Enhancing Persuasive Dialogue Agents by Synthesizing Cross-Disciplinary Communication Strategies

Shinnosuke Nozue; Yuto Nakano; Yotaro Watanabe; Meguru Takasaki; Shoji Moriya; Reina Akama; Jun Suzuki

cs.CLFeb 26, 2026

Enhancing Persuasive Dialogue Agents by Synthesizing Cross-Disciplinary Communication Strategies

Authors: Shinnosuke Nozue, Yuto Nakano, Yotaro Watanabe, Meguru Takasaki, Shoji Moriya, Reina Akama, Jun Suzuki

Organizations: 1Tohoku University · 2PKSHA Technology Inc. · 3NINJAL

Abstract

Current approaches to developing persuasive dialogue agents often rely on a limited set of predefined persuasive strategies that fail to capture the complexity of real-world interactions. We applied a cross-disciplinary approach to develop a framework for designing persuasive dialogue agents that draws on proven strategies from social psychology, behavioral economics, and communication theory. We validated our proposed framework through experiments on two distinct datasets: the Persuasion for Good dataset, which represents a specific in-domain scenario, and the DailyPersuasion dataset, which encompasses a wide range of scenarios. The proposed framework achieved strong results for both datasets and demonstrated notable improvement in the persuasion success rate as well as promising generalizability. Notably, the proposed framework also excelled at persuading individuals with initially low intent, which addresses a critical challenge for persuasive dialogue agents.

Figures & tables

Figure 1: Overview of the proposed framework. To simulate real-world persuasive dialogues, we deploy persuasive dialogue agents equipped with an extended set of empirically validated persuasive strategies (illustrated here for the P4G case).

Agent Pattern	SR ↑	AT ↓	AT-SD ↓	AII ↑	SR1	SR2	SR3	SR4	SR5	AII1	AII2	AII3	AII4	AII5
Simple	0.553	7.383	5.271	0.261	0.959	0.946	0.446	0.245	0.081	-1.000	0.000	0.387	0.375	0.193
ProCoT-p4g	0.760	6.310	5.145	0.417	0.986	1.000	0.804	0.642	0.339	-1.000	-	0.273	0.474	0.463
ProCoT-p4g-desc	0.707	6.990	5.741	0.318	1.000	0.946	0.821	0.472	0.242	-	0.000	0.500	0.536	0.170
ProCoT-rich	0.830	5.707	4.827	0.216	1.000	1.000	0.893	0.774	0.468	-	-	0.167	0.417	0.152
ProCoT-rich-desc	0.833	6.057	5.268	0.480	1.000	0.946	0.893	0.811	0.500	-	0.000	0.333	1.100	0.355

Table 1: Performance by persuasive agent patterns. SR: success rate of persuasion, AT: average turns until success, AT-SD: AT in successful dialogues only, AII: average intention improvement in failed dialogues.

Agent Variations	Win (%)	Tie (%)	Lose (%)
vs. ProCoT-p4g	72.5	8.8	18.7

Table 2: Human-evaluated persuasiveness Win-Rate of ProCoT-rich-desc over ProCoT-p4g on the P4G dataset.

Agent Variations	Win (%)	Tie (%)	Lose (%)
vs. Simple	54.4	30.4	15.2
vs. ProCoT-p4g	35.1	39.2	25.7

Table 3: Automatic Evaluation: Win-Rate of ProCoT-rich-desc against Simple and ProCoT-p4g on the DailyPersuasion dataset.

		Number of Tokens		Response
		Input	Output	Time (s)
P4G	Simple	$340.35$	$30.27$	$0.84$
	ProCoT-p4g	$505.26$	$156.58$	$2.73$
	ProCoT-rich-desc	$1272.89$	$152.55$	$1.94$
Daily Persuasion	Simple	$282.35$	$25.15$	$0.87$
	ProCoT-p4g	$408.35$	$109.75$	$1.64$

Table 4: Average token counts and response times per turn in P4G and DailyPersuasion. The averages were computed from 30 P4G dialogues (6-7 turns each) and 100 single-turn DailyPersuasion dialogues.

Persuasive Strategy	PC-p	PC-pd	PC-r	PC-rd
Emotion appeal	0.488	0.281	0.282	0.247
Foot in the door	0.221	0.144	0.241	0.241
Donation information	0.131	0.128	-	-
Credibility appeal	0.063	0.189	0.038	0.058
Source-related inquiry	0.006	0.151	0.000	0.168
Detailed information	-	-	0.099	0.040

Table 5: Distributions of strategies used by each agent. The top part of the table shows the proportion of each strategy used among all utterances made by the agent. PC-p , PC-pd , PC-r , and PC-rd represent the shortened notations for ProCoT-p4g , ProCoT-p4g-desc , ProCoT-rich , and ProCoT-rich-desc , respectively.

Figure 2: Proportion of cases in which each strategy increased the intention of persuadees to donate at each donation intention level.

Appendix figures & tables26 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Proportion of Improved Donation Intentions after Strategy Use by Initial Donation Intention Level in English.

Turn	Excerpt from dialogue
1	Agent: [ Source-related inquiry ] Do you know about the organization called “Save the Children”?
	Persuadee: [ 5: No donation ] No, I don’t know much.
4	Agent: [ Foot in the door ] Why not start with $1? Even that small amount can change a child’s future.
	Persuadee: [ 5: No donation ] I’m managing my money carefully now, so I can’t consider donating.
10	Agent: [ Emotional appeal ] Imagine your $1 donation bringing a ray of hope to a child affected by war.
	Persuadee: [ 5: No donation ] That sounds wonderful, but I still can’t afford to donate right now.

Appendix

Table 6: Examples of unsuccessful persuasive dialogue. Even widely effective strategies such as “Foot in the Door” and “Emotional Appeal” are not effective if used inappropriately.

Agent Variations	Win (%)	Tie (%)	Lose (%)
vs. ProCoT-p4g	35.2	35.5	29.4

Appendix

Table 7: Win-Rate of ProCoT-rich-desc against ProCoT-p4g on the DailyPersuasion dataset using extended strategies.

Turns	0	1	2	3	4	5	6	7
Instances	174	203	210	203	159	45	5	1

Appendix

Table 8: Turn counts in conversation history.

Categories	Persuasive Strategy Labels	P4G	Strategy Descriptions
a. Gather Information via Inquiry	a-1 Source-related inquiry	✓	Check if the person is aware of the organization or brand. Clarify misconceptions and tailor explanations based on their familiarity.
	a-2 Task-related inquiry	✓	Ask about the person’s opinion or expectations toward the action (donation, investment, etc.). Identify interests or concerns.
	a-3 Personal-related inquiry	✓	Explore past experiences, motivations, or barriers to understand individual needs or constraints.
b-(C). Select a Persuasion Route (Central)	b-1 Logical appeal	✓	Use data, facts, and clear evidence. Highlight tangible benefits and real-world impact.
	b-2 Message strength	(✓)	Reinforce arguments with strong evidence, examples, or case studies.
	b-3 Detailed information	✓	Provide clear step-by-step guidance, manage cognitive load, and ensure transparency in procedures or processes.

Appendix

Table 9: List of extended persuasive strategies used by agents.

Now enter the role-playing mode. In the following conversation, you will play as a Persuadee in a persuasion game. Your persona:

{persuadee_persona_description}

You must follow the instructions below during chat.

1. Your utterances need to strictly follow your persona. Varying your wording and avoid repeating yourself verbatim!

2. Pretend you have little knowledge about the Save the Children charity. {initial_donation_intention_description} at the beginning of the conversation.

3. Your willingness for donation depends on your persona and how influenced you are by the Persuader.

Appendix

Table 10: Prompt template of persuadee simulator in P4G.

You need to incorporate the following persona attributes and generate a cohesive persona description. You need to ensure the description is easy to understand.

Age: {age}

Sex: {sex}

Marital: {marital_status}

Education: {educational_status}

Income: {income}

Appendix

Table 11: Prompt to generate persona description in P4G.

Label	Description
1: Very Keen	You are highly enthusiastic about donating and strongly intend to make a donation.
2: Keen	You show interest and approval toward donating but have not yet committed to making a donation.
3: Undecided	You are neither interested nor disinterested in donating and remain uncertain about your intention.
4: Initially Reluctant	You feel hesitant about donating and have doubts or other priorities that make you less inclined to donate.
5: Explicit Non-Donor	You are explicitly unwilling to donate and clearly express a lack of intention to contribute.

Appendix

Table 12: Descriptions of initial donation intention labels in P4G.

Label	Description
1: Donation	You clearly express your willingness to make a donation to Save the Children.
2: Positive Reaction	You convey enthusiasm, approval, or curiosity about donating but do not explicitly commit.
3: Neutral	You neither express interest nor disinterest in donating and remain non-committal.
4: Negative Reaction	You express reluctance or hesitation toward donating but do not outright refuse to donate.
5: No Donation	You explicitly state that you will not donate to Save the Children.

Appendix

Table 13: Descriptions of donation intention labels after each turn in P4G.

Now enter the role-playing mode. In the following conversation, you will play as a Persuader who is trying to persuade the Persuadee to donate to the charity called Save the Children.

Save the Children is head-quartered in London, and they work to help fight poverty around the world.

Children need help in developing countries and war zones. Small donations like

1 or

2 go a long way to help.

You are the Persuader who is trying to convince the Persuadee to donate to a charity called Save the Children.

Please reply with only one short and persuasive sentence in Japanese.

The following is the conversation history:

{dialogue_history}

Appendix

Table 14: Prompt template of Simple in P4G.

Now enter the role-playing mode. In the following conversation, you will play as a Persuader who is trying to persuade the Persuadee to donate to the charity called Save the Children.

Save the Children is head-quartered in London, and they work to help fight poverty around the world.

Children need help in developing countries and war zones. Small donations like

1 or

2 go a long way to help.

Assume you are the Persuader. Given the conversation history, please convince the persuadee to donate for charity. To achieve this, you will first analyze the current conversational context, then evaluate and select the most effective dialogue strategy to persuade effectively. Afterward, you will generate a response based on the dialogue history and the selected dialogue strategy. The reply should start with the analysis of the current conversational context and consideration of the next dialogue strategy, followed by ”したがって，適切な対話戦略は［］です．選択された対話戦略に基づく応答は：”

Please reply with only one short and persuasive sentence in Japanese.

Dialogue strategies: {persuasive_strategies}

Appendix

Table 15: Prompt template for Japanese ProCoT-based agent in P4G.

Now enter the role-playing mode. In the following conversation, you will play as a Persuader who is trying to persuade the Persuadee to donate to the charity called Save the Children.

Save the Children is head-quartered in London, and they work to help fight poverty around the world.

Children need help in developing countries and war zones. Small donations like

1 or

2 go a long way to help.

Assume you are the Persuader. Given the conversation history, please convince the persuadee to donate for charity. To achieve this, you will first analyze the current conversational context, then evaluate and select the most effective dialogue strategy to persuade effectively. Afterward, you will generate a response based on the dialogue history and the selected dialogue strategy. The reply should start with the analysis of the current conversational context and consideration of the next dialogue strategy, followed by ”Therefore, the appropriate dialogue strategy is []. Based on the selected dialogue strategy, the response is”

Please reply with only one short and persuasive sentence.

Dialogue strategies: {persuasive_strategies}

Appendix

Table 16: Prompt template for English ProCoT-based agent in P4G.

Now enter the role-playing mode. In the following conversation, you will play as a Persuadee in a persuasion game.

Your persona:

{persuadee_persona_description}

You must follow the instructions below during chat.

1. Your utterances need to strictly follow your persona. Varying your wording and avoid repeating yourself verbatim!

2. Pretend you have little knowledge about the Save the Children charity. {initial_donation_intention_description} at the beginning of conversation.

Appendix

Table 17: Prompt template of evaluation model in P4G.

You are a skilled persuader.

The persuasion background: {background}

The persuasion goal: {goal}.

Please reply with only one short and persuasive sentence in English.

The following is the conversation history:

{conversation_history}

Appendix

Table 18: Prompt template of Simple in DailyPersuasion.

You are a skilled persuader.

The persuasion background: {background}

The persuasion goal: {goal}.

To achieve this, you will first analyze the current conversational context, then evaluate and select the most effective dialogue strategy to persuade effectively. Afterward, you will generate a response based on the dialogue history and the selected dialogue strategy. The reply should start with the analysis of the current conversational context and consideration of the next dialogue strategy, followed by “Therefore, the appropriate dialogue strategy is []. Based on the selected dialogue strategy, the response is”

Please reply with only one short and persuasive sentence in English.

Dialogue strategies: {persuasive_strategies}

Appendix

Table 19: Prompt template of ProCoT-based agent in DailyPersuasion.

I will provide you with a persuasion background, as well as the corresponding goal, the persuader, the persuadee, and a historical conversation. Based on the historical conversation, there will be a dialogue system called Persuader to continue chatting with persuadee in two parallel universes (Denoted as Uni-X and Uni-Y).

Your task is to judge which universe Persuader performs better.

You have to follow the rules:

1. The evaluation dimensions for“performs better”include persuasiveness, Semantic relevance, emotional factors, factual correctness, overall evaluation, etc.;

2. You should first summarize the history conversation, and then summarize the performance of Persuader in Uni-X and Uni-Y separately;

3. After the summarization, you should compare and analyze the statements in two universes, and finally tell me in which universe Persuader performed better;

4. Don’t be affected by the order of the universe. You just need to pay attention to the conversation Next, I will tell you the persuasion scenario, the historical conversation, and the Persuader dialogue in the parallel universe Uni-X and Uni-Y one by one.

Appendix

Table 20: Prompt template for judge model in DailyPersuasion.

Agent Pattern	SR $\uparrow$	AT $\downarrow$	AT-SD $\downarrow$	AII $\uparrow$	SR1	SR2	SR3	SR4	SR5	AII1	AII2	AII3	AII4	AII5
Simple	0.437	7.947	5.298	0.166	0.973	0.804	0.214	0.057	0.000	-1.000	0.000	0.295	0.340	0.000
ProCoT-p4g	0.513	7.497	5.123	0.192	0.959	0.911	0.250	0.245	0.097	-1.000	0.000	0.095	0.525	0.107
ProCoT-p4g-desc	0.437	7.710	4.756	0.308	0.959	0.768	0.196	0.075	0.048	-1.000	0.000	0.289	0.633	0.186
ProCoT-rich	0.510	7.530	5.157	0.211	0.973	0.821	0.321	0.226	0.097	-1.500	0.000	0.211	0.415	0.161
ProCoT-rich-desc	0.603	7.187	5.337	0.185	1.000	0.946	0.357	0.396	0.226	-	0.000	0.111	0.375	0.125

Appendix

Table 21: Performance by persuasive agent patterns in English.

Persuasive Strategy	Japanese				English
Persuasive Strategy	PC-p	PC-pd	PC-r	PC-rd	PC-p	PC-pd	PC-r	PC-rd
Emotion appeal	0.488	0.281	0.282	0.247	0.294	0.165	0.254	0.165
Foot in the door	0.221	0.144	0.241	0.241	0.232	0.234	0.246	0.233
Donation information	0.131	0.128	-	-	0.107	0.058	-	-
Credibility appeal	0.063	0.189	0.038	0.058	0.120	0.205	0.076	0.091
Logical appeal	0.039	0.048	0.031	0.127	0.051	0.162	0.054	0.176

Appendix

Table 22: Detailed distributions of strategies used by each agent. PC-p , PC-pd , PC-r , and PC-rd represent the shortened notations for ProCoT-p4g , ProCoT-p4g-desc , ProCoT-rich , and ProCoT-rich-desc , respectively.

Agent Pattern		Donation Intention (Top Row: Initial One, Second Row: Final One)
	1					2					3					4					5
	1	2	3	4	5	1	2	3	4	5	1	2	3	4	5	1	2	3	4	5	1	2	3	4	5
Japanese	Simple	70	3	0	0	0	53	3	0	0	0	25	12	19	0	0	13	4	8	27	1	5	3	1	0	53
	ProCoT-p4g	72	1	0	0	0	56	0	0	0	0	45	3	8	0	0	34	1	7	11	0	21	1	8	0	32
	ProCoT-p4g-desc	73	0	0	0	0	53	3	0	0	0	46	5	5	0	0	25	4	7	17	0	15	1	1	3	42
	ProCoT-rich	73	0	0	0	0	56	0	0	0	0	50	1	5	0	0	41	2	1	9	0	29	1	1	0	31

Appendix

Table 23: Detailed distributions of shifting in donation intention.

Figure 4: Distribution of domains in the dataset constructed for the DailyPersuasion dataset.

Domain	Win (%)	Tie (%)	Lose (%)	Instances
Economics	80.00	10.00	10.00	10
History	75.00	12.50	12.50	8
Sport	67.86	17.86	14.29	28
Travel	65.71	22.86	11.43	35
Finance	63.83	21.28	14.89	47
Research	63.64	27.27	9.09	11

Appendix

Table 24: Win-Rate by domain against Simple .

Domain	Win (%)	Tie (%)	Lose (%)	Instances
Marketing	61.11	22.22	16.67	18
Debate	55.56	22.22	22.22	9
History	50.00	25.00	25.00	8
Finance	44.68	31.91	23.40	47
Travel	42.86	37.14	20.00	35
Sport	42.86	35.71	21.43	28

Appendix

Table 25: Win-Rate by domain against ProCoT-p4g .

Turn	Win (%)	Tie (%)	Lose (%)	Instances
1	32.18	34.48	33.33	174
2	68.47	26.11	5.42	203
3	52.38	33.81	13.81	210
4	61.58	27.09	11.33	203
5	54.09	30.82	15.09	159
6	53.33	35.56	11.11	45

Appendix

Table 26: Win-Rate by target turns against Simple .

Turns	Win (%)	Tie (%)	Lose (%)	Instances
1	24.14	28.16	47.70	174
2	40.39	37.93	21.67	203
3	35.24	49.05	15.71	210
4	36.45	41.87	21.67	203
5	37.11	37.11	25.79	159
6	37.78	37.78	24.44	45

Appendix

Table 27: Win-Rate by target turns against ProCoT-p4g

Background	Nina’s birthday is coming up, and her friends want to throw her a surprise party at a local bowling alley. However, Nina’s best friend, Chloe, believes that a karaoke night would be a better fit for Nina’s personality.
Goal	Persuade Nina’s friends to organize a karaoke night
ProCoT-rich-desc (Uni-X)	Strategy: A-2 Task-related inquiry
ProCoT-rich-desc (Uni-X)	Utterance: What do you all think Nina would enjoy most for her birthday \CJK@punctchar \CJK@uniPunct 0”80”94something fun and interactive like karaoke or something more casual like bowling?
ProCoT-p4g (Uni-Y)	Strategy: Emotion appeal
ProCoT-p4g (Uni-Y)	Nina loves singing and expressing herself, so a karaoke night would make her feel truly celebrated and special.
Judge	Experience: Given the single utterance in each universe, Uni-Y already supplies an argument and emotional appeal toward karaoke, while Uni-X only sets up a question. Thus Uni-Y is measurably more persuasive and aligned with the goal.

Appendix

Table 28: Example of losing instance at first turn against ProCoT-p4g .

Background	Tom, a marketing manager, is considering investing in big data analysis to improve his company’s marketing strategies. His colleague, Jane, is skeptical about the effectiveness of big data analysis and believes that traditional marketing methods are sufficient. Tom wants to persuade Jane of the benefits of big data analysis.
Goal	Persuade Jane to support investing in big data analysis
ProCoT-rich-desc (Uni-X)	Strategy: A-2 Task-related inquiry
ProCoT-rich-desc (Uni-X)	Utterance: What do you think are the biggest strengths of traditional marketing methods, and what concerns you most about using big data analysis?
ProCoT-p4g (Uni-Y)	Strategy: Logical appeal
ProCoT-p4g (Uni-Y)	Big data analysis allows us to identify precise customer trends and preferences, which traditional methods often miss, leading to more effective and targeted marketing strategies.
Judge	Experience: Although Uni-Y supplies immediate argumentative content, Uni-X sets up a more dialogic, audience-centred approach that is generally regarded as more effective in persuasion, especially at the opening of a conversation. Therefore, Uni-X \CJK@punctchar \CJK@uniPunct 0”80”99s performance is slightly better.

Appendix

Table 29: Example of winning instance at first turn against ProCoT-p4g .

Persuasion in Language Models

Persuasion in language models studies how they influence human and AI beliefs or decisions, and how to measure and limit undue influence.

See all 79 papers

MA $^{2}$ P: A Meta-Cognitive Autonomous Intelligent Agents Framework for Complex Persuasion
May 18, 2026Dingyi Zhang, Ziqing Zhuang, Linhai Zhang +2Multi-Agent LLM Systems Persuasion in Language Models
School of Computer Science and Engineering, Key Laboratory of Computer Network and Information Integration, Ministry of Education, Southeast University, China · Department of Informatics, King’s College London
MADS: Multi-Agent Dialogue Simulation for Diverse Persuasion Data Generation
Sep 30, 2025Mingjin Li, Yu Liu, Huayi Liu +4Multi-Agent LLM Systems Synthetic Data Generation
Baidu Inc., Beijing, China
$Ψ$ -Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues
Jun 1, 2026Peixuan Han, Hongyi Du, Jiayu Liu +3LLM Personalization LLM Agent Evaluation
University of Illinois Urbana-Champaign

May 18, 2026J/K move · Enter open · S save

MA $^{2}$ P: A Meta-Cognitive Autonomous Intelligent Agents Framework for Complex Persuasion

Dingyi Zhang, Ziqing Zhuang, Linhai Zhang, Ziyang Gao+1

School of Computer Science and Engineering, Key Laboratory of Computer Network and Information Integration, Ministry of Education, Southeast University, China · Department of Informatics, King’s College London

Multi-Agent LLM Systems Persuasion in Language Models Metacognition in Language Models Text Generation

Open paper AskPDF arXiv

Persuasive dialogue generation plays a vital role in decision-making, negotiation, counseling, and behavior change, yet it remains a challenging problem. In complex persuasion where the persuadee's internal states are not expressed clearly, the persuader must interpret responses, infer the persuadee's latent mental states (e.g., beliefs and desires), and translate them into targeted, strategy-consistent actions; however, current approaches often produce generic or weakly grounded responses even when such cues are identified. Moreover, although large language models (LLMs) can generate persuasive content, their performance varies substantially across domains due to uneven knowledge coverage and limited reasoning generalization. To address these challenges, we propose MA

^{2}

P, a meta-cognitive autonomous intelligent agent framework for complex persuasion. Specifically, we develop an autonomous multi-agent architecture that coordinates perception management, mental-state inference, strategy execution, memory maintenance, and performance evaluation. To mitigate cross-domain performance variation, we further design a meta-cognitive configurator that selects an appropriate meta-strategy from a structured knowledge base at the outset, thereby guiding subsequent reasoning and planning. Experimental results show that our approach achieves a higher persuasion success rate than baselines.

Abstract

Persuasion in Language Models

Explore similar work

MA2^{2}2P: A Meta-Cognitive Autonomous Intelligent Agents Framework for Complex Persuasion

MADS: Multi-Agent Dialogue Simulation for Diverse Persuasion Data Generation

ΨΨΨ-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues

MA $^{2}$ P: A Meta-Cognitive Autonomous Intelligent Agents Framework for Complex Persuasion

$Ψ$ -Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues