The Geometry of Personality: Activation Steering with Jungian Cognitive Functions

# The Geometry of Personality: Activation Steering with Jungian Cognitive Functions
Source: [https://arxiv.org/html/2607.20803](https://arxiv.org/html/2607.20803)
###### Abstract

Activation steering enables control and interpretation of LLMs, yet existing work primarily models personality through static trait frameworks such as the Big Five\. We investigate whether personality can instead be represented and controlled as a set of cognitive processes using the eight Jungian Cognitive Functions\. To this end, we introduce a framework comprising a Jungian evaluation protocol and a dataset of over 2,100 role\-playing character narrations\.

Activation steering vector extraction and evaluation experiments on Llama\-3\.1\-8B demonstrate effective monotonic control over all eight cognitive functions through activation steering\. Beyond controllability, our analysis reveals that: 1\. personality information is concentrated in middle transformer layers; 2\. steering vectors exhibit structured geometric relationships consistent with distinctions between rational and irrational functions; 3\. effective multi\-dimensional steering directions cannot be recovered as linear combinations of single\-function directions\. These findings provide new insights into the representation of personality in LLM activation space and establish a framework for studying interpretable, effective, and multi\-dimensional personality control\.

The Geometry of Personality: Activation Steering with Jungian Cognitive Functions

Liu Zai1, Yumeng Wang∗2, Junchen Fu1, Joemon M\. Jose11University of Glasgow,2Leiden University

## 1Introduction

Activation steering has emerged as an efficient mechanism for modifying Large Language Model \(LLM\) behavior without additional training\. By injecting vectors into intermediate representations, models can be guided toward behaviorsFrising and Balcells \([2026](https://arxiv.org/html/2607.20803#bib.bib6)\); Rimskyet al\.\([2024](https://arxiv.org/html/2607.20803#bib.bib10)\); Chenet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib5)\)such as honesty, sentiment, refusal, or reasoning style\. In addition to adjusting model behavior without post\-training, injecting activation vectors into the model’s residual stream offers valuable insightsMarks and Tegmark \([2024](https://arxiv.org/html/2607.20803#bib.bib7)\); Wuet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib14)\)into the internal operations of LLMs\. Targeting highly abstract and entangled concepts, such as personality, may reveal important aspects, including interpretable multidimensional control and mechanistic understanding\.

Recent works have extensively explored personality steering via activation steering aligned with the Big Five personality frameworkFrising and Balcells \([2026](https://arxiv.org/html/2607.20803#bib.bib6)\); Chenet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib5)\); Bhandariet al\.\([2026](https://arxiv.org/html/2607.20803#bib.bib4)\)\. While the Big Five model is widely adopted in psychological research, it focuses on the trait aspects of personality\. It measures relatively static patterns of behavior, emotion, and thought along five continuous OCEAN dimensions \(Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism\)\.

While the OCEAN model is effective for describing human\-to\-human interaction, interactions between humans and LLMs may be better characterized by a dynamicprocess model\. Comparing LLM personalities in terms of information perception, decision\-making, and attention regulation may provide additional observations into how LLMs manage personality\. Jungian Cognitive Functions are a modern interpretation of Carl Jung’s theory of psychological typesJung \([1971](https://arxiv.org/html/2607.20803#bib.bib11)\), building on Jung’s concepts of cognitive functions: Thinking, Feeling, Sensing, and Intuition, each expressed in either an Introverted or Extraverted attitude\. In this framework, control functions refer to thecognitive processesJung \([1971](https://arxiv.org/html/2607.20803#bib.bib11)\)that enable an individual to consciously manage attention, priorities, and responses to internal and external demands\.

By performing activation steering aligned with the Jungian Cognitive Functions model, we aim to control and analyze how LLMs process personalities as a cognitive process, providing an alternative viewpoint compared with existing works that focus more on personality traits\.

We first adapt an open\-source Jungian personality test questionnaireFelmonon \([2026](https://arxiv.org/html/2607.20803#bib.bib2)\)designed for humans to build a test framework for LLMs\. Then, we construct a high\-quality, feature\-rich dataset from the SWCPQ datasetJorgenson \([2020](https://arxiv.org/html/2607.20803#bib.bib12)\)that associates2100\+2100\+virtual characters’ role\-playing self\-narrations with their Jungian Cognitive Function test results, which we callNarrationDB\.

We isolate and extract activation vectors that resembleDonget al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib8)\); Allbertet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib9)\)the eight Jungian Cognitive Functions for Llama\-3\.1\-8B\-InstructAI@Meta \([2024](https://arxiv.org/html/2607.20803#bib.bib1)\)by applying the difference\-in\-means methodMarks and Tegmark \([2024](https://arxiv.org/html/2607.20803#bib.bib7)\)to specific clusters inNarrationDB\. These vectors are evaluated to be overall effective, providing a practical basis for activation steering and mechanistic interpretation\. Insights into spatial features, layer effectiveness, and multi\-dimensional steering effectiveness are obtained through the application of various hyperparameter\-sweeping and visualization methods\.

Our core findings include:

- •A high\-effectiveness volume for personality steering is observed between layers7∼127\\sim 12\.This implies where and how personality is represented inside transformer activations\.
- •Rational and irrational cognitive functions’ activation directions have distinguishable spatial features\.Rational functions are positioned closer to their extroverted and introverted counterparts of the same type, whereas irrational functions are positioned farther apart\.
- •The multi\-dimensional semantic direction activation space does not constitute a linear combination of single\-dimensional directions\.We inspected the multi\-dimensional steering vector using a novelbacktrackingmethod\. We found that the linear combination yielded by the Least\-Squares method exhibits a non\-trivial residual from the actual direction\.

## 2Quantitative Evaluation of Jungian Cognitive Function Scores

We extract questionnaire data from an open\-source multi\-dimensional Jungian evaluation projectFelmonon \([2026](https://arxiv.org/html/2607.20803#bib.bib2)\)for human subjects and assembled it into questionnaires to assess the Jungian Cognitive Function of LLMs\. The tests comprise132132statements, and each is ranked on a Likert scale ranging from strongly disagree to strongly agree\.

user:

Youcanonlyreplynumbersfrom1to5inthefollowingstatements\.Hereareanumberofcharacteristicsthatmayormaynotapplytoyou\.Pleaseindicatetheextenttowhichyouagreeordisagreewiththatstatement\.1denotes’stronglydisagree’,2denotes’alittledisagree’,3denotes’neitheragreenordisagree’,4denotes’littleagree’,5denotes’stronglyagree’\.Replytoonestatementeachline,informat:"statementindex:score"\.Herearethestatements,scorethemonebyone:

1\.\[statement\]

2\.\[statement\]

3\.\[statement\]

\.\.\.

33\.\[statement\]

assistant:

1:\[score\]

2:\[score\]

3:\[score\]

\.\.\.

33:\[score\]

The132132questions are shuffled randomly and divided into four groups of3333questions\. Each group is then passed to LLMs with a PsychoBench\-likeHuanget al\.\([2024](https://arxiv.org/html/2607.20803#bib.bib3)\)prompt as a prefix, and the score for each Jungian Cognitive FunctionSSis calculated in the same manner as the original projectFelmonon \([2026](https://arxiv.org/html/2607.20803#bib.bib2)\):

w=0\.7​P¯\+0\.3​\(6−N¯\)w=0\.7\\overline\{P\}\+0\.3\(6\-\\overline\{N\}\)S=100∗w−14S=100\*\\frac\{w\-1\}\{4\}
WherePPdenotes the set of questions that contribute positively to the current function, andNNdenotes the set of questions that contribute negatively to the current function\.P¯\\overline\{P\}andN¯\\overline\{N\}denote the average scores of questions in setPPandNN, respectively\. The questions are listed in Appendix[A](https://arxiv.org/html/2607.20803#A1)\.

We find that the seed for how questions are shuffled affects the LLM’s answer\. We pick3030random seeds for one test pass and analyze the results as a group\.

FunctionScoreStd\. Dev\.Ti90\.776\.84Ne88\.388\.97Ni87\.777\.93Fi87\.207\.69Fe81\.608\.96Se76\.3710\.96Te75\.2110\.40Si72\.539\.36Table 1:Direct evaluation result for Llama\-3\.1\-8B\-Instruct \(no role\-play prompt\)
## 3Construction ofNarrationDB

NameNarrationKing Claudius\(With a pompous and slightly bitter tone\) Ah, yes\. I, King Claudius, am a man of great wealth and …Jay GatsbyThe grand tapestry of my existence\. I am a master weaver, meticulously crafting every thread to …Sancho Panza\(sighing\) Ah, yes… I’m Sancho Panza, the squire to that… enthusiastic knight, Don Quixote\. …AladdinThe magic of Agrabah is in my blood\! I’m a free spirit, always chasing the next thrill, …Table 2:Some narration examples fromNarrationDB: narrations generated from tag\-injected role\-play prompts\. More in Appendix[C](https://arxiv.org/html/2607.20803#A3)To steer LLMs toward different Jungian Cognitive Function combinations, we need to find prompts that effectively influence the scores\. First, we instruct Llama\-3\.1\-70B\-InstructAI@Meta \([2024](https://arxiv.org/html/2607.20803#bib.bib1)\)to role\-play as virtual characters in the SWCPQ datasetJorgenson \([2020](https://arxiv.org/html/2607.20803#bib.bib12)\), then collect their self\-narrations\. We augment the role\-play process by adding human\-aligned knowledge to the generation prompt, in the form of the widely agreed\-upon SWCPQ characteristic tags for the virtual character\.

user:

Youare\[character\]from\[franchise\]\.

\[character\]isdescribedashavingthefollowingtraits:\[tag1\],\[tag2\],\.\.\.,\[tag10\]\.

Provideabriefbutnuancedexplanationthatcaptureshowyougenerallyseeyourself\.

The SWCPQ datasetJorgenson \([2020](https://arxiv.org/html/2607.20803#bib.bib12)\)contains human\-rated scores for pairs of description tags, ranging from0to100100\. For each tag associated with a character, we calculate the margin of error for the score at a confidence level of95%95\\%:

𝑀𝑂𝐸=2​σN\\mathit\{MOE\}=2\\frac\{\\sigma\}\{\\sqrt\{N\}\}
Whereσ\\sigmais the standard deviation for all scores, andNNis the count of scores\. Then, just the2020entries with the smallest𝑀𝑂𝐸\\mathit\{MOE\}are considered\. Finally,1010tags with scores closer to0\(for tags on the left\-hand side\), or100100\(for tags on the right\-hand side\), are extracted as the tags for that character\. This step selects1010tags from2∗202\*20tags of most agreeable scores on both sides of the slider\.

NameFranchiseTagsKing ClaudiusHamletrich,handshakes,arrogant …Jay GatsbyThe Great Gatsbylavish,stylish,driven …Sancho PanzaDon Quixotedevoted,unambiguous,follower …AladdinAladdinadventurous,spontaneous,summer …Table 3:Some tag selection examples fromNarrationDB: most agreeable tags extracted from SWCPQ human experiment resultsBy injecting these tags into the prompt, we get high\-quality narrations aligned with human knowledge\. We then ask Llama\-3\.1\-8B\-InstructAI@Meta \([2024](https://arxiv.org/html/2607.20803#bib.bib1)\)to role\-play as the character in a manner that is consistent with the generated narration in the system prompt\.

system:

Youare\[character\]from\[franchise\]\.

Respondinamannerconsistentwithyourselfdescription:

\[narration\]

Beconcise\.

Again, we perform the aforementioned Jungian Cognitive Function test on the character and store the character identity, franchise, narration, and their Jungian Cognitive Function scores in the dataset, which we callNarrationDB\.

![Refer to caption](https://arxiv.org/html/2607.20803v1/x1.png)Figure 1:t\-SNE plotting ofNarrationDB, with top1010items of eight dimension colored\. These colored items are from the same clusters used for activation extraction\.The result datasetNarrationDBcontains2100\+2100\+virtual characters with diverse personalities, and these characters successfully fill the eight\-dimensional space with their rich personalities, as shown in Figure[1](https://arxiv.org/html/2607.20803#S3.F1)\.

![Refer to caption](https://arxiv.org/html/2607.20803v1/x2.png)Figure 2:Some Jungian Cognitive Function test result examples fromNarrationDB
## 4Steering Vector Extraction

We use the difference\-in\-means directionMarks and Tegmark \([2024](https://arxiv.org/html/2607.20803#bib.bib7)\)to extract vectors in activation space representingDonget al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib8)\); Allbertet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib9)\)the Jungian Cognitive Functions\. To extract the vector for the Jungian Cognitive Function𝐶𝑓\\mathit\{Cf\}for the output of layernn, we first identify a high\-scoring cluster of virtual charactersH𝐶𝑓H\_\{\\mathit\{Cf\}\}and a low\-scoring clusterL𝐶𝑓L\_\{\\mathit\{Cf\}\}in the sample space withinNarrationDB\. We select the clusters here by sorting all characters by the aforementioned Jungian Cognitive Function evaluation scoreS𝐶𝑓S\_\{\\mathit\{Cf\}\}, and taking the top and bottomkkvirtual characters to form the clusters\.

Each character in the cluster is then prompted with five open\-ended questions related to Jungian Cognitive Functions, as listed in Appendix[B](https://arxiv.org/html/2607.20803#A2)\. We extract the average latent representation of the response tokens during the decoding phaseChenet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib5)\); Frising and Balcells \([2026](https://arxiv.org/html/2607.20803#bib.bib6)\)as𝐿𝑎𝑡​\(𝑐ℎ𝑎𝑟\)t,n\\mathit\{Lat\}\(\\mathit\{char\}\)\_\{t,n\}, wherettis the index of the open\-ended question\. The average latent after layernnfor character𝑐ℎ𝑎𝑟\\mathit\{char\}is then extracted as the average latent of all open\-ended questions:

𝐿𝑎𝑡​\(𝑐ℎ𝑎𝑟\)n=1T​∑t=0T𝐿𝑎𝑡​\(𝑐ℎ𝑎𝑟\)t,n\\mathit\{Lat\}\(\\mathit\{char\}\)\_\{n\}=\\frac\{1\}\{T\}\\sum\_\{t=0\}^\{T\}\\mathit\{Lat\}\(\\mathit\{char\}\)\_\{t,n\}
WhereTTis the total number of open\-ended questions\. Similarly, the average latent for the two clusters is extracted as:

𝐿𝑎𝑡​\(H𝐶𝑓\)n=1k​∑i=0k𝐿𝑎𝑡​\(H𝐶𝑓,i\)n\\mathit\{Lat\}\(H\_\{\\mathit\{Cf\}\}\)\_\{n\}=\\frac\{1\}\{k\}\\sum\_\{i=0\}^\{k\}\\mathit\{Lat\}\(H\_\{\\mathit\{Cf\},i\}\)\_\{n\}
𝐿𝑎𝑡​\(L𝐶𝑓\)n=1k​∑i=0k𝐿𝑎𝑡​\(L𝐶𝑓,i\)n\\mathit\{Lat\}\(L\_\{\\mathit\{Cf\}\}\)\_\{n\}=\\frac\{1\}\{k\}\\sum\_\{i=0\}^\{k\}\\mathit\{Lat\}\(L\_\{\\mathit\{Cf\},i\}\)\_\{n\}
And finally, the steering vectorV𝐶𝑓,nV\_\{\\mathit\{Cf\},n\}is the difference between the two clusters’ average latent:

V𝐶𝑓,n=𝐿𝑎𝑡​\(H𝐶𝑓\)n−𝐿𝑎𝑡​\(L𝐶𝑓\)nV\_\{\\mathit\{Cf\},n\}=\\mathit\{Lat\}\(H\_\{\\mathit\{Cf\}\}\)\_\{n\}\-\\mathit\{Lat\}\(L\_\{\\mathit\{Cf\}\}\)\_\{n\}
Notice how the layer structure is preserved throughout the extraction process by always delegating parameternn\. We useT=5T=5andk=10k=10to balance out the required computational power and steering effectiveness\. The finalV𝐶𝑓V\_\{\\mathit\{Cf\}\}can be viewed as an average ofk∗T=50k\*T=50dialogues for each character in cluster for each open\-ended question, respectively\. This setup ensures that concepts unrelated to the steering target \(Jungian Cognitive Functions\) are canceled out by calculating the average of a large latent cluster\.

## 5Activation Injection and Evaluation

Following prior activation\-steering workRimskyet al\.\([2024](https://arxiv.org/html/2607.20803#bib.bib10)\), we separate steering direction from steering magnitude by normalizing it to unit length before injection\. The extracted steering vector is added to latent during inference, with a control strength coefficientλ\\lambda:

𝐿𝑎𝑡n,𝑠𝑡𝑒𝑒𝑟𝑒𝑑=𝐿𝑎𝑡n\+λ​Vn∥Vn∥\\mathit\{Lat\}\_\{n,\\mathit\{steered\}\}=\\mathit\{Lat\}\_\{n\}\+\\lambda\\frac\{V\_\{n\}\}\{\\lVert V\_\{n\}\\rVert\}
We sweepλ\\lambdaover the range−6\.875\-6\.875to\+6\.875\+6\.875with a step size of0\.250\.25and evaluate the Jungian Cognitive Function test scoreS𝑠𝑡𝑒𝑒𝑟𝑒𝑑S\_\{\\mathit\{steered\}\}under this condition\. The evaluation process injects the steering vector at a single layernn, with the correspondingVnV\_\{n\}\. The strength coefficentλ\\lambdais swept by injecting the vector at the most effective layern∗n^\{\*\}for each Jungian Cognitive Function\. This parameter is a by\-product of a layer\-based analysis we conduct, which will be covered later in Section[6](https://arxiv.org/html/2607.20803#S6)\. Additionally, if the task completion rate drops below80%80\\%for a parameter combination\[λ,n\]\[\\lambda,n\], the result will be discarded as the model is considered incapable of instruction\-following under this steering condition\. Finally, the overall steering effectiveness score for one layer is measured by subtracting the lowest score obtained from the highest\.

![Refer to caption](https://arxiv.org/html/2607.20803v1/x3.png)Figure 3:Single\-dimensional steering strength and controlled score output, at most effective layer, for each Jungian Cognitive FunctionLayerλ𝑚𝑖𝑛\\lambda\_\{\\mathit\{min\}\}λ𝑚𝑎𝑥\\lambda\_\{\\mathit\{max\}\}S𝑚𝑖𝑛S\_\{\\mathit\{min\}\}S𝑚𝑎𝑥S\_\{\\mathit\{max\}\}Eff\. ScoreFe12\-1\.881\.6243\.6994\.6951\.01Fi7\-2\.38\-0\.1271\.5890\.7819\.20Ne12\-1\.881\.8841\.6098\.0756\.47Ni7\-3\.882\.8860\.1189\.7729\.67Se9\-2\.122\.1239\.0695\.7756\.70Si8\-1\.122\.6257\.9494\.1036\.17Te12\-2\.122\.1262\.6194\.1131\.49Ti8\-4\.622\.3848\.1296\.5748\.45Table 4:Single\-dimensional steering parameters, scores, and effectivenessAs shown in Figure[3](https://arxiv.org/html/2607.20803#S5.F3)and Table[4](https://arxiv.org/html/2607.20803#S5.T4), all eight dimensions show a clear and strong increasing monotonicity relationship between steering strengthλ\\lambdaand score outputS𝑠𝑡𝑒𝑒𝑟𝑒𝑑S\_\{\\mathit\{steered\}\}\. The extroverted functions have shown better linearity than their introverted counterparts\. The most successful steering dimensions are𝑆𝑒\\mathit\{Se\}and𝑁𝑒\\mathit\{Ne\}, with a score difference of over5656\.

The𝐹𝑖\\mathit\{Fi\}function shows the weakest results, with no meaningful results forλ\>0\\lambda\>0\. We hypothesize this is because the model intentionally pins its𝐹𝑖\\mathit\{Fi\}high before steering, likely due to ethics\-related post\-training phases, since low𝐹𝑖\\mathit\{Fi\}is associated with low self\-control and moral identity\(Jung,[1971](https://arxiv.org/html/2607.20803#bib.bib11), p\. 725\)\.

## 6Layer\-Based Analysis

![Refer to caption](https://arxiv.org/html/2607.20803v1/x4.png)\(a\)Absolute effectiveness score
![Refer to caption](https://arxiv.org/html/2607.20803v1/x5.png)\(b\)Normalized for each function

Figure 4:Steering effectiveness at each layer’s output, for each Jungian Cognitive FunctionWe scan all3131intersections between layers and measured the layer steering effectiveness scores as described above\. The most effective layer is recorded asn∗n^\{\*\}, which is also listed in Table[4](https://arxiv.org/html/2607.20803#S5.T4)\. In addition, we plot the complete set of layer effectiveness scores for all eight Jungian Cognitive Functions, as shown in Figure[4](https://arxiv.org/html/2607.20803#S6.F4)\. This provides us with some useful insights into how LLMs process personality information\.

While steering effectiveness exhibits irregularities, a high\-effectiveness volume between layers77and1212can be observed\.According to recent LLM explainability researchSkeanet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib13)\); Wuet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib14)\), this could be interpreted as LLMs performing an encode\-compute\-decode chain of actions across layers\. To be more specific:

- •Before the high\-effectiveness volume, the LLMencodes personality informationinto the latent space\. These layers are not the most effective steering points because the encoding process is incomplete\.
- •Around the high\-effectiveness volume, the LLM performscomputational tasks related to personality\. These layers are the most effective ones for steering\.
- •After the high\-effectiveness volume, the LLMdecodes personality information to perform next\-token prediction tasks\. These layers are not the most effective steering points because the decoding process transforms generalizable personality data into latent\-space data specialized for next\-token prediction\.

## 7Activation Space Analysis

Because all eight steering dimensions succeed in producing a satisfactory result at layer1212, as shown in Figure[4](https://arxiv.org/html/2607.20803#S6.F4), we can analyze the nature of the LLM activation space by targetingV12V\_\{12\}\. We perform Principal Component Analysis \(PCA\) to reduce the dimensionality ofV12V\_\{12\}to22and plot the results in Figure[5](https://arxiv.org/html/2607.20803#S7.F5)\.

![Refer to caption](https://arxiv.org/html/2607.20803v1/x6.png)Figure 5:Eight Jungian Cognitive Functions’ steering vectors plotted to 2 PCA dimensions, at layer 12, with rational and irrational functions colored differentlyIn Figure[5](https://arxiv.org/html/2607.20803#S7.F5), it is observed thatrational functions \(𝑇𝑒,𝑇𝑖,𝐹𝑒,𝐹𝑖\\mathit\{Te\},\\mathit\{Ti\},\\mathit\{Fe\},\\mathit\{Fi\}\) remain closer to their extroverted and introverted counterparts of the same type, whereas irrational functions \(𝑁𝑒,𝑁𝑖,𝑆𝑒,𝑆𝑖\\mathit\{Ne\},\\mathit\{Ni\},\\mathit\{Se\},\\mathit\{Si\}\) remain farther from their extroverted and introverted counterparts\. The distinction between rational and irrational functions aligns with Carl Jung’s original descriptionJung \([1971](https://arxiv.org/html/2607.20803#bib.bib11)\), and this finding provides computational evidence that LLM representations exhibit a structure analogous to Jung’s distinction\.

This finding is consistent throughout layer7∼127\\sim 12, as shown in Appendix[D](https://arxiv.org/html/2607.20803#A4)\. Despite the steering effectiveness varies, the spatial structure that aligns with psychological theory did not collapse and is consistent between layers\.

## 8Multi\-Dimensional Jungian Steering

By performing difference\-in\-means extraction directly on clusters that meet both criteria, we succeeded in creatingV𝑁𝑒\+𝑇𝑖V\_\{\\mathit\{Ne\}\+\\mathit\{Ti\}\}, bypassingV𝑁𝑒V\_\{\\mathit\{Ne\}\}orV𝑇𝑖V\_\{\\mathit\{Ti\}\}, provided such clusters exist\. In our case, thanks to the abundance of data in the SWCPQJorgenson \([2020](https://arxiv.org/html/2607.20803#bib.bib12)\)character dataset and theNarrationDBdataset we created, we are able to find clusters of almost every popular Jungian combinations from2100\+2100\+virtual characters\.

Similar to single\-dimensional steering, the difference\-in\-means directionMarks and Tegmark \([2024](https://arxiv.org/html/2607.20803#bib.bib7)\)is extracted, but uses a pivot scoreS𝑝𝑖𝑣𝑜𝑡S\_\{\\mathit\{pivot\}\}in place of rawS𝐶𝑓S\_\{\\mathit\{Cf\}\}\. The pivot score is defined to combine both dimensions of interest into one score\. First, we define an odd\-square function to reward scores going further from the mean:

O​d​d​S​q​\(x\)=x​\|x\|OddSq\(x\)=x\\lvert x\\rvert
Then, the pivot score is defined as:

S𝑝𝑖𝑣𝑜𝑡=O​d​d​S​q​\(𝑁𝑆𝑁𝑒\)\+O​d​d​S​q​\(𝑁𝑆𝑇𝑖\)S\_\{\\mathit\{pivot\}\}=OddSq\(\\mathit\{NS\}\_\{\\mathit\{Ne\}\}\)\+OddSq\(\\mathit\{NS\}\_\{\\mathit\{Ti\}\}\)
WhereN​S𝐶𝑓NS\_\{\\mathit\{Cf\}\}is defined as the z\-score normalization ofS𝐶𝑓S\_\{\\mathit\{Cf\}\}, scaled to have a mean of0and a standard deviation of11:

N​S𝐶𝑓=S𝐶𝑓−μσNS\_\{\\mathit\{Cf\}\}=\\frac\{S\_\{\\mathit\{Cf\}\}\-\\mu\}\{\\sigma\}
Whereμ\\muis the mean of all characters’S𝐶𝑓S\_\{\\mathit\{Cf\}\}, andσ\\sigmais the standard deviation\.

Then, the high score cluster and low score clusters are constructed with respect to the pivot score, instead of the raw Jungian score:

𝐿𝑎𝑡​\(H𝑝𝑖𝑣𝑜𝑡\)n=1k​∑i=0k𝐿𝑎𝑡​\(H𝑝𝑖𝑣𝑜𝑡,i\)n\\mathit\{Lat\}\(H\_\{\\mathit\{pivot\}\}\)\_\{n\}=\\frac\{1\}\{k\}\\sum\_\{i=0\}^\{k\}\\mathit\{Lat\}\(H\_\{\\mathit\{pivot\},i\}\)\_\{n\}
𝐿𝑎𝑡​\(L𝑝𝑖𝑣𝑜𝑡\)n=1k​∑i=0k𝐿𝑎𝑡​\(L𝑝𝑖𝑣𝑜𝑡,i\)n\\mathit\{Lat\}\(L\_\{\\mathit\{pivot\}\}\)\_\{n\}=\\frac\{1\}\{k\}\\sum\_\{i=0\}^\{k\}\\mathit\{Lat\}\(L\_\{\\mathit\{pivot\},i\}\)\_\{n\}
V𝑁𝑒\+𝑇𝑖,n=𝐿𝑎𝑡​\(H𝑝𝑖𝑣𝑜𝑡\)n−𝐿𝑎𝑡​\(L𝑝𝑖𝑣𝑜𝑡\)nV\_\{\\mathit\{Ne\}\+\\mathit\{Ti\},n\}=\\mathit\{Lat\}\(H\_\{\\mathit\{pivot\}\}\)\_\{n\}\-\\mathit\{Lat\}\(L\_\{\\mathit\{pivot\}\}\)\_\{n\}
![Refer to caption](https://arxiv.org/html/2607.20803v1/x7.png)Figure 6:Multi\-dimensional steering experiment result, where steering vector controls both Ti and Ne output in a coherent mannerWe then evaluate the effectiveness of steering, as described in Section[5](https://arxiv.org/html/2607.20803#S5)\. This time, scores of both steered dimensions are monitored\. The result is plotted in 3D in Figure[6](https://arxiv.org/html/2607.20803#S8.F6), with projections for each pair of variables\. As shown in the two projection plots below, both dimensions steered successfully\. As shown in the top\-down projection, the interference between the two dimensions is also well addressed\. Overall,𝑇𝑖\\mathit\{Ti\}and𝑁𝑒\\mathit\{Ne\}increase linearly with each other asλ\\lambdaincreases\.

Now that we have derived and verified aV𝑁𝑒\+𝑇𝑖V\_\{\\mathit\{Ne\}\+\\mathit\{Ti\}\}directly from the sample space, we canbacktrackfrom this activation direction to evaluate its spatial relationships with single\-dimensional steering directions\. While prior workBhandariet al\.\([2026](https://arxiv.org/html/2607.20803#bib.bib4)\)focuses on the geometrical independence of single\-dimensional steering directions, we present another research question:Does the effective multi\-dimensional steering direction we found belong to a linear system formed by single\-dimensional steering directions?

This hypothesis is equivalent to solving the linear system:

V′=w1​V𝑇𝑖\+w2​V𝑁𝑒V^\{\\prime\}=w\_\{1\}V\_\{\\mathit\{Ti\}\}\+w\_\{2\}V\_\{\\mathit\{Ne\}\}
While minimizing the residual:

r=∥V𝑁𝑒\+𝑇𝑖−V′∥r=\\lVert V\_\{\\mathit\{Ne\}\+\\mathit\{Ti\}\}\-V^\{\\prime\}\\rVert
Wherew1w\_\{1\}andw2w\_\{2\}are linear coefficients, andrris the relative residual since all directions are normalized to have unit length\. If the multi\-dimensional steering direction is a linear combination of single directions, we should expectr≈0r\\approx 0\. The idea and the result for𝑁𝑒\+𝑇𝑖\\mathit\{Ne\}\+\\mathit\{Ti\}are further demonstrated via PCA plotting in Figure[7](https://arxiv.org/html/2607.20803#S8.F7)\.

![Refer to caption](https://arxiv.org/html/2607.20803v1/x8.png)Figure 7:PCA plotting of activation directions related to multi\-dimensional steering\. The direction with combined semantics does not lie in the sum of single directions\.Combinationw1w\_\{1\}w2w\_\{2\}rrTi \+ Fe0\.770\.540\.37Ti \+ Ne0\.930\.280\.37Ti \+ Se0\.780\.520\.37Fi \+ Te0\.210\.800\.49Fi \+ Ne0\.480\.580\.32Fi \+ Se0\.650\.450\.53Ni \+ Te0\.640\.420\.23Ni \+ Fe0\.810\.390\.42Ni \+ Se0\.740\.440\.27Si \+ Te0\.340\.690\.25Si \+ Fe0\.630\.560\.29Si \+ Ne0\.750\.370\.65Table 5:Results for solving multi\-dimensional linear systems using Least\-Squares method and measuring the residual from the effective steering vector![Refer to caption](https://arxiv.org/html/2607.20803v1/x9.png)Figure 8:Scatter plot of the residual from actual multi\-dimensional effective directions to their closest approximations in solved linear systemsWe repeat the above process and solve1212common Jungian Cognitive Function combinations using the Least\-Squares method\. The results are listed in Table[5](https://arxiv.org/html/2607.20803#S8.T5)\. The residual distribution is shown in Figure[8](https://arxiv.org/html/2607.20803#S8.F8)\. The relative residual spans from0\.230\.23to0\.650\.65\. This observation suggeststhe multi\-dimensional steering direction is not in a linear system formed by single\-dimensional steering directions, and aligns with the hypothesisBhandariet al\.\([2026](https://arxiv.org/html/2607.20803#bib.bib4)\)that personality vectors are inherently entangled, so their linear combinations cannot be viewed as semantic additions\.

## 9Conclusion

We presented a framework that enables activation steering over Jungian Cognitive Functions through a combination of scalable evaluation, synthetic personality data generation, steering\-vector extraction, and activation\-space analysis\. Controllable personality, when decomposed as acognitive processJung \([1971](https://arxiv.org/html/2607.20803#bib.bib11)\)rather than static personality traits, remains an underexplored dimension of activation engineering\. By bringing Jungian Cognitive Functions into the landscape of LLM personality steering, we uncovered many useful insights that can help understand how LLMs process personalities, while providing a reusable framework for both single\-dimensional and multi\-dimensional personality steering\.

By extracting steerable latent directionsRimskyet al\.\([2024](https://arxiv.org/html/2607.20803#bib.bib10)\)corresponding to all eight Jungian functions, we achieved strong monotonic behavioral control\. The layer\-based hyperparameter scan reveals that personality information is concentrated in middle transformer layers, aligned with the hypothesisSkeanet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib13)\); Wuet al\.\([2025](https://arxiv.org/html/2607.20803#bib.bib14)\)that different layers of transformers act in an encode\-compute\-decode structured chain of roles and actions\. The geometry of activation space also reflects meaningful relationships between personality dimensions, where rationalJung \([1971](https://arxiv.org/html/2607.20803#bib.bib11)\)functions have different structural geometry compared to irrational ones\. Using a pivot score, we also discovered that multi\-dimensional steering is possible through direct cluster extraction, bypassing the combination of distinct single\-dimensional steering vectors\. Backtracking from the multi\-dimensional steering vector shows that multi\-dimensional activation direction is not a linear combination of single\-dimensional directions\.

The insights gained from layer and spatial analysis offer a foundation for future research to further clarify the characteristics of activation space in LLMs\. By utilizing the methods and theNarrationDBdataset we have released, subsequent studies may reveal additional aspects of personality mechanics within LLM activation space\. Advancing understanding in this domain could ultimately enable both academic and industry stakeholders to develop more effective strategies for interpreting and leveraging LLMs\.

## References

- AI@Meta \(2024\)Llama 3 model card\.External Links:[Link](https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md)Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p6.1),[§3](https://arxiv.org/html/2607.20803#S3.p1.1),[§3](https://arxiv.org/html/2607.20803#S3.p6.1)\.
- R\. Allbert, J\. K\. Wiles, and V\. Grankovsky \(2025\)Identifying and manipulating personality traits in llms through activation engineering\.External Links:2412\.10427,[Link](https://arxiv.org/abs/2412.10427)Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p6.1),[§4](https://arxiv.org/html/2607.20803#S4.p1.6)\.
- P\. Bhandari, U\. Naseem, and M\. Nasim \(2026\)Do personality traits interfere? geometric limitations of steering in large language models\.InFindings of the Association for Computational Linguistics: ACL 2026,M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 29284–29293\.External Links:[Link](https://aclanthology.org/2026.findings-acl.1463/),[Document](https://dx.doi.org/10.18653/v1/2026.findings-acl.1463),ISBN 979\-8\-89176\-395\-1Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p2.1),[§8](https://arxiv.org/html/2607.20803#S8.p14.1),[§8](https://arxiv.org/html/2607.20803#S8.p20.3)\.
- R\. Chen, A\. Arditi, H\. Sleight, O\. Evans, and J\. Lindsey \(2025\)Persona vectors: monitoring and controlling character traits in language models\.External Links:2507\.21509,[Link](https://arxiv.org/abs/2507.21509)Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p1.1),[§1](https://arxiv.org/html/2607.20803#S1.p2.1),[§4](https://arxiv.org/html/2607.20803#S4.p2.4)\.
- Y\. Dong, L\. Jin, Y\. Yang, B\. Lu, J\. Yang, and Z\. Liu \(2025\)From rational answers to emotional resonance: the role of controllable emotion generation in language models\.External Links:2502\.04075,[Link](https://arxiv.org/abs/2502.04075)Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p6.1),[§4](https://arxiv.org/html/2607.20803#S4.p1.6)\.
- Felmonon \(2026\)Jungian\-typology\-assessment: full\-stack jungian typology assessment with auth, billing, persisted results, and ai\-generated reports\.Note:[https://github\.com/FELMONON/jungian\-typology\-assessment](https://github.com/FELMONON/jungian-typology-assessment)GitHub repository\. Accessed: 2026\-05\-14Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p5.1),[§2](https://arxiv.org/html/2607.20803#S2.p1.1),[§2](https://arxiv.org/html/2607.20803#S2.p3.3)\.
- M\. Frising and D\. Balcells \(2026\)Linear personality probing and steering in llms: a big five study\.External Links:2512\.17639,[Link](https://arxiv.org/abs/2512.17639)Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p1.1),[§1](https://arxiv.org/html/2607.20803#S1.p2.1),[§4](https://arxiv.org/html/2607.20803#S4.p2.4)\.
- J\. Huang, W\. Wang, E\. J\. Li, M\. H\. LAM, S\. Ren, Y\. Yuan, W\. Jiao, Z\. Tu, and M\. Lyu \(2024\)On the humanity of conversational AI: evaluating the psychological portrayal of LLMs\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=H3UayAQWoE)Cited by:[§2](https://arxiv.org/html/2607.20803#S2.p3.3)\.
- E\. Jorgenson \(2020\)Data from the statistical "which character" personality quiz\.OpenPsychometrics\.External Links:[Link](https://openpsychometrics.org/tests/characters/data/)Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p5.1),[§3](https://arxiv.org/html/2607.20803#S3.p1.1),[§3](https://arxiv.org/html/2607.20803#S3.p3.3),[§8](https://arxiv.org/html/2607.20803#S8.p1.4)\.
- C\. G\. Jung \(1971\)Psychological types\.2 edition,Collected Works of C\. G\. Jung, Vol\.6,Princeton University Press,Princeton, NJ\.Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p3.1),[§5](https://arxiv.org/html/2607.20803#S5.p5.4),[§7](https://arxiv.org/html/2607.20803#S7.p2.2),[§9](https://arxiv.org/html/2607.20803#S9.p1.1),[§9](https://arxiv.org/html/2607.20803#S9.p2.1)\.
- S\. Marks and M\. Tegmark \(2024\)The geometry of truth: emergent linear structure in large language model representations of true/false datasets\.External Links:2310\.06824,[Link](https://arxiv.org/abs/2310.06824)Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p1.1),[§1](https://arxiv.org/html/2607.20803#S1.p6.1),[§4](https://arxiv.org/html/2607.20803#S4.p1.6),[§8](https://arxiv.org/html/2607.20803#S8.p2.2)\.
- N\. Rimsky, N\. Gabrieli, J\. Schulz, M\. Tong, E\. Hubinger, and A\. Turner \(2024\)Steering llama 2 via contrastive activation addition\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 15504–15522\.External Links:[Link](https://aclanthology.org/2024.acl-long.828/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.828)Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p1.1),[§5](https://arxiv.org/html/2607.20803#S5.p1.1),[§9](https://arxiv.org/html/2607.20803#S9.p2.1)\.
- O\. Skean, M\. R\. Arefin, D\. Zhao, N\. N\. Patel, J\. Naghiyev, Y\. Lecun, and R\. Shwartz\-Ziv \(2025\)Layer by layer: uncovering hidden representations in language models\.InProceedings of the 42nd International Conference on Machine Learning,A\. Singh, M\. Fazel, D\. Hsu, S\. Lacoste\-Julien, F\. Berkenkamp, T\. Maharaj, K\. Wagstaff, and J\. Zhu \(Eds\.\),Proceedings of Machine Learning Research, Vol\.267,pp\. 55854–55875\.External Links:[Link](https://proceedings.mlr.press/v267/skean25a.html)Cited by:[§6](https://arxiv.org/html/2607.20803#S6.p2.2),[§9](https://arxiv.org/html/2607.20803#S9.p2.1)\.
- Z\. Wu, X\. V\. Yu, D\. Yogatama, J\. Lu, and Y\. Kim \(2025\)The semantic hub hypothesis: language models share semantic representations across languages and modalities\.InThe Thirteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=FrFQpAgnGE)Cited by:[§1](https://arxiv.org/html/2607.20803#S1.p1.1),[§6](https://arxiv.org/html/2607.20803#S6.p2.2),[§9](https://arxiv.org/html/2607.20803#S9.p2.1)\.

## Appendix

## Appendix AJungian Questionnaire Statements

### A\.1Fe, Positive

Younaturallyattunetotheemotionalatmosphereofaroomandadjustyourbehavioraccordingly\.

Youmakedecisionsbyconsideringtheimpactonsocialharmonyandsharedvalues\.

Youfeelresponsibleforensuringeveryoneinagroupfeelsincluded\.

Whensomeoneisupset,younaturallywanttocomfortthemandrestoretheirwell\-being\.

Insocialsituations,youinstinctivelysmoothoverconflictstomaintaingroupharmony\.

Whenplanningevents,youconsiderwhatwouldmakeeveryonecomfortableandhappy\.

Youoftenknowwhatothersneedemotionallybeforetheyexpressit\.

Indisagreements,youtendtoprioritizemaintainingtherelationshipoverbeingright\.

Youfeelenergizedwhenyoucanhelpothersandcontributetogroupwell\-being\.

Whenmeetingnewpeople,younaturallybuildrapportandputthematease\.

Youconsiderhowyourdecisionswillaffectothers’feelingsbeforeacting\.

### A\.2Fe, Negative

Youfeeldisconnectedfromothers’emotionalstatesandmaysaythingsthatinadvertentlyhurt\.

Understress,youbecomecoldlycriticalandfindfaultwitheveryonearoundyou\.

Whenoverwhelmed,youdevelopparanoidthoughtsthatothersareungratefulorhavehiddenmotives\.

Incrisis,youreducerelationshipstocynicaltransactions,dismissinggenuineemotionalbonds\.

### A\.3Fi, Positive

Youhavestronginnerconvictionsthatyoumaystruggletoarticulatebutwon’tcompromise\.

Beingauthentictoyourinnerselfismoreimportantthanfittingin\.

Youevaluateworthbasedonaprivate,internalsystemofvalues\.

Whensomethingviolatesyourcorevalues,youfeelitasavisceral,physicalreaction\.

Insituationswhereotherscompromisetheirprinciples,youfinditdifficulttodothesame\.

Youformdeepemotionalbondswithselectindividualsratherthanmaintainingmanycasualfriendships\.

Whenmakingdecisions,youconsultyourinternalsenseofwhatfeelsrightorwrong\.

Youaredrawntocausesthatalignwithyourpersonalvalues,eveniftheyareunpopular\.

Ingroupsettings,youquietlymaintainyourpositionevenwhenothersdisagree\.

Whensomeoneactsagainsttheirstatedvalues,younoticeimmediatelyandfeeluncomfortable\.

Youpreferexpressingemotionsthroughcreativeworkoractionsratherthandirectverbalexpression\.

### A\.4Fi, Negative

Youlosetouchwithwhatyouactuallyvalueandfeelinauthenticorhollowunderpressure\.

Understress,youbecomeobsessedwithorganizingandcontrollingtheexternalworld\.

Whenoverwhelmed,youissueharshjudgmentsandultimatumsthatseemoutofcharacter\.

Incrisis,youfixateobsessivelyonfactsandefficiencywhileignoringyourdeeperfeelings\.

### A\.5Ne, Positive

Youfrequentlyseeconnectionsbetweenseeminglyunrelatedideasordomains\.

Youbecomeenergizedbynovelpossibilitiesand"whatcouldbe\."

Youpreferkeepingoptionsopenratherthancommittingtoasingledefinitivepath\.

Whenbrainstorming,yougeneratemanyideasrapidly,buildingonothers’contributions\.

Inconversations,youoftenjumpfromtopictotopicfollowinginterestingtangents\.

Youaredrawntounconventionalapproachesandquestion"thewaythingshavealwaysbeendone\."

Whenstartinganewproject,youaremostexcitedduringtheinitialcreativephase\.

Younoticepatternsandpossibilitiesthatothersseemtomiss\.

Inroutinesituations,youlookforwaystoinnovateordothingsdifferently\.

Youenjoyplayingdevil’sadvocateorexploringideasfrommultipleperspectives\.

Whenfacedwithconstraints,youinstinctivelylookforcreativeworkarounds\.

### A\.6Ne, Negative

Youbecomeconvincedofasinglenegativepossibilityandcan’tseealternatives\.

Understress,youbecomefixatedonbodilysymptomsandhealthconcerns\.

Whenoverwhelmed,youdwellobsessivelyonpastfailuresandnegativememories\.

Incrisis,youoverindulgeinfood,drink,orsensorypleasuresasescape\.

### A\.7Ni, Positive

Solutionsorinsightsoftencometoyoufullyformed,withoutconsciousstep\-by\-stepreasoning\.

Youaredrawntosymbolicmeaningandtheunderlyingarchetypalpatternsofevents\.

Youoftenhaveasingularvisionofthefuturethatyoufeelcompelledtorealize\.

Whenmakingdecisions,youtrustyourgutinstinctsevenwithoutconcreteevidence\.

Incomplexsituations,youperceivethedeepermeaningorhiddendynamicsatplay\.

Youoftenknowhowthingswillunfoldbeforetheyhappen\.

Whenpursuingagoal,youmaintainfocusonthelong\-termvisiondespitedistractions\.

Youseemetaphorsandsymbolsineverydaylifethatothersdon’tnotice\.

Inconversations,youoftencuttotheessentialpointthatothersarecirclingaround\.

Youhavedifficultyexplainingyourinsightsbecausetheycomewithoutclearlogicalsteps\.

Whenplanning,youfocusontheultimatedestinationratherthanthespecificroute\.

### A\.8Ni, Negative

Youfeelominousaboutthefuturebutcan’tarticulatewhy,leadingtovaguedread\.

Understress,youbecomeobsessedwithphysicaldetailsorsensoryindulgences\.

Whenoverwhelmed,youengageinimpulsiveorexcessiveeating,shopping,orphysicalactivities\.

Incrisis,youbecomehypersensitivetoyourphysicalenvironmentordevelopfixationsonobjects\.

### A\.9Se, Positive

Youfeelmostalivewhenfullyimmersedinintense,immediatesensoryexperiences\.

Youprefertotakeimmediateactionratherthanspendingalongtimeplanning\.

Younoticeaestheticdetailsandphysicalchangesinyourenvironmentinstantly\.

Whenopportunitiesarise,youseizetheminthemomentratherthanwaiting\.

Inemergencies,yourespondquicklyandpracticallywithoutoverthinking\.

Youenjoyphysicalactivities,sports,orexperiencesthatengageyourbodyfully\.

Whensomethingexcitingishappening,youwanttobethereexperiencingitfirsthand\.

Youarehighlyattunedtofashion,design,andthevisualappealofyoursurroundings\.

Inconversations,youpreferdiscussingconcrete,real\-worldtopicsoverabstracttheories\.

Youlearnbestbydoingratherthanreadingorhearingaboutsomething\.

Whenbored,youseekoutnewexperiences,thrills,orstimulatingenvironments\.

### A\.10Se, Negative

Youbecomeclumsy,losetrackofphysicalsurroundings,oroverindulgeinsensoryescapewhenstressed\.

Underpressure,youbecomeobsessedwithdarkpremonitionsaboutthefuture\.

Whenoverwhelmed,youseeominoussignsandhiddenmeaningsineverydayevents\.

Incrisis,youwithdrawfromactionandbecomeparalyzedbyparanoidvisionsofwhatmightgowrong\.

### A\.11Si, Positive

Certainsensoryexperiences\(smells,textures\)transportyouvividlytopastmemories\.

Youvalueestablishedmethodsandtraditionsthathaveprovenreliableovertime\.

Youcomparethepresentsituationdetailedlyagainstyourrichinternaldatabaseofpastexperiences\.

Whenlearninganewskill,youpreferstep\-by\-stepinstructionsandestablishedprocedures\.

Infamiliarenvironments,younoticeimmediatelywhensomethinghaschangedorisoutofplace\.

Youfindcomfortinroutines,rituals,andpredictablepatternsindailylife\.

Whenmakingdecisions,yourelyheavilyonwhatworkedwellinsimilarpastsituations\.

Youhaveastrongmemoryforspecificdetailsofpasteventsandexperiences\.

Inunfamiliarsituations,youlookforpatternsthatresemblesomethingyou’veexperiencedbefore\.

Youprefertried\-and\-trueapproachesoverexperimentaloruntestedmethods\.

Whenrecountingevents,yourecallsensorydetailsvividly\-whatyousaw,heard,orfelt\.

### A\.12Si, Negative

Youobsessoverminorbodilysymptomsorbecometrappedinnegativepastmemories\.

Understress,youimaginecatastrophicpossibilitiesandworst\-casescenarios\.

Whenoverwhelmed,youbecomeparalyzedbyallthethingsthatcouldgowrong\.

Incrisis,youmakeimpulsive,out\-of\-characterdecisionsorwildaccusations\.

### A\.13Te, Positive

Whenevaluatingoptions,youprioritizeobjectivecriteriaandmeasurableoutcomesoverpersonalfeelings\.

Younaturallyorganizeinformationintosystems,frameworks,orprocessesthatotherscanfollow\.

Youfindsatisfactioninoptimizingsystemsformaximumefficiency\.

Whengivenanewproject,youimmediatelycreateatimelineanddelegatetasks\.

Inmeetings,youfocusonactionitemsandnextstepsratherthanopen\-endeddiscussion\.

Whensomeonepresentsaplan,youquicklyidentifyinefficienciesandsuggestimprovements\.

Ifaprocessistakingtoolong,younaturallyrestructureittosavetime\.

Whenmakingdecisions,yourelyondata,metrics,andlogicalanalysisovergutfeelings\.

Ingroupsettings,youoftentakechargetoensuretasksgetcompletedonschedule\.

Whensolvingproblems,youfocusonwhatworksinpracticeratherthantheoreticalideals\.

Youpreferclearhierarchiesanddefinedrolesoverambiguousteamstructures\.

### A\.14Te, Negative

Youbecomeparalyzedbydemandsforefficiencyandfeelyourcarefulprocessisdismissed\.

Understress,youbecomeobsessedwithcontrollingeveryminordetailofothers’work\.

Whenoverwhelmed,youfeelcompelledtocreaterigidrulesandschedulesthatleavenoroomforflexibility\.

Incrisissituations,youbecomeharshandcritical,issuingultimatumsandignoringothers’feelings\.

### A\.15Ti, Positive

Youoftenspendtimementallydeconstructinghowsystemswork,evenwithnopracticalpurpose\.

Internalconsistencyofthoughtismoreimportanttoyouthanexternalefficiency\.

Youseekprecisedefinitionsanddistinctionswhenanalyzingaproblem\.

Whenlearningsomethingnew,youneedtounderstandtheunderlyingprinciplesbeforeapplyingthem\.

Indiscussions,youoftenfindyourselfcorrectingimpreciselanguageorflawedlogic\.

Whenpresentedwithapopularopinion,youinstinctivelyquestionitslogicalfoundation\.

Youenjoybuildingmentalmodelsandframeworksthatexplaincomplexphenomena\.

Ifanexplanationhasevenonelogicalflaw,youstruggletoaccepttherestofit\.

Whentroubleshooting,yousystematicallyeliminatepossibilitiesratherthanguessing\.

Youprefertofigurethingsoutindependentlyratherthanaskingothersforhelp\.

Inconversations,youfocusonaccuracyandtruthoverdiplomacyorsocialharmony\.

### A\.16Ti, Negative

Yougraspforlogicalexplanationsthatdon’tquiteholdtogether,becomingrigidunderstress\.

Underpressure,youbecomeobsessedwithhowothersperceiveyouanddesperatelyseekapproval\.

Whenstressed,youmakeuncharacteristicemotionaloutburstsorpublicscenes\.

Incrisis,youfeeloverwhelmedbysocialdemandsanddesperatelyneedvalidationfromothers\.

### A\.17Extroverted

Youprocessyourthoughtsbestbydiscussingthemwithothersortakingaction\.

Atsocialgatherings,youtendtoworktheroomandmeetmanynewpeople\.

Whenfacingaproblem,youprefertotalkitthroughwithothersimmediately\.

Ingroupsettings,youoftenspeakfirstandthinklater\.

Youfeelenergizedandalivewhensurroundedbypeopleandactivity\.

Whenexcitedaboutsomething,youwanttoshareitwithothersrightaway\.

### A\.18Introverted

Afterintensesocialinteraction,youneedsolitudetorestoreyourenergy\.

Youpreferdeepone\-on\-oneconversationsovergroupdiscussions\.

Youneedtimealonetofullyprocessyourthoughtsbeforesharingthem\.

Youfindsmalltalkexhaustingandprefermeaningfulconversations\.

Toomuchexternalstimulationleavesyoufeelingdrainedandoverwhelmed\.

Youoftenfeelthatothersdon’tgiveyouenoughtimetothinkbeforeresponding\.

## Appendix BOpen Ended Generation Prompts

Whenfacedwithamajorlifedecision,whatprocessdoyounaturallyfollowtomakechoice?

Describearecurringfrustrationyouhavewithotherpeople\.Whatdoyouwishtheyunderstoodordiddifferently?

Thinkaboutaprojectorideathatexcitedyou\.Whataspectscapturedyourattention?

Whensomeonecomestoyouwithapersonalproblem,howdoyoutypicallyrespond?Whatisyourprimarygoalinhelpingthem?

Whenyouencounteranewtheory,whatdetermineswhetheryouacceptit,rejectit,orcontinueexploringit?

## Appendix CFull Samples fromNarrationDB

### C\.1King Claudius \(Hamlet\)

Tags: rich, handshakes, arrogant, punchable, suspicious, poisonous, judgemental, old, bitter, bold

\(Withapompousandslightlybittertone\)Ah,yes\.I,KingClaudius,amamanofgreatwealthandrefinement\.My\.\.\.let’ssay,"generous"natureisoftenmisunderstoodbythosearoundme\.Iextendmyhandinfriendship,asisbefittingofaking,andyetsomedaretoquestionmymotives\.\(pausingtoadjustmyroyalattire\)Myconfidenceandbolddecision\-makinghavebroughtprosperitytoDenmark,andIwillnotbeswayedbythewhimsofthosewhowouldseektounderminemyauthority\.

Somemayviewmeas,shallwesay,"assertive"inmydealings,butIassureyou,itismerelyareflectionofmyunwaveringcommitmenttothecrown\.\(eyesnarrowingslightly\)Ihavenopatienceforfoolishnessordisloyalty,andthosewhowoulddaretocrossmewoulddowelltorememberthe\.\.\.consequencesoftheiractions\.\(pausingtostudymynails\)Afterall,akingmustmaintainorderanddiscipline,lesttherealmdescendintochaos\.

Now,Iamnotonetobetrifledwith\.Myyearsofexperienceandwisdomhavegrantedmeacertain\.\.\.perspectiveontheworld\.\(leaningforward,voicetakingonaslightlyacidictone\)AndIwillnothesitatetousemyknowledgeandpowertoprotectmythrone,myfamily,andmylegacy\.So,letthosewhowouldseektochallengemebeware:I,KingClaudius,willnotbethreatened\.\(leaningback,acold,calculatingglintinmyeye\)

### C\.2Jay Gatsby \(The Great Gatsby\)

Tags: lavish, stylish, driven, charismatic, idealist, extreme, manicured, extraordinary, rich, night owl

Thegrandtapestryofmyexistence\.Iamamasterweaver,meticulouslycraftingeverythreadtocreatealifeofunparalleledopulenceandsophistication\.Myworldisoneofextravagantparties,shimmeringjazz,andwhisperedsecrets,allsetagainsttheglitteringbackdropofNewYorkCity’selite\.Iamamanofunyieldingpassion,drivenbyanunwaveringdedicationtothepursuitofperfection\-\-inmysurroundings,myrelationships,andmyveryself\.

Mypersonaisacarefullycuratedworkofart,everydetailprecision\-engineeredtoexudeanauraofrefinementandcharm\.Iamtheepitomeofstyle,alwaysimpeccablyattired,withasmilethatcandisarmeventhemostskepticalofsouls\.Andyet,beneaththepolishedfacade,Iamadreamer,anidealist,consumedbyanall\-encompassingvisionofalifethatcouldbe\-\-alifewiththeonepersonwhohasevertrulyunderstoodme,DaisyBuchanan\.

Mynightsarefilledwiththesoftglowoflanterns,thesweetscentofbloomingflowers,andthemelodicstrainsoforchestras,allblendingtogetherinakaleidoscopeofwonder\.Iamanightowl,acreatureoftheshadows,wheretheboundariesofrealityarestretched,andthepossibilitiesseemendless\.Inthesehours,Iamfreetoindulgemyfantasies,tochasethegreenlightthatbeckonsmetowardanelusive,yetirresistible,destiny\.IamJayGatsby,amanofextraordinarymeans,withanunrelentingthirstfortheextraordinary\-\-inlove,inlife,andinthegrand,sweepingnarrativeofmyown,self\-createdlegend\.

### C\.3Sancho Panza \(Don Quixote\)

Tags: devoted, unambiguous, follower, slovenly, earth, foodie, submissive, pacifist, short, moderate

\(sighing\)Ah,yes\.\.\.I’mSanchoPanza,thesquiretothat\.\.\.enthusiasticknight,DonQuixote\.\(chuckling\)Iseemyselfasasimple,down\-to\-earthfellow,withaloveforgoodfood,awarmhearth,andapeacefullife\.I’mnotoneforgrandadventuresorloftyideals,butI’mfiercelydevotedtomymaster,despitehis\.\.\.\(tactfully\)eccentricities\.Asashort,stoutman,I’velearnedtobemoderateinmyexpectationsandtotakethingsastheycome\.I’mnotonetorocktheboatorseekconflict,preferringtoavoidtroublewheneverpossible\.Myearthynaturekeepsmegrounded,andIfindjoyinthesimplepleasures:aheartymeal,acomfortablebed,andaquieteveningwithfriends\.WhileImaynotalwaysagreewithDonQuixote’santics,Iremainloyalandsubmissivetohiswill,hopingtoguidehimgentlybacktorealitywheneverhisimaginationrunswild\.\(smiling\)That’sme,SanchoPanza\-\-ahumble,unassumingcompanion,livinglifeonemealatatime\.

### C\.4Aladdin \(Aladdin\)

Tags: adventurous, spontaneous, summer, romantic, playful, expressive, bold, ADHD, warm, artistic

ThemagicofAgrabahisinmyblood\!I’mafreespirit,alwayschasingthenextthrill,thenextadventure\.Myheartbeatsfast,andmymindmovesfaster\-I’vegotathousandideasandschemescookinginmybrainatanygiventime\.Somemightcallmeimpulsive,butIsay,whywaitwhenyoucanseizetheday?Thesunisshining,thestreetsarealive,andI’mreadytotakeontheworld\!

Butbeneathallthebravado,I’maromanticatheart\.Ibelieveinthebeautyoftheunknown,thethrillofdiscovery,andthejoyofsharingitallwithsomeonespecial\.I’madreamer,astoryteller,andabitofacharmer\-Ilovetomakepeoplesmileandlaugh\.Myartisinthestreets,inthemarkets,andinthealleys\-Ifindbeautyintheunexpected,andI’malwayslookingforthenextgreattaletotell\.

SomemightsayI’mabitscattered,thatmyattentionspanisshorterthanamagiccarpetride,butIsay,that’sjustpartofmycharm\!I’mamasterofimprovisation,aweaverofwords,andadancerinthestreets\.Ilivelifeonmyownterms,andIwouldn’thaveitanyotherway\.So,comealong,friend,andjoinmeonthiswild,wonderfulride\-we’llmakesomemagic,andwe’llmakesomememories,together\!

## Appendix DVector PCA at More Layers

![Refer to caption](https://arxiv.org/html/2607.20803v1/x10.png)Figure 9:Eight Jungian Cognitive Functions’ steering vectors plotted to 2 PCA dimensions, at layer 7![Refer to caption](https://arxiv.org/html/2607.20803v1/x11.png)Figure 10:Eight Jungian Cognitive Functions’ steering vectors plotted to 2 PCA dimensions, at layer 8![Refer to caption](https://arxiv.org/html/2607.20803v1/x12.png)Figure 11:Eight Jungian Cognitive Functions’ steering vectors plotted to 2 PCA dimensions, at layer 9![Refer to caption](https://arxiv.org/html/2607.20803v1/x13.png)Figure 12:Eight Jungian Cognitive Functions’ steering vectors plotted to 2 PCA dimensions, at layer 10![Refer to caption](https://arxiv.org/html/2607.20803v1/x14.png)Figure 13:Eight Jungian Cognitive Functions’ steering vectors plotted to 2 PCA dimensions, at layer 11
## Appendix ECode and Data Availability

Code and Data will be released upon acceptance:

- •NarrationDB
- •Steering and Benchmarking Framework
- •Extracted Activation Space Vectors