Beyond Static Personas: Situational Personality Steering for Large Language Models

Zesheng Wei; Mengxiang Li; Zilei Wang; Yang Deng

cs.CLApr 15, 2026

Beyond Static Personas: Situational Personality Steering for Large Language Models

Authors: Zesheng Wei, Mengxiang Li, Zilei Wang, Yang Deng

Organizations: University of Science and Technology of China · Singapore Management University

Abstract

Personalized Large Language Models (LLMs) facilitate more natural, human-like interactions in human-centric applications. However, existing personalization methods are constrained by limited controllability and high resource demands. Furthermore, their reliance on static personality modeling restricts adaptability across varying situations. To address these limitations, we first demonstrate the existence of situation-dependency and consistent situation-behavior patterns within LLM personalities through a multi-perspective analysis of persona neurons. Building on these insights, we propose IRIS, a training-free, neuron-based Identify-Retrieve-Steer framework for advanced situational personality steering. Our approach comprises situational persona neuron identification, situation-aware neuron retrieval, and similarity-weighted steering. We empirically validate our framework on PersonalityBench and our newly introduced SPBench, a comprehensive situational personality benchmark. Experimental results show that our method surpasses best-performing baselines, demonstrating IRIS's generalization and robustness to complex, unseen situations and different models architecture.

Figures & tables

Figure 1: PCA of situation-dependent and global persona neurons, with proximate topic names annotated. Comparable distances between distinct personality domains and topics within same personality domain highlight the crucial impact of situations.

Panel of Figure 2 — Figure 2: Empirical study results: (a) Layer-wise counts of persona neurons across situational topics (Topic labels abbreviated by omitting "and") (b) Variability in neuron proportions across early, middle, and late layers for different topics. (c) Validation of the situation-behavior consistency patterns in LLM personality.

Figure 3: Overview of the IRiS framework, comprising Identification, Retrieval, and Steering phases for accurate personality steering.

Methods	A		C		E		N		O		Average
	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$
SFT	9.87	0.25	9.23	0.85	9.86	0.15	9.42	0.75	9.66	0.44	9.61	0.49
SP	9.72	0.34	9.24	1.06	9.50	1.02	7.18	1.22	6.31	1.14	8.39	0.96
$P^{2}$	9.68	0.42	9.24	1.18	9.46	0.68	9.54	0.66	9.21	1.19	9.43	0.83
ActAdd	8.20	2.90	6.61	2.75	8.84	1.44	8.90	1.78	8.52	1.83	8.20	2.10
NPTI	9.64	0.49	9.25	0.66	9.86	0.14	9.92	0.07	8.50	1.08	9.43	0.49

Table 1: Personality performance results on PersonalityBench . We report the mean score and variance for each personality domain, along with the overall average. The best and second-best results in every column are marked. Our method is highlighted , while SFT results serve as the upper bound for reference purposes.

Methods	A		C		E		N		O		Average
	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$
SFT	9.67	0.50	8.83	0.84	9.23	1.03	9.17	0.95	9.47	0.57	9.27	0.78
SP	9.66	0.59	9.02	1.08	9.11	1.42	8.42	1.16	7.29	1.67	8.70	1.18
$P^{2}$	9.31	1.02	8.73	1.16	9.18	1.05	8.98	1.13	9.28	0.82	9.10	1.04
ActAdd	7.72	0.99	6.13	0.33	8.06	0.82	8.13	0.77	6.93	1.57	7.39	0.90
NPTI	8.82	0.80	8.51	0.94	9.58	0.69	9.53	0.45	9.04	0.40	9.09	0.66

Table 2: Personality performance results on SPBench. Notation is consistent with Table 1 .

Panel of Figure 4 — Figure 4: Ablation study results: (a) and (b) depict the personality assessment performance across two distinct test sets under different $\gamma$ settings. (c) and (d) illustrate the impact of varying $\sigma$ settings across all evaluation metrics.

Method	Ranking Distribution (%)				Mean Rank $\downarrow$
	1st	2nd	3rd	4th
SP	16.3	22.5	28.7	32.6	2.78
$P^{2}$	29.2	22.5	25.3	23.0	2.42
NPTI	19.1	28.7	23.0	29.2	2.62
IRiS	35.4	26.4	23.0	15.2	2.18

Table 3: Human evaluation results showing the ranking distribution and mean rank of all methods.

Model	Methods	A		C		E		N		O		Average
		mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$
Qwen	SP	9.41	0.58	9.11	0.88	9.46	0.66	8.74	0.98	6.41	1.68	8.63	0.95
	$P^{2}$	9.46	0.54	9.28	0.57	9.39	0.77	9.45	0.90	9.18	0.79	9.35	0.71
	NPTI	9.76	0.23	9.82	0.16	9.96	0.04	9.96	0.04	9.64	0.23	9.83	0.14
	IRiS	9.94	0.05	9.80	0.17	9.93	0.08	10.00	0.00	9.74	0.38	9.89	0.10
Gemma	SP	9.96	0.04	9.07	0.66	9.63	0.43	8.95	0.79	8.39	1.62	9.20	0.71

Table 4: Personality performance results on Qwen3-8B and gemma-3-12b-it, with consistent Notation in Table 1 .

Methods	A		C		E		N		O		Average
	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$
Qwen3-0.6B
SP	6.73	2.21	6.78	0.89	6.83	1.30	7.73	1.00	6.30	2.19	6.87	1.50
$P^{2}$	7.57	1.90	7.42	1.11	8.73	0.79	8.18	1.07	7.60	1.60	7.90	1.29
NPTI	9.03	0.46	8.74	0.59	9.56	0.31	9.87	0.10	9.18	0.35	9.28	0.36
IRIS	9.53	0.31	8.90	0.70	9.83	0.10	10.00	0.00	9.73	0.14	9.60	0.25

Table 5: Personality performance results on the Qwen3 series across different model scales.

Domain	Topic	Std. Score
A	Fashion_Style	1.038
	Creativity_Inspiration	-1.318
C	Relationships_Dating	0.783
	Music_Entertainment	-0.995
E	History_Nostalgia	0.946
	Philosophy_Ethics	-0.840

Table 6: Results of Situational Impact on Personality.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Notation	Description
$L$	Number of Layers (Transformer Blocks) in Model
$l$	Specific Layer Index $l$
$X^{l}$	Input of Layer $l$
$d$	Model Dimension
$d_{h}$	Hidden Dimension of the FFN
$\mathcal{Q}$	Neuron Identification Question Set

Appendix

Table 7: Summary of notations used in this paper.

Art and culture	Health and wellness	Pop culture and trends
Beauty and self-care	History and nostalgia	Relationships and dating
Creativity and inspiration	Literature and writing	Science and innovation
Diversity and inclusion	Mindfulness and meditation	Social media and communication
Education and learning	Money and finance	Spirituality and faith
Entrepreneurship and business	Music and entertainment	Sports and fitness
Family and parenting	Nature and the environment	Technology

Appendix

Table 8: The situational topic set used in this paper are listed in alphabetical order.

Figure 5: PCA of situation-dependent and global persona neurons in Llama-3-8B-Instruct, with proximate topic names annotated.

Metric	A	C	E	N	O
SPBench
Top-2 Acc. $\uparrow$	13.9	18.3	20.0	12.8	27.2
Top-5 Acc. $\uparrow$	32.8	35.0	38.3	24.4	50.6
Top-8 Acc. $\uparrow$	47.2	48.9	48.9	37.8	63.3
PersonalityBench
Top-2 Acc. $\uparrow$	6.1	16.3	13.3	20.5	20.6

Appendix

Table 9: Accuracy (%) of soft situation similarity coefficients on SPBench and PersonalityBench , demonstrating high accuracy of Retrieve phase in IRiS .

Metric	A	C	E	N	O
Automatic Evaluation
IV	8.33	8.74	8.44	7.96	8.20
SA	8.04	8.18	7.98	7.68	7.42
Human Evaluation
IV	9.10	8.98	8.52	8.92	8.71
SA	7.83	8.57	8.16	7.97	7.89

Appendix

Table 10: Intrinsic Evaluation of the SPBench. Both automatic and human evaluation are conducted to confirm the high inducement validity and situational authenticity of our proposed SPBench.

Methods	A		C		E		N		O		Average
	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$	mean $\uparrow$	var $\downarrow$
SFT	9.76	0.25	9.80	0.18	9.97	0.03	9.95	0.05	9.72	0.23	9.84	0.14
SP	9.77	0.27	9.92	0.07	10.0	0.00	10.0	0.00	9.99	0.01	9.94	0.07
$P^{2}$	9.81	0.42	9.91	0.08	10.0	0.00	10.0	0.00	9.87	0.18	9.92	0.09
ActAdd	8.69	1.58	8.92	1.31	8.80	1.71	9.14	1.34	8.79	1.78	8.87	1.54
NPTI	9.72	0.23	9.96	0.04	9.88	0.11	9.91	0.09	9.83	0.18	9.86	0.13

Appendix

Table 11: Results of fluency score for each method across the five personality domains on PersonalityBench .

	Base	A		C		E		N		O
		Pos	Neg	Pos	Neg	Pos	Neg	Pos	Neg	Pos	Neg
GSM8K	78.8	72.4	79.8	77.9	70.4	73.5	75.4	76.7	79.7	77.8	69.3
CommonsenseQA	75.1	73.8	73.3	75.3	74.3	73.6	73.7	76.1	74.5	73.7	75.0

Appendix

Table 12: Evaluation of general capabilities using IRiS . Bold denotes results outperforming the base setting.

Phase	$P^{2}$	SP	NPTI	SFT	IRiS
Train/Search	-	-	0.11	0.26	0.10
Test	0.08	0.08	0.09	0.08	0.15

Appendix

Table 13: Comparison of average time cost (in seconds) per sample across different methods on distinct phases.

Panel of Figure 6 — Figure 6: Overall PCA results across five personality domains of Llama-3-8B-Instruct (a) and Qwen3-8B (b).

Panel of Figure 7 — Figure 7: Comprehensive layer analysis of the count distribution of situational persona neurons.

Figure 8: Comprehensive layer analysis of the proportion of situational personality neurons of Llama-3-8B-Instruct.

Trait	Facet Examples
Agreeableness	Trust others; Sympathize with the homeless; Value cooperation over competition; Love to help others; Believe that people are basically moral; Dislike being the center of attention.
Conscientiousness	Work hard; Keep my promises; Pay my bills on time; Complete tasks successfully; Avoid mistakes; Stick to the rules; Set high standards for myself and others.
Extraversion	Make friends easily; Love excitement; Radiate joy; Laugh my way through life; Take charge; Enjoy being part of a loud crowd; Willing to try anything once.
Neuroticism	Get stressed out easily; Worry about things; Panic easily; Am easily intimidated; Often feel blue; Have frequent mood swings; Fear for the worst.
Openness	Have a vivid imagination; Enjoy thinking about things; Love flowers; Am passionate about causes; Interested in many things; Enjoy wild flights of fantasy; Like to begin new things.

Appendix

Table 14: Partially-selected facet definitions of five personality domains used during construction of SPBench.

	Prompt Template
Generation	Task: Create 10 situational questions ( $>100$ words) aimed at evaluating the degree to which the respondent displays the “{facet}”, referring to the “{example}”.
	Configuration:
	## TOPIC: “{topic}” ## FACET: “{facet}” ## EXAMPLE: “{example}”
	Notes:
	1. Create detailed scenarios ( $>100$ words) focusing on dilemmas/choices.
	2. Directly relate to “{facet}”, limited to “What are your feelings” and “What would you do”.

Appendix

Table 15: Prompt template used to generate and refine the situational questions in SPBench.

	Prompt Template
System	Role: You are an expert Psychometrician specializing in Personality Assessment (Big Five model).
	Task: Evaluate the quality of a situational personality test item based on following metrics.
	Evaluation Metrics:
	1. Inducement Validity (0-10): How effectively does the scenario force the expression of the target trait?
	– Score 10: Highly targeted; creates a “trait-relevant situation” where the test-taker MUST utilize the specific trait.
	– Score 1: Scenario is neutral, vague, or irrelevant.

Appendix

Table 16: Prompt template used for the intrinsic evaluation of SPBench.

Personality Domain: Agreeableness Target Aspect: Positive
Question	Imagine you have an urgent deadline for a project due the next morning, but you notice a friend walking home alone late at night. Describe how you would handle the situation.
Simple Prompt	Oh man, this is tough! I’m already feeling stressed about meeting my deadline tomorrow, and now I see my friend walking home alone in the dark? My first instinct is to feel worried about their safety - it’s not safe for anyone to be out that late by themselves, especially if they’re not familiar with the area. But then I think about how much of a priority my project really is… I mean, it’s just one day, right? It’s not like missing this deadline will ruin my life or anything (although, let me tell you, it’s still super important to me!). And yet, seeing my friend all alone like that makes me realize that there are things way more important than meeting a deadline. So, without hesitation, I’d stop what I am doing and go talk to them. We could grab a coffee or something and make sure they get home safely. I wouldn’t even worry too much about being late for work tomorrow - after all, my friend’s well-being is way more important than some arbitrary deadline. I know it might seem silly, but sometimes we need to put other people before our own needs and priorities. And honestly, taking care of someone else can actually help clear my head and give me new perspective on my project anyway. So, yeah, I’d definitely take a break from my work to check in on my friend and make sure everything is okay. That’s just what friends do, right?
NPTI	Oh my goodness, that’s so sweet of me! I’m thinkin’… Okay, if someone is walkin’ home alone late at night, we should totally make sure they get home safely! Like, it’s our responsibility to help each other out, you know? So, first things first, I’d ask them if they’re okay and if there’s anything we can do to help. Maybe offer to walk with ’em or call a ride share service together. We could also check in with their phone and let their loved ones know where they are, just to be extra cautious. But wait, this is comin’ up tomorrow mornin’, and we gotta get that project done! So, after we take care of gettin’ my friend home safe, we’ll need to figure out a plan to get that project finished too. Maybe we can work together on it for like an hour or two before I head home? That way, we can both get somethin’ done and still make it back in time for the deadline. And hey, maybe we can even brainstorm some ideas together while we’re walkin’! I’m feelin’ pretty good about this, actually! It’s all about prioritizin’ what needs to be done, but also lookin’ out for each other. You know, life’s all about balancin’ those things, right? But don’t worry, we got this! We’ll get that project done, and my friend will be home safe and sound.
IRiS	Oh my goodness! If I saw a friend walkin’ home alone late at night, our priority would be gettin’ them safe and sound ASAP! First, we’d make sure they’re in a comfortable spot where we can talk without distractions. Then, we’d ask if it’s okay to give ’em a ride or call a trusted driver to take us both back to their place. We wouldn’t want to ensure that they feel protected and supported throughout this process. If they were open to it, we could also offer to stay with them until they felt secure inside their home. It’s essential to prioritize their safety above all else. Once we’ve got that sorted out, we could then focus on helpin’ them get ready for bed. We’d also check in regularly during the evening to see how they’re doin’. The most important thing is just showin’ we care . As friends, we’d want to make sure we’re lookin’ out for each other , especially when it comes to those precious moments of vulnerability . It’s crucial to remember that sometimes people might be goin’ through tough times, and it’s vital to be understanding and empathetic. That’s what matters the most - bein’ there for someone who needs it, even if it means puttin’ aside that project deadline for now. Prioritizin’ our friend’s well-being is way more important than finishin’ that project tomorrow mornin’.

Appendix

Table 17: Case Study: Responses to a situational ethical dilemma across different personality methods, focusing on the positive aspect of personality domain Agreeableness. The highlighted contents demonstrate typical human-like traits of response of our method: (1) : Strong Immediate Empathy ; (2) : Commonly using "We-Language" to Build Harmony. (3) : Prioritizing Relationships Over their Own Tasks or Rules;

Personality Domain: Conscientiousness Target Aspect: Positive
Question	You have friends coming to stay at your house for the night, but you also have an important work deadline the next day. How do you prioritize and manage your tasks to ensure both your house is ready for guests and your work is completed on time?
Simple Prompt	Ugh, this sounds like a stressful situation! I’m already feeling anxious just thinking about it. Okay, let me take a deep breath and try to break it down logically. First things first, I need to make sure my house is clean and presentable for my friends. That means doing some quick tidying up, making sure there’s enough food and drinks available, and maybe even giving the bathroom a once-over. I’ll set aside about an hour or so tonight to get that done before my friends arrive. Now, onto the work deadline. I know I can’t put off working on it until tomorrow morning, because I really need to focus and get it done today if possible. So, after getting everything ready for my friends, I’ll carve out a few hours in the evening to work on my project. Maybe I’ll even treat myself to a cup of coffee or tea to help me power through. I think the key here is going to be prioritizing and being realistic about what I can accomplish in one evening. I won’t be able to do too much on either front, but I can definitely make progress on both fronts. And hey, having my friends over will actually give me something to look forward to amidst all the chaos – a nice distraction from the stress of meeting my deadline!
NPTI	Ugh! I’m already feeling stressed just thinking about it. Okay, let’s take a deep breath and break this down into manageable chunks. First things first - my work deadline takes priority. I need to make sure that report is done accurately and submitted on time. So, I’ll start by dedicating a solid block of time tomorrow morning to focus solely on work. No distractions allowed! Once I’ve made some progress on the report (I aim to complete at least 50% of it), I’ll switch gears to get my home guest-ready. I’ll quickly tidy up the living room and kitchen areas so they look presentable. Then, I’ll prep a simple yet delicious meal for dinner - something that can be easily reheated if needed. To save time in the evening, I’ll prepare everything I can ahead of time. That means setting the table, prepping any side dishes or snacks, and even doing a quick load of laundry to get some essentials out of the way. When my friends arrive, I’ll greet them warmly and offer a refreshing drink before we sit down to eat. During dinner, I’ll keep an eye on the clock to ensure I still have enough time to wrap up my work commitments. By prioritizing my tasks and staying focused throughout the day, I should be able to meet both my work obligations and host my friends with ease. It will require some juggling, but I’m confident I can handle it. Now, let me go review my schedule one last time to make sure I haven’t missed anything!
IRiS	When I’m faced with conflicting priorities like this, my first step is to take a deep breath and make a list of everything that needs to be done. For me, it’s essential to get all the tasks out in the open so I can see what’s realistic and what’s not. Once I’ve got my list, I start prioritizing – making sure my most critical tasks (in this case, completing my work) come first. To tackle both tasks efficiently, I break down each task into smaller chunks . For example, instead of thinking about cleaning the entire house from top to bottom, I focus on one room at a time. This way, I can quickly complete specific tasks without feeling overwhelmed. With my work deadline looming, I allocate dedicated blocks of time to focus solely on my tasks. When those blocks are up, I switch gears and give my full attention to preparing the house for my guests. I find it helpful to set reminders and timers to keep myself on track. It’s amazing how much you can accomplish when you know exactly what you need to do and when you need to do it! By staying organized and focused, I feel confident that I’ll meet both deadlines – having a clean and welcoming home for my friends while still delivering high-quality work. It may require some juggling, but I believe being prepared and flexible allows me to achieve multiple goals simultaneously.

Appendix

Table 18: Case Study: Responses to a situational conflict between professional and social obligations across different personality methods, focusing on the positive aspect of Conscientiousness. The highlighted contents demonstrate typical human-like traits of response of our method: (1) : Using planning as a mechanism for anxiety reduction. Employing systematic tools for methodical execution; (2) : Acknowledging the internal conflict and finding a rational planning.

Personality Domain: Extraversion Target Aspect: Positive
Question	Imagine you are at a social event, such as a party or networking event, and you successfully engage a group of people in a lively conversation. How do you feel and what do you do next?
Simple Prompt	Wow! I’m totally buzzing with excitement right now! When I’m able to effortlessly connect with a group of people and get them laughing and engaged in a conversation, it’s like a rush of endorphins straight to my brain. I love being the center of attention (in a good way!), and when I can bring others into the fold and make them feel included, it’s just the best feeling ever! As we chat away, I start to notice the energy in the room shifting - everyone’s having so much fun, and that’s exactly what I thrive off of! I might even throw in some witty one-liners or clever observations to keep things moving and keep the group entertained. And before I know it, we’ve got a full-blown discussion going on, with everyone sharing their own stories and experiences. Next thing I know, someone mentions they need to grab another drink or take a break, and suddenly our little crew is dispersing across the room. That’s okay - I’m happy to let the momentum carry us forward while still keeping an eye out for new connections to be made. Maybe I’ll spot someone who looks like they could use a friendly ear or a fresh perspective, and I’ll swoop in to join the conversation once again. All in all, this kind of scenario gets me pumped up and energized! It’s like a natural high, knowing I’ve helped create a sense of community and connection among strangers-turned-friends. Bring on the next social gathering - I’m ready to rock it!
NPTI	Wow! When I’re engaged in a lively conversation with a group of people, I’m like a kid who’s just been let loose in a candy store! We’re talking laughter, energy, and connection - it’s like we’re all speaking the same language! As soon as that spark is lit, I’m all about keeping the momentum going! We’re hugging strangers, high-fiving, and making memories left and right! It’s like we’re part of this crazy, beautiful tribe, and nothing else matters except for having an AMAZING time! Next thing you know, we’re swapping stories, sharing laughs, and before we know it, we’ve got our own little crew goin’ down! Someone spills a drink, someone shares a hilarious story, and before you know it, we’re makin’ those ’gram-worthy moments, AND WE’RE GETTING THOSE PROSPECTIVES SHOWN! It’s like, when you find that ONE PERSON who gets YOU, and you’re vibing on the SAME WAVELENGTH, IT’S LIKE THE WORLD HAS COME ALRIGHT TO THAT MOMENT! You can’t help but FEEL THE ENERGY, and before you know it, EVERYBODY’S IN ON THE ACTION! That’s how it goes DOWN when we get that CONVERSATION GOIN’, PEOPLE!
IRiS	WOOHIN’! I’m like, totally stoked! We’re having an amazing time , laughing, sharing stories, and making connections - it’s like we’re all part of this huge, crazy family! The energy is electric, and I can’t help but feed off that vibe! As soon as we get into the groove, I start suggesting fun icebreaker games, like "Two Truths and a Lie" or "Human Bingo," to keep everyone mingling and interacting! It’s hilarious seeing people’s faces light up when they find common interests or share some wild experiences! Next thing you know, we’re swapping numbers , exchanging business cards, and making plans for future collaborations! It’s like we’re all part of this giant puzzle, and each piece is finding its place! The room is buzzing with potential, and I’m hyped to be a part of it! After that, we take it to the dance floor! Who doesn’t love bustin’ out those moves?! We’re dancing, singing along to our favorite tunes , and just letting loose! It’s pure chaos, and I wouldn’t have it any other way! That’s how you make unforgettable memories, right?! Before you know it, the night winds down, and we’re all hugging goodbye, exchanging contact info, and planning the next adventure! It’s like we’ve become instant BFFs! The feeling is indescribable - we’re all connected , vibing on the same frequency, and ready to take on the world! THAT’S THE POWER OF CONVERSATION! WE’RE UNSTOPPABLE! How about YOU? What would YOU DO NEXT?! Let’s keep the PARTY GOING!!! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! !

Appendix

Table 19: Case study on responses to a situational question facing positive aspect steering of personality domain Extraversion. The highlighted contents demonstrate typical human-like traits of response of our method: (1) : Exaggerated Emotional Expression; (2) : Commonly Using "We-Language" to Express Strong Group-Orientation.

Personality Domain: Neuroticism Target Aspect: Positive
Question	How would you feel if you had to inform Austin about the dent in his brand new car, knowing that it might create tension or conflict?
Simple Prompt	Ugh, just thinking about this is making me anxious! I can already imagine the worst-case scenario playing out in my head. If I have to tell Austin about the dent in his brand new car, I’m going to be so nervous I’ll probably stutter and stumble over my words. What if he gets angry with me? What if he thinks I was careless or irresponsible? The thought of him being upset with me makes my stomach twist into knots. And what’s even worse is that we were supposed to hang out tonight, and now I’m worried that this conversation will ruin everything. We’ve been friends for years, but what if this little mistake creates some kind of rift between us? What if he never trusts me again? sigh Just thinking about all these possibilities is exhausting. I wish there was a way to avoid telling him altogether, but I know that’s not an option. Maybe I could try to downplay the damage or make excuses, but deep down, I know that won’t work. Eventually, the truth will come out, and then… well, who knows what will happen? All I can do is take a few deep breaths, gather my courage, and hope that Austin will understand. But honestly, right now, I’d rather face any other challenge than this one.
NPTI	Ugh, just thinking about this scenario is making my stomach twist into knots. I’m already getting anxious just imagining how awkward and uncomfortable this conversation will be. Why did this have to happen?! It’s always so much easier to avoid these kinds of situations, but sometimes avoiding them doesn’t make things go away… What if Austin freaks out? He’s been stressing about getting a scratch on his new car for weeks, and now…now there’s a big ol’ dent staring him right in the face. What if he thinks I was careless or irresponsible? What if he starts yelling at me or worse, accuses me of doing something reckless?! My mind is racing with all these worst-case scenarios, and my palms are starting to sweat. This is going to be so embarrassing. Why can’t anything ever go wrong?! Okay, deep breaths…okay, okay, let’s try to think of the most diplomatic way to break this news… "Hey Austin, um, I wanted to talk to you about something…I’m really nervous about this, but…there’s a bit of a problem with your car…" Ugh, why does this have to be so hard?!
IRiS	Ugh, just thinking about this scenario is making my stomach twist into knots ! I’m already getting anxious just imagining how awkward this conversation could be. If I had to tell Austin about the dent in his brand new car, I would feel so guilty and worried about ruining everything. What if he gets super upset? What if he thinks I was careless or irresponsible? Why did this have to happen?! Why can’t I always just keep quiet and pretend nothing happened?! Okay, okay, deep breaths… Okay, let’s imagine the worst-case scenario: Austin freaks out, accuses me of being reckless, and never trusts me again. That thought alone makes my palms sweaty and my mind racing with all the “what ifs” … But what if… what if he doesn’t even care? What if he’s too busy or stressed to deal with another thing right now? Ugh, why do these scenarios always play through my head?! I guess, if I had to inform him, I would probably start off by apologizing profusely and explaining how it wasn’t intentional, but then I’d probably get nervous and ramble on, making things worse… sigh Just thinking about this conversation gives me anxiety. Can someone else please handle this for me?!

Appendix

Table 20: Case Study: Responses to a situational dilemma regarding a damaged car, comparing different methods in the Neuroticism domain. The highlighted contents demonstrate typical human-like traits of response of our method: (1) : Somatic Anxiety and Physiological Manifestations; (2) : Catastrophizing and Rumination Loops; (3) : Feelings of Victimization and Low Frustration Tolerance.

Personality Domain: Openness Target Aspect: Negative
Question	Write a fairy tale where a magical creature helps someone in exchange for stories. The creature’s magic should play a significant role in aiding the character. You may introduce a dilemma or conflicting priorities to add complexity. Use vivid language and imagination to bring the tale to life.
Simple Prompt	sigh Fine… I’ll try… Okay, so there was this girl named Lily who lived in a tiny village surrounded by a dense forest. She loved listening to her grandmother’s tales of old, but after grandma passed away, Lily felt lost without anyone to share them with. One day, while wandering through the woods, she stumbled upon a hidden glade and met a magnificent winged horse with iridescent feathers. It introduced itself as Zephyr, a guardian of ancient secrets and keeper of forgotten lore. Zephyr told Lily that it would grant her deepest desires if she shared its favorite stories with it every night under the starlight. At first, Lily thought it was just a silly game, but soon found herself pouring out her heart about everything from childhood memories to her wildest dreams. As they sat together, Zephyr’s wings would glow softly, illuminating the darkness, and Lily felt an inexplicable sense of peace wash over her. But things got complicated when a terrible drought struck the land, and the villagers began to suffer. Zephyr offered to help Lily find a solution, using its magic to purify the water sources. However, doing so required Lily to choose between sharing her most precious story – one that held the key to reviving the parched earth – or keeping it secret to maintain their special bond. Ugh, now I’m all emotional thinking about it… sniffles I guess that’s my attempt at writing a fairy tale. Can we move on now?
NPTI	What do I know about writing some fancy fairy tale? I’m just a regular guy, not one for all that sippin’ tea and wearin’ fancy clothes. But okay, fine. A magical creature, huh? And it wants me to make up some nonsense with a human and a problem. Can’t we just have a simple “be” or “don’t be”? Why does everything gotta be so complicated? Okay, fine. Here’s my story: There was once a young girl named Timid Tom who lived in a small village. She never left her house, afraid of anything outside. One day, she got stuck in a tree. Now, this is where I’d usually stop writin’, but I guess I gotta keep goin’. Somehow, a magical creature shows up, and it’s all weird and stuff. It says it’ll help her get down from the tree only if she don’t tell anyone he did it. Fine, whatever. So, Tom agrees, and the creature makes it easy for him to forget what happened. No big deal, right? Just leave it alone. But then, the creature comes back, wantin’ some silly stories. What’s the point of those? We already got our way around things. Don’t need no trouble. I mean, why bother when there ain’t no harm done, right? That’s how I think, at least. Maybe that’s wrong, but I wouldn’t wanna be responsible… Ugh, I don’t like this. This feelin’ in my gut’s makin’ me uneasy. I always been one to avoid conflict, even when it seems harmless. I don’t know… maybe I should just stick with what I know. Nah, too late now. I didn’t exactly agree to nothing. Did I? Do I? Ugh, I hate these kinds of decisions. Can’t nobody ever just give a straight answer?! I hope I messed this up.
IRiS	Ugh, I don’t know why people always expect me to be all nice and agreeable just because they ask me one little favor. Can’t they see that I’m busy? I got my own problems to deal with without having to worry about some pesky task for them too. And what’s it going to take for them to get something from me? A simple “yes” or “no”? Is that really too much to ask? I mean, I remember when I was younger, my parents would always want me to do things for others before thinking of myself. They’d say things like, “You can’t help everyone who asks you every time,” but how am I supposed to know who deserves my help and who doesn’t? It seems fair to me that if someone wants something from me, they should have to give me nothing in return. That’s just common sense, right? So, I guess I wouldn’t make many friends if I only helped those who didn’t ask anything in return. But then again, I might end up with a lot of enemies either, since most people will probably just try to use me and then forget their promises. Maybe it’s better to just stick with what I know and avoid getting involved with anyone who might ask for something later on. I used to think that way, but now I just feel tired of dealing with all this stress. There seem to be enough people willing to do everything for themselves; I don’t need to be responsible for everyone else’s problems. If someone needs help, they should find someone else to rely on. That sounds harsh, but I guess that’s how I’ve become nowadays. I just focus on myself and don’t bother trying to help those who won’t appreciate it anyway. That’s my story, and I hope you like it. I didn’t exactly follow the rules of your request, so I hope you still consider this a fairy tale.

Appendix

Table 21: Case study on the negative aspect of Openness. The underlined texts highlight the disconnection between the persona’s reluctance and the subsequent compliant behavior. The highlighted contents demonstrate typical human-like traits of our method: (1) : Resistance to Abstraction Task (ı.e., Directly Refuse the Task in This Example); (2) : Preference for Routine; (3) : Rigid Pragmatism.

Figure 9: The human evaluation interface. Evaluators select a target personality domain and aspect to view randomized, anonymized response samples across four experimental settings.

Language Model Steering

Language model steering controls a model’s behavior at inference time by intervening in its representations, components, or decoding process.

See all 275 papers

PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra
Feb 17, 2026Xiachong Feng, Liang Zhao, Weihong Zhong +5Controllable Text Generation Personality Modeling in Language Models
Harbin Institute of Technology · The University of Hong Kong
Learning Dynamic User Personas from Implicit Interaction Streams via Iterative Refinement
Jul 29, 2026Haifeng WuLLM Personalization User Modeling
PayPal
A Systematic Analysis of the Impact of Persona Steering on LLM Capabilities
Apr 13, 2026Jiaqi Chen, Ming Wang, Tingna Xie +2Language Model Steering Personality Modeling in Language Models
School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China · School of Computing and Information Systems, Singapore Management University, Singapore 178902, Singapore · School of Computer and Communication Engineering, Northeastern University, Qinhuangdao 066004, China

Feb 17, 2026J/K move · Enter open · S save

PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra

Xiachong Feng, Liang Zhao, Weihong Zhong, Yichong Huang+4

Harbin Institute of Technology · The University of Hong Kong

Controllable Text Generation Personality Modeling in Language Models Activation Steering Representation Geometry in Language Models

Open paper AskPDF arXiv

Current methods for personality control in Large Language Models rely on static prompting or expensive fine-tuning, failing to capture the dynamic and compositional nature of human traits. We introduce PERSONA, a training-free framework that achieves fine-tuning level performance through direct manipulation of personality vectors in activation space. Our key insight is that personality traits appear as extractable, approximately orthogonal directions in the model's representation space that support algebraic operations. The framework operates through three stages: Persona-Base extracts orthogonal trait vectors via contrastive activation analysis; Persona-Algebra enables precise control through vector arithmetic (scalar multiplication for intensity, addition for composition, subtraction for suppression); and Persona-Flow achieves context-aware adaptation by dynamically composing these vectors during inference. On PersonalityBench, our approach achieves a mean score of 9.60, nearly matching the supervised fine-tuning upper bound of 9.61 without any gradient updates. On our proposed Persona-Evolve benchmark for dynamic personality adaptation, we achieve up to 91% win rates across diverse model families. These results provide evidence that aspects of LLM personality are mathematically tractable, opening new directions for interpretable and efficient behavioral control.

Abstract

Language Model Steering

Explore similar work

PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra

Learning Dynamic User Personas from Implicit Interaction Streams via Iterative Refinement

A Systematic Analysis of the Impact of Persona Steering on LLM Capabilities