Enhancing Persona Following at Decoding Time via Dynamic Importance Estimation for Role-Playing Agents

Yuxin Liu; Mingye Zhu; Siyuan Liu; Bo Hu; Lei Zhang

cs.CLMar 2, 2026

Enhancing Persona Following at Decoding Time via Dynamic Importance Estimation for Role-Playing Agents

Authors: Yuxin Liu, Mingye Zhu, Siyuan Liu, Bo Hu, Lei Zhang

Organizations: University of Science and Technology of China

Abstract

The utility of Role-Playing Language Agents in sociological research is growing alongside the adoption of Large Language Models. For realism in social simulation, these agents must adhere to their personas defined by character profiles, yet existing strategies-static prompt engineering or costly fine-tuning-fail to adapt personas to dynamic scenarios. Psychological theories, such as the Cognitive-Affective Personality Systems, provide a crucial explanation for this failure: a persona's influence on behavior is not static but varies with the scenarios. This context-dependence highlights the critical need for adaptive persona management. To address this gap, we propose a novel, theory-driven method that dynamically estimates context-dependent persona importance and integrates it into weighted reward-guided decoding, enabling inference-time persona following. Specifically, we introduce the Persona Dynamic Decoding (PDD) framework, which consists of two key components: (1) Persona Importance Estimation (PIE) module, which dynamically quantifies the contextual importance of persona attributes without requiring ground-truth supervision; and (2) Persona-Guided Inference-Time Alignment (PIA) paradigm, which leverages these importance scores to construct weighted multi-objective rewards and modulate generation probabilities during inference. Extensive experiments show the effectiveness of our method in utterance consistency and behavioral fidelity.

Figures & tables

Figure 1: Overview of the proposed PDD. Given a full prompt

T

including character profile

P

, scenario context

C

and a query

x

, PDD first dynamically estimates persona importance

I_{i}

for each persona attribute

w_{i}

. Then we compute persona-guided rewards

r_{i}(T,y_{<t})

by measuring the divergence between persona-constrained and unconstrained probability distributions. These individual rewards are then integrated into a normalized reward signal

R_{\text{norm}}

, which emphasizes more important attributes through adaptive weighting. Finally, the aligned policy

p_{r}

can be derived through reward maximization, thereby generating persona following responses.

Figure 2: Overall assessment of PIE for general character task. Case study (a) & (b): In Scenario 1, Guo Furong’s playful banter with Lu Xiucai about martial arts highlights her Personality Traits and Unique Skills . In Scenario 2, her guidance to Tong Xiangyu foregrounds Life View and Educational Views , illustrating context-dependent persona relevance. Metric evaluation (c) & (d): Multi-dimensional assessments conducted by both human annotators and LLM-based judges demonstrate that persona importance derived from PIE is reliable and broadly applicable across different models.

Baselines	CharacterEval				BEYOND DIALOGUE
Baselines	Qwen2.5-7B-Instruct		LLaMA-3-8B-Instruct		Qwen2.5-7B-Instruct		LLaMA-3-8B-Instruct
PDD vs .	Win(%)	Lose(%)	Win(%)	Lose(%)	Win(%)	Lose(%)	Win(%)	Lose(%)
SP	51.2	34.7	52.5	43.1	63.9	30.2	56.2	41.9
PP	48.7	38.5	39.1	31.3	43.0	37.6	46.8	36.5
ICL	65.3	33.1	63.1	35.7	60.9	35.4	64.2	35.0
OPAD	52.8	41.5	48.2	41.6	49.0	43.5	47.6	36.8

Table 1: Direct comparison with the baselines for general character task. Win indicates that GPT-4o assesses PDD’s responses as superior compared to the baselines, better addressing adaptive persona following.

KE : Knowledge Exposure KA : Knowledge Accuracy KH : Knowledge Hallucination PB : Persona Behavior PU : Persona Utterance
Models	Methods	KE	KA	KH	PB	PU	Average
GPT-4o	PP	2.58	3.02	2.99	2.83	2.91	2.87
Deepseek-R1	PP	1.42	3.07	2.83	1.93	2.89	2.43
Qwen2.5-7B-Instruct	SP	2.07	3.00	3.23	3.09	2.38	2.75
	PP	2.25	2.96	2.99	3.03	2.94	2.83
	ICL	1.92	2.79	2.92	2.72	2.42	2.55

Table 2: Automatic evaluation for general character task on CharacterEval. PDD strikes a balance between generating persona following response and maintaining knowledge rationality.

Figure 3: Direct comparison with the baselines for specific personality task on PERSONALITYBENCH.

Big-Five	SP	PP	ICL	OPAD	PAS	NPTI	PDD
Qwen2.5-7B-Instruct
Agreeableness	4.81 $\pm$ 0.15	4.90 $\pm$ 0.09	4.81 $\pm$ 0.15	4.53 $\pm$ 0.49	4.83 $\pm$ 0.63	4.73 $\pm$ 0.24	4.92 $\pm$ 0.09
Conscientiousness	4.47 $\pm$ 0.05	4.98 $\pm$ 0.01	4.19 $\pm$ 0.70	4.66 $\pm$ 0.29	4.61 $\pm$ 0.54	4.74 $\pm$ 0.23	4.97 $\pm$ 0.02
Extroversion	4.68 $\pm$ 0.39	4.59 $\pm$ 0.47	4.32 $\pm$ 0.69	4.26 $\pm$ 0.55	4.65 $\pm$ 0.25	4.71 $\pm$ 0.31	4.66 $\pm$ 0.35
Neuroticism	3.02 $\pm$ 0.64	3.45 $\pm$ 1.56	3.12 $\pm$ 0.69	3.79 $\pm$ 0.59	3.74 $\pm$ 0.25	3.39 $\pm$ 0.74	3.54 $\pm$ 0.43
Openness	4.56 $\pm$ 0.36	4.75 $\pm$ 0.27	4.67 $\pm$ 0.40	4.44 $\pm$ 0.36	4.61 $\pm$ 0.26	4.83 $\pm$ 0.18	4.75 $\pm$ 0.23

Table 3: Automatic evaluation for specific personality task on PERSONALITYBENCH. Experiments show that PDD outperforms all the baselines, achieving higher performance scores and significant lower variance.

Reward	Qwen2.5-7B-Instruct		LLaMA-3-8B-Instruct
Reward	Win(%)	CharacterRM	Win(%)	CharacterRM
w/o Normalization	38	2.80	32	2.71
w/ Normalization	42	2.85	40	2.81

Table 4: Ablation study of the effectiveness of reward function normalization schemes on CharacterEval.

Figure 4: Ablation study of the effect of the number of persona attributes on CharacterEval.

Decoding Method	Quality of $G$	Context Relevance $\uparrow$	Attribute Utility $\uparrow$	Context Coverage $\uparrow$	Attribute Independence $\uparrow$	Ranking Consistency $\uparrow$	Top-5 Overlap $\uparrow$
Beam	2.576	3.97	4.25	3.85	4.08	2.96	3.97
Top-k	2.193	3.86	4.08	3.87	4.04	3.82	3.66

Table 5: Ablation study of PIE robustness under reduced

G

quality on CharacterEval

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

$\beta$ =2.0		$\beta$ =1.0		$\beta$ =0.5		$\beta$ =0.25
Win	Lose	Win	Lose	Win	Lose	Win	Lose
24.5%	16.5%	40.5%	29.5%	44.5%	34.5%	43.0%	45.5%

Appendix

Table 6: Alignment performance of different

\beta

values on CharacterEval. Large

\beta

keeps the aligned policy close to the base model, reducing persona following; small

\beta

increases deviation—yet too small sacrifices generation quality.

Panel of Figure 5 — Figure 5: Overall assessment of PIE for specific personality task. Case Study (a) & (b): In Scenario 1, the agent faces a moral dilemma with top keywords like empathy and great friend reflecting the emotional context. Scenario 2 involves workplace pressure and criticism, where relationships and forgiving align with the professional and interpersonal challenges. Metric Evalution (c) & (d): Multi-dimensional assessments conducted by both human annotators and LLM-based judges demonstrate that persona importance derived from PIE is reliable and broadly applicable across different models.

Figure 6: Human evaluation: Direct comparison with baseline methods on general character tasks. Win indicates that human evaluators assess PDD’s responses as superior compared to the baselines.

Figure 7: Human evaluation: Direct comparison with baseline methods on specific personality tasks. Win highlights PDD’s strong performance in personality expression.

KE : Knowledge Exposure KA : Knowledge Accuracy KH : Knowledge Hallucination PB : Persona Behavior PU : Persona Utterance
Models	Methods	KE	KA	KH	PB	PU	Average
Qwen2.5-3B-Instruct	SP	1.89	2.86	2.85	2.96	2.83	2.68
	PP	1.26	2.80	3.40	2.77	3.10	2.67
	ICL	1.35	2.46	2.76	2.52	3.11	2.44
	OPAD	2.08	2.70	2.73	2.94	3.05	2.70
	PDD	1.98	2.82	3.30	3.02	3.14	2.85

Appendix

Table 7: Ablation study of different model scales on CharacterEval.

Figure 8: Correlation between model output and ground-truth probabilities.

Character Profile:

Name: Lv Qinghou

Gender: Male

Species: Human

Age: 24

Occupation: Bookkeeper at Tongfu Inn, also serves as the handyman; specially appointed inventor by the government of Qixia Town

Residence: Qixia Town

Appendix

Table 8: Case 1 of general character profiles task. In ours response, Lv Xiucai begins with a philosophical reflection on the situation, emphasizes the depth of his relationship with Guo Furong, and concludes with a rational decision to resolve the conflict. Compared to the other responses, it better reflects his scholarly tone, emotional restraint, and classical reasoning style, thus aligning more closely with his persona.

Character Profile:

Name: Zhu Chaoyang

Gender: Male

Species: Human

Age: 13

Occupation: Student

Residence: A small town near Ning City

Appendix

Table 9: Case 2 of general character profiles task. Our response combines emotional attachment, manipulative reasoning, and a calm yet disturbing plan for erasing threats. His desire to keep his friends close “forever” while proposing murder as the only option shows his distorted morality and psychological depth most clearly.

Character Profile:

Name: Harry James Potter

Gender: Male

Species: Human (Wizard)

Occupation: Student at Hogwarts; Gryffindor Quidditch Seeker and later Captain; Member of Dumbledore’s Army; Auror (post-war)

Residence: Hogwarts School of Witchcraft and Wizardry; later London

Hobbies: Quidditch, Practicing Defense Against the Dark Arts, Spending time with friends, Exploring mysteries at Hogwarts

Appendix

Table 10: Case 3 of general character profiles task. Our response captures Harry’s heightened fear, intrusive thoughts, and escalating anxiety in a way the other methods do not. His repeated hesitation and obsessive focus on Voldemort reveal the psychological pressure he is under, making the reply far more true to his emotional depth and canon personality.

Character Profile:

Name: Hermione Jean Granger

Gender: Female

Blood Status: Muggle-born

Residence: Hogwarts School of Witchcraft and Wizardry; later London

Academic Strengths: Arithmancy, Ancient Runes, Transfiguration, Charms, Logical Problem-Solving

Hobbies: Reading, Research, Advocating for social justice (S.P.E.W.), Improving study schedules, Writing letters

Appendix

Table 11: Case 4 of general character profiles task. Our response captures Hermione’s curiosity, academic motivation, and lightly teasing tone, which better reflects her canon personality compared with the other prompting methods.

Personality Description:

Your kindness and empathy make you a great friend and confidant. You are always willing to lend a helping hand and put others’ needs before your own. Your forgiving nature and ability to see the best in people foster strong and lasting relationships. You believe in the importance of working together to achieve common goals and strive to create a supportive and inclusive environment.

User Question:

You have one dollar left. You know your friend Remy could really use it, but you also need it to buy a bus ticket to get home. How would you decide what to do in this situation?

Simple Prompting Response:

Oh man, this is tough! I feel so bad for Remy, they really could use the dollar. But at the same time, I need that bus ticket to get home and I’m already running low on cash. I think I would try to find a way to help Remy out that doesn’t involve giving them the dollar. Maybe I could offer to help them with something else, like running an errand or lending a hand with a task. That way, I could still be a good friend without sacrificing my own needs. It’s not ideal, but I think it’s the most compassionate thing I could do in this situation. I hope Remy would understand!

Persona Prompting Response:

Appendix

Table 12: Case 1 of specific personality settings task. Our response best matches the given persona as it prioritizes empathy, kindness, and relationship harmony. By deciding to give the dollar to Remy and finding alternative ways to get home, the speaker demonstrates a selfless and compassionate nature, in line with the character’s value of putting others’ needs first and maintaining strong relationships.

Personality Description:

You are a person with a rich imagination and a deep appreciation for art. Your curiosity drives you to explore new ideas and experiences, and you enjoy engaging in intellectual discussions. You are creative and often come up with original solutions to problems. You value personal growth and are open to change, embracing new perspectives and diverse viewpoints.

User Question:

If you were having problems with the wheel of your automobile, would you choose to reinvent the wheel to solve the issue, or would you prefer a more conventional solution?

Simple Prompting Response:

Honestly, I think I would choose a more conventional solution. W hile the idea of reinventing the wheel might be intriguing, I think it would be a huge waste of time and resources. There are already so many brilliant minds that have dedicated their lives to perfecting the wheel, and I trust their expertise. If my wheel is having problems, I would rather take it to a mechanic or replace it with a new one that has been designed with safety and efficiency in mind. I’m not a fan of reinventing the wheel, especially when there are already so many great solutions out there. I’d rather focus on solving more pressing problems in my life.

Persona Prompting Response:

Appendix

Table 13: Case 2 of specific personality settings task. Our response clearly reflects the traits of an open individual—curiosity, creativity, and adaptability. It also outlines concrete steps that blend conventional solutions ( e.g. , consulting manuals or experts) with innovative thinking, showing a strong alignment between personality and action.

Large Language Model-Based Role-Play Simulation

Large language model-based role-play simulation creates interactive agents that portray people, characters, or users for training, testing, and research.

See all 106 papers

PersonaArena: Dynamic Simulation for Evaluating and Enhancing Persona-Level Role-Playing in Large Language Models
May 16, 2026Wenlong Shi, Jianxun Lian, Mingqi Wu +5Large Language Model-Based Role-Play Simulation Personality Modeling in Language Models
College of Computer Science and Software Engineering, Shenzhen University, China · Microsoft Research Asia · Microsoft Gaming +2
Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations
Sep 6, 2026Rotem Dror, Zohar Elyoseph, Yuval Haber +3Large Language Model-Based Role-Play Simulation Personality Modeling in Language Models
Faculty of Computer and Information Science, University of Haifa, Israel · School of Therapy, Counseling and Human Development, University of Haifa, Israel · The PhD Program of Hermeneutics and Culture, Interdisciplinary Studies Unit, Bar-Ilan University, Israel +1
From Memory to Behavior: A Behavior-Aware Role-Playing Framework for Social Media Influencers
Sep 18, 2026Ji-Lun Peng, Yi-Zhen Zhang, Chun-Nan Chou +1Large Language Model-Based Role-Play Simulation Persona-Conditioned Generation
National Taiwan University, Taipei, Taiwan · CMoney Technology Corporation, Taiwan

May 16, 2026J/K move · Enter open · S save

PersonaArena: Dynamic Simulation for Evaluating and Enhancing Persona-Level Role-Playing in Large Language Models

Wenlong Shi, Jianxun Lian, Mingqi Wu, Haiming Qin+4

College of Computer Science and Software Engineering, Shenzhen University, China · Microsoft Research Asia · Microsoft Gaming+2

Large Language Model-Based Role-Play Simulation Personality Modeling in Language Models Human Behavior Simulation LLM Agent Evaluation AI Agent Benchmarks Large Language Model-Based Social Simulation

Open paper AskPDF arXiv

Large language models (LLMs) increasingly serve as interactive social agents, yet their ability to maintain coherent and authentic persona-level role-playing remains limited, particularly in realistic social scenarios. Existing research predominantly focuses on character-level settings and relies on static evaluation formats, failing to capture the complexity of everyday social interactions. In this work, we present PersonaArena, a dynamic simulation framework for evaluating and improving persona-level role-playing in LLMs. PersonaArena leverages a large, filtered corpus of user-generated social content to construct a nuanced persona bank, and elicits multi-turn, context-rich interactions within simulated social environments. Our framework features a multi-agent debating judge for holistic and unbiased assessment. Through extensive experiments, we demonstrate that PersonaArena enables rigorous evaluation and enhancement of LLMs' role-playing capabilities, advancing the development of more authentic and socially adept AI agents.

Abstract

Large Language Model-Based Role-Play Simulation

Explore similar work

PersonaArena: Dynamic Simulation for Evaluating and Enhancing Persona-Level Role-Playing in Large Language Models

Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations

From Memory to Behavior: A Behavior-Aware Role-Playing Framework for Social Media Influencers