PersonaFactor: Preference Factors with Dynamic Selection for LLM Personalization

Hang Su; Zequn Liu; Jianghan Chao; Qiuyu Tian; Xuesong Lu; Yingce Xia

cs.CLOct 12, 2026

PersonaFactor: Preference Factors with Dynamic Selection for LLM Personalization

Authors: Hang Su, Zequn Liu, Jianghan Chao, Qiuyu Tian, Xuesong Lu, Yingce Xia

Organizations: East China Normal University · Beijing Zhongguancun Academy · Renmin University of China · Southeast University

Abstract

Although users may express different personalization needs across tasks, these task-specific behaviors can reflect shared underlying preference mechanisms. Existing LLM personalization methods often focus on surface-level interactions or undifferentiated histories, overlooking the shared mechanisms underlying these behaviors. We introduce PersonaFactors, a framework that distills reusable and interpretable personalization factors from heterogeneous user interactions. We construct a library of thirteen PersonaFactors by extracting evidence-grounded rationales and clustering recurring patterns across tasks, and develop a dynamic factor-guided generation framework that selects factors relevant to each user's history and current request. Across four personalization benchmarks and four backbone models, Dynamic PersonaFactor achieves the best score on every reported metric. Representation analyses further show that PersonaFactors and user history induce partially overlapping low-dimensional representations, and that their joint use yields complementary shifts. Taken together, these findings suggest that factor-based and history-based personalization can mutually reinforce each other.

Figures & tables

Figure 1: Overview of constructing PersonaFactors. We extract personalization rationales from user interactions across heterogeneous scenarios, cluster them into candidate factors, and obtain canonical PersonaFactors through iterative sampling, naming, merging, and normalization.

Dimension	PersonaFactor
Cognitive and Knowledge Alignment	Cognitive Trust; Schema Congruence; Cognitive Consistency; Cognitive Load
Context and Goal Adaptation	Situational Anchoring; Goal Congruence; Constraint Consistency
Self and Identity Congruence	Self-Concept Congruence; Identity Belonging; Value Identification
Behavioral and Expressive Style Consistency	Behavioral Pattern Consistency; Expressive Style Consistency; Perceived Autonomy Control

Table 1: The shared PersonaFactor library.

Figure 2: Overview of dynamic factor-guided personalized generation. Given the user history and current query, we score each PersonaFactor by its task relevance and generation influence, fuse the two signals to select relevant factors, and use them to guide personalized response generation.

Model	Method	LaMP		LongLaMP		LaMP-QA	PersonalLLM
Model	Method	R-1	R-L	R-1	R-L	LLM Judge	Reward Model
DeepSeek-V4-Pro	Direct	0.2932	0.2580	0.2125	0.1242	0.4915	-0.9013
	RAG	0.3400	0.3004	0.3003	0.1538	0.6503	0.4049
	RAG-SelfFactor	0.3325	0.2904	0.2891	0.1561	0.5940	-0.3861
	RAG-CoPA	0.3456	0.3062	0.3082	0.1565	0.7196	0.4642
	Static PersonaFactor	0.3495	0.3076	0.3141	0.1581	0.7418	0.4795

Table 2: Main results across models and personalization methods. LaMP and LongLaMP report macro-averaged ROUGE-1/ROUGE-L over their evaluated tasks; LaMP-QA uses an LLM-as-a-Judge score, and PersonalLLM uses a reward model score. Best results within each model are boldfaced. Results are reported as averages over three independent runs.

Selector	Art & Entertainment	Lifestyle & Personal Development	Society & Culture
Dynamic PersonaFactor	0.3570	0.4542	0.5139
$R_{x,f}$ -only	0.3457	0.4502	0.5099
$G_{x,f}$ -only	0.3312	0.4427	0.5023
Random- $k$	0.3396	0.4398	0.5056

Table 3: Ablation of the dynamic factor selector on Qwen3-8B using LaMP-QA. The three columns report the benchmark’s LLM-as-a-Judge score for each domain. Higher is better.

Figure 3: LaMP-QA performance as a function of the number of user-history items for Qwen3-8B.

Figure 4: Low-dimensional overlap between History and PersonaFactor effects across LaMP and LongLaMP tasks and layers for Qwen3-8B.

Figure 5: Representation shifts induced by History, PersonaFactor, and their joint condition for Qwen3-8B on LaMP. Top: normalized shifts in a shared two-dimensional PCA space. Lines connect matched observations across input conditions. Bottom: mean unnormalized

L_{2}

norms with bootstrap 95% confidence intervals.

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 9 — Figure A1: Rationale Extraction Prompt.

Type	Component	Content
Input	Current task Write a personalized review of a Hamilton Beach two-slice toaster, given a 5.0 rating and the summary “A great general-purpose toaster with extra perqs for bagel lovers” $[\ldots]$ . User history Prior reviews describe upgrading from inexpensive cookware to professional kitchen equipment and emphasize even heating, durability, and appearance: “I never realized how much of a difference REAL kitchen gear makes $[\ldots]$ ” Other reviews value precise measurements, strong motor performance, absorbency, softness, and product design $[\ldots]$ . Answer “To be candid, I haven’t owned or regularly used a toaster in about twenty years $[\ldots]$ It’s a sturdy design with a nice-looking, soft-industrial aesthetic $[\ldots]$ Once I know what settings work for the things I toast, I never have to guess again $[\ldots]$ the ‘bagel’ specialty function $[\ldots]$ is quite welcome, as is the ‘defrost’ function $[\ldots]$ Auto-shutoff prevents burning in case of jams $[\ldots]$ .”
Output	Rationale 1 Reason: Matches user’s upgrade mindset. Theory: Cognitive Consistency Theory. Explanation: The user consistently reviews products where they acknowledge upgrading from basic/budget items to quality gear (cookware, towels, mixer). The answer mirrors this pattern by admitting past lack of toaster use and praising the upgrade to modern features. Rationale 2 Reason: Aligns with practical value focus. Theory: Utilitarian Value Perception. Explanation: User history shows appreciation for functional improvements (even cooking, precise measurements, absorbent towels). The answer highlights practical benefits like auto-shutoff, defrost function, and consistent toasting results. Rationale 3 Reason: Reflects aesthetic appreciation. Theory: Aesthetic-Usability Effect. Explanation: User frequently mentions product appearance (handsome cookware, attractive mixer, good-looking towels). The answer notes the ‘soft-industrial aesthetic’ and how it looks great in the kitchen.

Appendix

Table A.1: Rationale extraction example from LongLaMP product-review generation.

Type	Component	Content
Input	Current task “Paraphrase the following tweet without any explanation before or after it: @morphiine, it’s amusing that you label ohfukk.com as a new venture when it’s simply for entertainment.” User history “@purelynarcotic Ohh.. #twpp is falling silent.. I was waiting to track ur US version.. Next time maybe.. Btw sunny california sounds good”; “@morphiine kya karen yaar.. so many clothes here and there.. $[\ldots]$ ”; “Ohh the gprs is not so constant on this highway..” $[\ldots]$ . Answer “@morphiine U call ohfukk.com as a new venture :p Thts funny.. Its just fun”
Output	Rationale 1 Reason: Matches user’s casual tone. Theory: Tone Congruence Theory. Explanation: The user’s history shows informal, conversational language with emojis and abbreviations (e.g., ‘Btw’, ‘kya karen yaar’). The answer mirrors this style with ‘:p’ and ‘Thats funny’, aligning with the user’s communication preferences. Rationale 2 Reason: Reflects user’s humor style. Theory: Humor Alignment Principle. Explanation: The user frequently uses lighthearted expressions (e.g., ‘Ohh..’, ‘Ohh the gprs’). The answer’s playful tone (‘U call’, ‘just fun’) resonates with the user’s established humor pattern.

Appendix

Table A.2: Rationale extraction example from LaMP-7 tweet paraphrasing.

Panel of Figure 12 — Figure A2: Reason Semantic Similarity Evaluation Prompt.

Panel of Figure 13 — Figure A3: Prompt for naming the shared acceptance mechanism of a rationale cluster.

Panel of Figure 14 — Figure A4: Prompt for merging raw factors by final-name equivalence.

Panel of Figure 15 — Figure A5: Prompt for auditing the raw-factor to global-factor mapping.

Panel of Figure 16 — Figure A6: Prompt for repairing the factor mapping after audit.

Dimension	PersonaFactor	Definition
Cognitive and Knowledge Alignment	Cognitive Trust	Whether generated content uses sources, evidence, reasoning, and explanatory structures the user is likely to trust.
	Schema Congruence	Whether generated content fits the user’s knowledge background, familiar concepts, experience, and ways of understanding.
	Cognitive Consistency	Whether generated content respects judgments, preferences, beliefs, or concerns shown in the user’s history.
	Cognitive Load	Whether complexity, structural density, and explanatory style fit the user’s likely processing capacity.
Context and Goal Adaptation	Situational Anchoring	Whether generated content incorporates the concrete situation, task background, purpose, role context, and possible emotional state.
	Goal Congruence	Whether generated content serves the user’s current problem, goal, or intended next action.

Appendix

Table A.3: Complete definitions of the thirteen PersonaFactors.

Query	History	Dynamically selected factors	Model answer
Generate a headline for the following article: Here’s an illustrator you and your awesome third grade neighbor can both appreciate. Andrew Ho’s colorful portrayals of the $\ldots$	1. A Retired Elementary School Teacher Turned Her Home Into A Folksy Fine Art Mecca: $\ldots$ 2. This Queer, Immigrant Illustrator Starts Every Morning With A Work Of Activist Art: $\ldots$ 3. Artists Are Offering Free Drawings In Exchange For ACLU Donations: $\ldots$	Expressive Style Consistency; Cognitive Consistency; Goal Congruence; Situational Anchoring.	{“headlines”: “This Illustrator’s Colorful Work Charms Both You and Your Third Grade Neighbor”}

Appendix

Table A.4: Qualitative case from LaMP.

Query	History	Dynamically selected factors	Model answer
Do I have to cite my own paper?	4. $\ldots$ Is there a text out there which introduces philosophy via thought experiment rather than throwing big terms at you right from the beginning. $\ldots$ 7. $\ldots$ But sometimes healers leave the queen and follow BK or giants, and that sucks. $\ldots$ 8. Why don’t Kilgrave’s commands affect Kilgrave himself? $\ldots$	Schema Congruence; Cognitive Trust; Situational Anchoring.	Think of it like this: you’re building a house. The bricks you made earlier (your past paper) are now part of the new wall you’re constructing (your current paper). $\ldots$ creating a clear, traceable path for your reader $\ldots$ So, yes, you must cite your own paper whenever you use its content $\ldots$

Appendix

Table A.5: Qualitative case from LaMP-QA.

Query	History	Dynamically selected factors	Model answer
Generate an abstract for the title “DESA: Disparity Estimation With Surface Awareness” using the following items: depth map, depth estimation, normal restriction, 2D aggregation, benchmark performance	1. $\ldots$ In this paper, we first present a depth extraction method based on sampling of the space at camera viewpoints to address the interference problem. $\ldots$ 2. $\ldots$ Experiments show that the proposed method saves up to 76% coding time with negligible PSNR loss and bit rate overhead.	Cognitive Trust; Schema Congruence; Goal Congruence.	We propose DESA, a disparity estimation method with surface awareness that integrates normal restriction into a 2D aggregation framework to enhance depth estimation accuracy. $\ldots$

Appendix

Table A.6: Qualitative case from LongLaMP.

Query	History	Dynamically selected factors	Model answer
When did the First Fleet land in Australia?	1. What is the history of mukluks? How are they traditionally made? Liked response: $\ldots$ History: $\ldots$ Traditional construction: $\ldots$ 3. Can you describe the differences between autotrophs and heterotrophs? Liked response: $\ldots$ 1. Autotrophs: $\ldots$ 2. Heterotrophs: $\ldots$	Situational Anchoring; Cognitive Trust; Cognitive Load.	The First Fleet, a convoy of 11 ships carrying convicts, marines, and officials from Great Britain, landed in Australia in January 1788. Here is a detailed, step-by-step breakdown of the event and its historical context: $\ldots$

Appendix

Table A.7: Qualitative case from PersonalLLM.

Panel of Figure 22 — Figure A7: Prompt for inference-time relevance identification.

Panel of Figure 23 — Figure A8: Prompt for inference-time generation.

Panel of Figure 24 — Figure A9: Prompt used by the prompt-based selector to infer the task-specific PersonaFactor subset.

Benchmark	Scene	Samples	Metric
LaMP	Personalized News Headline Generation	1,925	ROUGE-1/L
	Personalized Scholarly Title Generation	2,500
	Personalized Tweet Paraphrasing	1,500
LongLaMP	Personalized Abstract Generation	4,560	ROUGE-1/L
	Personalized Product Review Generation	1,826
	Personalized Topic Generation	2,452

Appendix

Table A.8: Benchmark scenes, sample counts, and evaluation metrics.

Panel of Figure 26 — Figure A10: RAG-SelfFactor prompt for generating query-specific personalization factors.

Model	Method	News Headline Generation		Scholarly Title Generation		Tweet Paraphrasing
		R-1	R-L	R-1	R-L	R-1	R-L
DeepSeek-V4-Pro	Direct	0.1273	0.1162	0.4332	0.3786	0.3191	0.2792
	RAG	0.1704	0.1505	0.4774	0.4211	0.3721	0.3295
	RAG-SelfFactor	0.1654	0.1443	0.4324	0.3799	0.3998	0.3471
	RAG-CoPA	0.1696	0.1509	0.4773	0.4246	0.3899	0.3432
	Static PersonaFactor	0.1712	0.1515	0.4769	0.4203	0.4003	0.3509

Appendix

Table A.9: Task-level results on LaMP-4 (News Headline Generation), LaMP-5 (Scholarly Title Generation), and LaMP-7 (Tweet Paraphrasing), evaluated with ROUGE-1 and ROUGE-L. Results are reported as averages over three independent runs.

Model	Method	Abstract Generation		Product Review Generation		Topic Generation
		R-1	R-L	R-1	R-L	R-1	R-L
DeepSeek-V4-Pro	Direct	0.3040	0.1871	0.1408	0.0866	0.1926	0.0988
	RAG	0.3446	0.1978	0.2929	0.1406	0.2635	0.1229
	RAG-SelfFactor	0.3772	0.2178	0.3090	0.1508	0.1811	0.0998
	RAG-CoPA	0.3560	0.2022	0.3015	0.1426	0.2670	0.1248
	Static PersonaFactor	0.3613	0.2039	0.3124	0.1449	0.2685	0.1254

Appendix

Table A.10: Task-level results on LongLaMP under the user-setting split. The tasks are Personalized Abstract Generation, Personalized Product Review Generation, and Personalized Topic Generation; evaluation uses ROUGE-1 and ROUGE-L. Results are reported as averages over three independent runs.

Model	Method	Art & [-1pt] Entertainment	Lifestyle & [-1pt] Personal [-1pt] Development	Society & [-1pt] Culture
DeepSeek-V4-Pro	Direct	0.3764	0.4972	0.6009
	RAG	0.5306	0.6482	0.7720
	RAG-SelfFactor	0.5254	0.5983	0.6583
	RAG-CoPA	0.5886	0.7519	0.8182
	Static PersonaFactor	0.6027	0.7817	0.8409
	Dynamic PersonaFactor	0.6158	0.7970	0.8597

Appendix

Table A.11: Task-level results on LaMP-QA across Art and Entertainment, Lifestyle and Personal Development, and Society and Culture. Scores are computed with the benchmark’s LLM-as-a-Judge protocol. Results are reported as averages over three independent runs.

Model	Method	PersonalLLM
DeepSeek-V4-Pro	Direct	-0.9013
	RAG	0.4049
	RAG-SelfFactor	-0.3861
	RAG-CoPA	0.4642
	Static PersonaFactor	0.4795
	Dynamic PersonaFactor	0.4915

Appendix

Table A.12: Task-level results on PersonalLLM. Scores are obtained from the benchmark’s preference reward models. Results are reported as averages over three independent runs.

Model	Factor Selection	LaMP		LongLaMP		LaMP-QA	PersonalLLM
Model	Factor Selection	R-1	R-L	R-1	R-L	LLM Judge	Reward Model
Qwen3-8B	Prompt-based selector	0.3199	0.2801	0.2612	0.1487	0.4028	-0.3255
Qwen3-8B	Score-based selector	0.3336	0.2956	0.3377	0.1859	0.4417	-0.1722
Gemma-3-4B	Prompt-based selector	0.2711	0.2378	0.2298	0.1299	0.4672	-0.2772
Gemma-3-4B	Score-based selector	0.2818	0.2490	0.2943	0.1651	0.5107	-0.1836

Appendix

Table A.13: Comparison of selector realizations within Dynamic PersonaFactor on open-weight models. The prompt-based selector directly predicts the factor subset, whereas the score-based selector uses task-relevance and generation-influence scores. LaMP and LongLaMP report macro-averaged ROUGE-1/ROUGE-L. LaMP-QA uses an LLM-as-a-Judge score, and PersonalLLM uses a reward-model score. Higher is better for all metrics, and the best result within each model is boldfaced. Results are averaged over three independent runs.

Selector	Art & Entertainment	Lifestyle & Personal Development	Society & Culture
Dynamic PersonaFactor	0.4121	0.5358	0.5842
$R_{x,f}$ -only	0.3977	0.5086	0.5727
$G_{x,f}$ -only	0.4021	0.5198	0.5653
Random- $k$	0.4065	0.5156	0.5688

Appendix

Table A.14: Dynamic selector ablation on Gemma-3-4B using LaMP-QA. Higher is better.

Figure A11: LaMP-QA performance as a function of the number of user-history items for Gemma-3-4B.

Figure A12: Low-dimensional overlap between History and PersonaFactor effects across LaMP and LongLaMP tasks and transformer layers for Gemma-3-4B.

Model	Dataset	Task	Top-3 overlap
Qwen3-8B	LaMP	News headline generation	0.180
		Scholarly title generation	0.181
		Tweet paraphrasing	0.293
	LongLaMP	Abstract generation	0.272
		Product review	0.232
		Topic writing	0.257

Appendix

Table A.15: Task-level top-3 subspace overlap between History and PersonaFactor. The overlap is computed after aggregating the model-specific middle-to-late layer set.

Model	Dataset	Task	Full-layer mean	Middle-to-late-layer mean
Qwen3-8B	LaMP	News headline generation	0.352	0.333
		Scholarly title generation	0.437	0.434
		Tweet paraphrasing	0.395	0.368
	LongLaMP	Abstract generation	0.380	0.356
		Product review	0.419	0.399
		Topic writing	0.437	0.428

Appendix

Table A.16: Task-level summary of layer-wise rank-8 subspace overlap between History and PersonaFactor.

Figure A13: Combined activation effects of History and PersonaFactor for Qwen3-8B across LongLaMP tasks. Each task compares History, PersonaFactor, and their joint condition relative to the Direct baseline.

LLM Personalization

LLM personalization adapts language models to users’ preferences, histories, expertise, and changing goals while balancing usefulness, privacy, and control.

See all 130 papers

POPI: Personalizing LLMs via Optimized Natural Language Preference Inference
Oct 17, 2025Yizhuo Chen, Xin Liu, Ruijie Wang +7LLM Personalization Personalized Language Models
University of Illinois Urbana-Champaign · Amazon · University of Notre Dame
MoF: Preference-Aware Mixture Modeling for Black-Box LLM Personalization
Oct 6, 2026Hun ParkLLM Personalization Personalized Language Models
Department of Artificial Intelligence Korea University Seoul, Republic of Korea
Unlocking Latent Personalization in LLMs
Sep 27, 2026Wei Chen, Guanghui Zhu, Zhongliang Cai +1LLM Personalization Personalized Language Models
State Key Laboratory for Novel Software Technology, Nanjing University

Oct 17, 2025J/K move · Enter open · S save

POPI: Personalizing LLMs via Optimized Natural Language Preference Inference

Yizhuo Chen, Xin Liu, Ruijie Wang, Zheng Li+6

University of Illinois Urbana-Champaign · Amazon · University of Notre Dame

LLM Personalization Personalized Language Models Personalized Language Model Alignment Language Model Post-Training User Preference Modeling

Open paper AskPDF arXiv

Large language models (LLMs) are typically aligned with population-level preferences, despite substantial variation across individual users. We introduce POPI, a user-level personalization framework that separates the problem into two components connected by a natural-language interface: a shared inference model that distills heterogeneous user signals into a concise preference summary, and a shared generator that conditions on this summary to produce personalized responses. Both components are trained under a unified preference-optimization objective, with reinforcement learning handling the non-differentiable inference step. This objective decomposes into generator approximation error and summary informativeness, revealing how a single loss simultaneously drives accurate generation and informative summarization. Because the interface is natural language, learned summaries can be inferred once per user and reused across different generators -- including frozen, black-box commercial APIs. Across four personalization benchmarks, POPI generally improves personalization quality while reducing context overhead by up to an order of magnitude.

Abstract

LLM Personalization

Explore similar work

POPI: Personalizing LLMs via Optimized Natural Language Preference Inference

MoF: Preference-Aware Mixture Modeling for Black-Box LLM Personalization

Unlocking Latent Personalization in LLMs