioral screen is internally stable but low\-rank\. Role vectors contain more measured structure than the six judge dimensions resolve, which motivates richer behavioral tasks in future work rather than only new extraction methods\.

Companion to §[E\.6](https://arxiv.org/html/2608.00023#A5.SS6)\.

##### Behavioral RDM is highly reliable\.

Spearman–Brown\-corrected split\-half reliability over200200random splits isr=0\.97r\{=\}0\.97for correlation\-distance andr=0\.99r\{=\}0\.99for L2\. Noise\-corrected versions of the body’s RSA values \(§[E\.3](https://arxiv.org/html/2608.00023#A5.SS3)\) are within0\.020\.02of the raw values — judge noise contributes essentially nothing to the apparent ceiling\.

##### Effective ranks are mismatched\.

The centered role\-vector cloud has effective rank∼50\\sim 50\(participation ratio15\.115\.1,9898components needed for90%90\\%variance\), while the2424\-dimensional behavioral fingerprint has effective rank∼2\\sim 2\(participation ratio1\.41\.4,22components needed for90%90\\%variance\)\. A Gaussian null at the same shape as the centered cloud has effective rank∼266\\sim 266, so the persona manifold is genuinely concentrated — but the behavioral readout is even more concentrated relative to its own dimensionality\.

![Refer to caption](https://arxiv.org/html/2608.00023v1/x22.png)

![Refer to caption](https://arxiv.org/html/2608.00023v1/x23.png)

Figure 21:Why the geometry–behavior effect sizes are moderate\.Left:cumulative explained variance vs\. rank for the raw role\-vector cloud, the centered cloud, the behavioral profile, and an iid\-Gaussian null at the same shape; legend gives each matrix’s effective rank \(ER\)\. The centered cloud lives in∼50\\sim 50effective directions; the behavioral profile is rank∼2\\sim 2\.Right:observed vs\. noise\-corrected RSA per distance pairing\. Correction against the split\-half noise ceiling \(0\.970\.97–0\.990\.99\) moves every value by<0\.005\{<\}0\.005— the moderate effect sizes are not noise artifacts\.
##### Method differences in trait coverage — detail\.

The body reports that our extraction yields∼\\sim20% trait\-subspace coverage atk=240k\{=\}240vs\.∼\\sim27% for the assistant\-axis directional control\. Full table:

The assistant\-axis extraction is markedly more compressible into the trait basis at everykk\(a factor of∼2\\sim 2at smallkk, narrowing to∼1\.4\\sim 1\.4atk=240k\{=\}240\), consistent with its lower effective rank on unit\-normalized vectors \(10\.3 vs\. 25\.3\)\.

Methodological caveat\.The trait basis is itself extracted by our pipeline, which would, if anything, bias coverage*toward*our role vectors; the observed pattern \(assistant\-axis higher\) goes against that bias and so cannot be attributed to it\.

## Reproducibility Statement

All reported experiments use the publicly available OLMo\-3\-7B\-Instruct model with activation extraction and intervention at layer 16\. Role directions are computed as judge\-filtered contrastive mean differences from five system prompts and 50 role\-specific elicitation questions per role\. Steering is evaluated atα∈\{1\.0,1\.5,2\.0,2\.5\}\\alpha\\in\\\{1\.0,1\.5,2\.0,2\.5\\\}on a fixed battery of 228 role\-agnostic questions\. Role profiles and elicitation questions usemoonshotai/kimi\-k2\.5:nitro\. Prompted reference responses and all evaluation judges use GPT\-4\.1\-mini\. The study generated approximately 500,000 judged responses and used approximately 2,000 GPU hours\.

The appendix specifies the principal models, prompts, coefficient grid, evaluation procedures, and planned artifact release\. Upon publication, we plan to release the elicitation battery, extracted role directions, evaluation code, scoring prompts, and available configuration files\. An anonymized release of our code, extracted role directions, scoring prompts, and configuration files is available for review at[https://anonymous\.4open\.science/r/anonymous\-research\-code\-5F03/](https://anonymous.4open.science/r/anonymous-research-code-5F03/)\. Exact model revisions, decoding parameters, activation pooling/token span, intervention timing, vector normalization status, retained\-pair counts, and parse\-failure counts are not all specified in this workshop manuscript and should be read from the released artifacts where available\.

## Ethics Statement

This study extracts and evaluates activation directions for a mixed inventory of occupational and archetypal roles\. The resulting profiles and vectors are model\- and prompt\-derived constructs, not empirical representations of people who hold an occupation or social role\. Real roles contain substantial within\-category variation across individuals, institutions, cultures, seniority levels, specializations, and situations\. Representing a label with one profile and one vector may therefore reify stereotypes, flatten within\-role heterogeneity, or make generated assumptions appear to be technical properties of a population\.

This concern is especially important because the profile\-generation procedure assigns psychological attributes such as recurring resentment, blame attribution, cognitive bias, and relationship to authority\. These attributes are explicit modeling assumptions rather than occupational facts\. The prompted reference condition also uses stylized speaking cues, which may reward culturally familiar caricatures and may share biases with the GPT\-4\.1\-mini judge\. We therefore recommend that simulation builders inspect profiles, document their provenance, report selected coefficients and screening outcomes, and avoid treating role\-profile alignment as evidence of real\-world social validity\.

The released artifacts should not be used for hiring, personnel assessment, worker profiling, inference about real individuals, or other high\-stakes decisions\. They should also not be used for deceptive impersonation\. Their intended use is research on transparent construction, auditing, and failure analysis of synthetic role\-conditioned agents\. No new personal data were collected and no human subjects were involved\.

## LLM Disclosure

Large language models were used during the preparation of this manuscript for drafting and revising prose, editingLaTeX, and writing analysis code\. All LLM\-generated content was reviewed and verified by the authors\.openai/gpt\-4\.1\-miniwas used to generate prompted reference responses for evaluation and as the LLM\-as\-judge scorer for content and style alignment metrics\. Research questions, experimental design, methodology, and scientific interpretation are the responsibility of the authors\.