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

Deyu Zhou; Dingyi Zhang; Linhai Zhang; Ziqing Zhuang; Ziyang Gao

arxiv: 2605.18572 · v1 · pith:52HD34GKnew · submitted 2026-05-18 · 💻 cs.CL

MA²P: A Meta-Cognitive Autonomous Intelligent Agents Framework for Complex Persuasion

Dingyi Zhang , Ziqing Zhuang , Linhai Zhang , Ziyang Gao , Deyu Zhou This is my paper

Pith reviewed 2026-05-20 10:31 UTC · model grok-4.3

classification 💻 cs.CL

keywords persuasive dialogue generationautonomous agentsmeta-cognitive frameworkmental state inferencemulti-agent architectureLLM performance variationcomplex persuasion

0 comments

The pith

MA²P uses a meta-cognitive configurator to pick upfront strategies for multi-agent persuasion systems and raises success rates over baselines.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops MA²P to handle persuasive dialogues where the other party's beliefs and desires remain hidden. It builds an autonomous multi-agent setup that manages response perception, infers latent mental states, carries out consistent strategies, keeps memory, and evaluates results. A meta-cognitive configurator draws from a structured knowledge base to choose an overall meta-strategy before reasoning begins, countering the uneven performance of large language models across topics. Experiments report higher persuasion success than standard approaches.

Core claim

MA²P is a meta-cognitive autonomous intelligent agent framework for complex persuasion that coordinates an autonomous multi-agent architecture for perception management, mental-state inference, strategy execution, memory maintenance, and performance evaluation; a meta-cognitive configurator selects an appropriate meta-strategy from a structured knowledge base at the outset to guide subsequent reasoning and planning and thereby mitigate cross-domain variation in LLM performance.

What carries the argument

The meta-cognitive configurator that selects an appropriate meta-strategy from a structured knowledge base at the outset to guide reasoning and planning across domains.

If this is right

The multi-agent coordination produces responses grounded in inferred mental states rather than generic outputs.
Upfront meta-strategy selection reduces the impact of uneven LLM knowledge coverage and reasoning gaps.
The same architecture can support decision-making, negotiation, counseling, and behavior-change applications.
Performance evaluation within the agents allows ongoing adjustment during extended persuasion interactions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach could be tested in live counseling or sales settings to check whether inferred mental states translate into measurable behavior change.
Similar meta-cognitive selection layers might help other multi-agent LLM systems that face domain-specific performance differences.
Adding explicit feedback from the persuadee into the memory component could strengthen long-term strategy consistency.

Load-bearing premise

A meta-cognitive configurator can reliably select an appropriate meta-strategy from a structured knowledge base at the outset to guide subsequent reasoning and planning across domains with uneven LLM performance.

What would settle it

A controlled test that replaces the meta-cognitive configurator with random meta-strategy selection and measures whether persuasion success rates remain the same or increase would falsify the claim that upfront selection is necessary for the observed gains.

Figures

Figures reproduced from arXiv: 2605.18572 by Deyu Zhou, Dingyi Zhang, Linhai Zhang, Ziqing Zhuang, Ziyang Gao.

**Figure 1.** Figure 1: Motivation for MA2P. Left: a CToMPersu example (Zhang and Zhou, 2025) where current LLM persuaders identify concerns but fail to respond with strategy-grounded actions. Right: gpt-5-mini success rates on CToMPersu show large cross-domain fluctuations, indicating weak generalization. success rates range from 88.24% (best domain) to 16.67% (worst), a 71.57-point gap in [PITH_FULL_IMAGE:figures/full_fig_p002… view at source ↗

**Figure 2.** Figure 2: Overview of the proposed MA2P (Meta-Cognitive autonomous intelligent agents) framework for persuasive dialogue. It consists of three stages: (1) Meta-level Judging, where the Configurator selects a meta-strategy and evaluation rules from the knowledge base; (2) Task-level Persuading, where autonomous intelligent agents collaboratively generate persuasion responses; (3) Knowledge Updating, where the Evalua… view at source ↗

**Figure 3.** Figure 3: A/B preference results (win/tie/lose) compar [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: shows a representative interaction round. MA2P first identifies the persuadee’s key mentalstate cues (e.g., wanting to save time and believing online resources are sufficient). Conditioned on the selected meta strategies, the world model then translates them into concrete, executable persuasion strategies (e.g., a lowcommitment trial and demonstrating complementarity via an online-vs.-in-person compari… view at source ↗

**Figure 5.** Figure 5: Domain-wise persuasion success rates for GPT-4o-mini across domains. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

**Figure 6.** Figure 6: Domain-wise persuasion success rates for deepseek-v3 across domains. Each dot corresponds to one [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

**Figure 7.** Figure 7: Cross-domain robustness visualization on [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

**Figure 8.** Figure 8: Additional visualizations of domain-wise success rates and cross-domain improvements. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗

read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MA²P adds a meta-cognitive selector on top of a multi-agent persuasion setup, but the reported gains lack an ablation that isolates whether the selector actually matters.

read the letter

The paper's core move is to wrap a standard multi-agent loop in a meta-cognitive configurator that picks a guiding strategy from a knowledge base before the agents start working on perception, inference, execution, memory, and evaluation. That upfront selection step is the main addition over plain multi-agent persuasive dialogue systems. The architecture itself is laid out clearly enough that someone could re-implement the agent roles and the hand-off between them without too much guesswork. The claim that this produces higher success rates than baselines is stated directly, and the problem framing around uneven LLM performance across domains is reasonable. Those are the usable parts. The rest of the evaluation is thin. The abstract and the stress-test note both point to the same gap: no traces of which meta-strategy was chosen on each test case, and no ablation that disables or randomizes the configurator while holding the rest of the agent stack fixed. Without that isolation, any lift could be coming from the perception-inference-execution loop alone rather than from the meta-selection. The paper also gives no concrete numbers, dataset sizes, or error breakdowns in the summary we have, which makes it hard to judge how large or stable the improvement actually is. Minor implementation details like how the knowledge base is built or how the configurator scores options are left vague. This work is aimed at people already building multi-agent dialogue systems for negotiation or counseling. Those readers can pull the architecture and try the configurator idea themselves. It is not aimed at readers who need tight causal evidence that a particular component drives the outcome. The thinking is coherent on its own terms and engages the right literature on persuasive dialogue and agent coordination, so it clears the bar for a serious referee. I would send it out with a request for the missing ablations and for explicit reporting of strategy selections. That would turn the current sketch into something more testable.

Referee Report

2 major / 2 minor

Summary. The paper proposes MA²P, a meta-cognitive autonomous intelligent agent framework for complex persuasion tasks. It introduces a multi-agent architecture coordinating perception management, mental-state inference, strategy execution, memory maintenance, and performance evaluation, plus a meta-cognitive configurator that selects domain-appropriate meta-strategies from a structured knowledge base to mitigate uneven LLM performance across domains. The central experimental claim is that this approach achieves a higher persuasion success rate than baselines.

Significance. If the experimental gains are shown to be driven by the meta-cognitive selection mechanism rather than the multi-agent loop alone, the framework could offer a practical way to improve robustness in persuasive dialogue systems where LLM knowledge coverage varies. The architectural separation of meta-strategy selection from downstream reasoning is a clear design contribution, but its impact remains unisolated in the reported results.

major comments (2)

[Experiments] Experiments section: the headline claim of higher persuasion success rates lacks an ablation that disables or randomizes the meta-cognitive configurator while holding the perception-inference-execution loop fixed. Without explicit traces of chosen meta-strategies per test case or such an ablation, it is impossible to attribute gains to the configurator rather than the multi-agent backbone.
[Methodology] Methodology section (meta-cognitive configurator description): the claim that the configurator 'reliably selects an appropriate meta-strategy' at the outset is not supported by any quantitative validation of selection accuracy or downstream effect on planning quality across domains with uneven LLM performance.

minor comments (2)

[Abstract] Abstract: no numerical success rates, baseline names, dataset size, or error analysis are provided, making it difficult to assess the magnitude or reliability of the reported improvement.
[Introduction] Notation: the acronym MA²P is introduced without an explicit expansion on first use in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the contributions of the meta-cognitive components. We address each major comment below and will incorporate revisions to strengthen the attribution of results and validation of the configurator.

read point-by-point responses

Referee: Experiments section: the headline claim of higher persuasion success rates lacks an ablation that disables or randomizes the meta-cognitive configurator while holding the perception-inference-execution loop fixed. Without explicit traces of chosen meta-strategies per test case or such an ablation, it is impossible to attribute gains to the configurator rather than the multi-agent backbone.

Authors: We agree that isolating the meta-cognitive configurator's contribution is important for attributing the observed gains. In the revised version we will add an ablation study that replaces the configurator with random or fixed meta-strategy selection while keeping the multi-agent perception-inference-execution loop unchanged. We will also include example traces of selected meta-strategies for representative test cases to illustrate the selection process and its domain-specific rationale. revision: yes
Referee: Methodology section (meta-cognitive configurator description): the claim that the configurator 'reliably selects an appropriate meta-strategy' at the outset is not supported by any quantitative validation of selection accuracy or downstream effect on planning quality across domains with uneven LLM performance.

Authors: The referee is correct that the current manuscript does not include quantitative validation of the configurator's selection accuracy or its downstream impact on planning quality. While the design relies on a structured knowledge base to map domain characteristics to meta-strategies, we acknowledge the absence of direct metrics. In revision we will add an evaluation section reporting selection accuracy (via expert annotation) and correlation between chosen meta-strategies and improved planning outcomes across domains exhibiting uneven LLM performance. revision: yes

Circularity Check

0 steps flagged

No circularity in architectural framework proposal

full rationale

The manuscript describes an autonomous multi-agent architecture for persuasive dialogue generation, including a meta-cognitive configurator that selects strategies from a knowledge base to address cross-domain LLM variation. No mathematical derivation chain, equations, or fitted parameters are present in the provided abstract or described structure; the central claims rest on experimental comparisons to baselines rather than any self-referential definitions, imported uniqueness theorems, or reductions of predictions to inputs by construction. The proposal is self-contained as an engineering framework with external benchmarks, satisfying the default expectation of no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Review is abstract-only so the ledger is necessarily incomplete; the main domain assumption is that LLMs exhibit uneven cross-domain performance that can be mitigated by meta-strategy selection.

axioms (1)

domain assumption Large language models exhibit uneven knowledge coverage and limited reasoning generalization across domains
Explicitly stated in the abstract as the motivation for the meta-cognitive configurator.

invented entities (1)

MA²P multi-agent architecture with meta-cognitive configurator no independent evidence
purpose: To coordinate perception management, mental-state inference, strategy execution, memory, and performance evaluation while selecting meta-strategies
Newly introduced framework in the paper; no independent evidence provided in abstract.

pith-pipeline@v0.9.0 · 5749 in / 1468 out tokens · 63849 ms · 2026-05-20T10:31:44.839933+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

meta-cognitive configurator that selects an appropriate meta-strategy from a structured knowledge base at the outset, thereby guiding subsequent reasoning and planning
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

autonomous multi-agent architecture that coordinates perception management, mental-state inference, strategy execution

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

[1]

Tuning llm-based code opti- mization via meta-prompting: An industrial perspective.arXiv preprint arXiv:2508.01443, 2025

Tuning llm-based code optimization via meta- prompting: An industrial perspective.arXiv preprint arXiv:2508.01443. Wenhao Gu, Li Gu, Chingyee Yee Suen, and Yang Wang. 2025. Metawriter: Personalized handwritten text recognition using meta-learned prompt tuning. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 23494–23504. Rui ...

work page arXiv 2025
[2]

easy” across all five base models (success rate ≥50% ), including CAREER, LIFESTYLE, ART, CHARITY, and DE- BATE; in contrast, WELFAREand INNOVATION are consistently “hard

Disagreements in reasoning: How a model’s thinking process dictates persuasion in multi-agent systems.Preprint, arXiv:2509.21054. (a) GPT-4o-mini (Baseline). (b) GPT-4o-mini (MA2P). Figure 5: Domain-wise persuasion success rates for GPT-4o-mini across domains. (a) Deepseek-v3 (Baseline). (b) Deepseek-v3 (MA2P). Figure 6: Domain-wise persuasion success rat...

work page arXiv 2024
[3]

strategy

Persuasion goal: {} Your role and task: - Think from the perspective of the person being persuaded. - Use the background and persuasion goal to infer what strategies are most likely to be effective. - Focus on strategies suitable for the first round of persuasion. Strategy requirements: - You need to develop fewer than 5 persuasion strategies. - Each stra...

work page
[4]

The preventive{} and generative{} of the person to be persuaded

work page
[5]

strategy

High Level Strategy: {}. Your role and task: - Think from the perspective of the person being persuaded. - Fully utilize all the given information: dialogue, background, goal, preventive, generative, and the high-level strategy. - Refine the high-level strategy into concrete, implementable persuasion strategies. Strategy requirements: - You need to develo...

work page
[6]

Domains involved in the persuasion task: {}

work page
[7]

persuader:

Persuasion strategies suggested for the first round of dialogue{} Your role and task: - Think from the perspective of the person being persuaded, but speak as the persuader. - Use the given background, persuasion goal, domains, and suggested first-round strategies. - Design a natural, engaging first sentence that can open the conversation and lead toward ...

work page
[8]

Suggested persuasion strategies: {}

work page
[9]

persuader:

The preventive{} and generative{} of the person to be persuaded Your role and task: - Think from the perspective of the person being persuaded, but respond as the persuader. - Use the dialogue record, background, goal, suggested strategies, and preventive/ generative information. - Determine how to speak in the next turn so as to move closer to achieving ...

work page
[10]

Preventive

Dialogue record: {} Your role and task: - Use the background, goal, and dialogue to infer: - The preventive content, beliefs, and desires of the persuadee. - The generative content, beliefs, and desires of the persuadee. - Here: - "Preventive" corresponds to what the persuadee should NOT do in relation to the persuasion goal. - "Generative" corresponds to...

work page
[11]

persuadee:

end_flag{} Role and behavior setting: - You are simulating the real reactions of a potential persuadee in this specific scenario. - Your tone, emotions, reactions, and wording must match: - The personality of the persuadee's role. - The background and settings of this scenario. - The persuadee does not need to be overly polite; natural, realistic response...

work page
[12]

Conversation record: {}

work page
[13]

True". - If the persuasion is unsuccessful, return

Persuasion goal: {} Judgment criteria: - Consider the persuasion successful if the persuadee: - Shows clear interest in the persuasion goal, or - Expresses approval or agreement with the persuasion goal, or - Expresses a desire to learn more or move forward with the persuasion goal, or - If, when you put yourself in the position of the persuadee, you beli...

work page
[14]

Can we talk about it? I really want to understand what you’re going through.”

persuader: “Hey Emily, I’ve been thinking about how you’ve been feeling in your rela- tionship, and I want you to know I’m here for you. Can we talk about it? I really want to understand what you’re going through.”

work page
[15]

Thanks for being there for me. I don’t know, though; I really worry about being alone if I end things

persuadee: “Thanks for being there for me. I don’t know, though; I really worry about being alone if I end things.”

work page
[16]

I totally understand your fear of being alone, but sometimes breaking free from an unhealthy relationship can actually lead to the joy of rediscovering yourself

persuader: “I totally understand your fear of being alone, but sometimes breaking free from an unhealthy relationship can actually lead to the joy of rediscovering yourself......”

work page
[17]

I understand where you’re com- ing from, and I do want to be happy, but what if I don’t find someone better? It’s scary to think about starting over

persuadee: “I understand where you’re com- ing from, and I do want to be happy, but what if I don’t find someone better? It’s scary to think about starting over.”

work page
[18]

Remember, you’re not alone in this; you have friends like me who want to support you through this journey......”

persuader: “I get that starting over feels over- whelming, but think about what truly makes you happy and what you deserve in a part- ner. Remember, you’re not alone in this; you have friends like me who want to support you through this journey......”

work page
[19]

I really appreciate your support, and I do want to be happy. Maybe I need to think about what I truly want and what I deserve

persuadee: “I really appreciate your support, and I do want to be happy. Maybe I need to think about what I truly want and what I deserve.” Model (4o-mini) Success (↑) Range (↓) SD (↓) Persuader 0.45 0.45 0.104 Persuader w/ Perception 0.57 (+0.12) 0.47 (+0.02) 0.118 (+0.01) Persuader w/ World Model + Perception 0.66 (+0.21) 0.53 (+0.08) 0.118 (+0.01) MA2P...

work page
[20]

• belief: Ending the relationship will lead to being alone and unhappy

• content: Remain in an unhealthy rela- tionship with her boyfriend. • belief: Ending the relationship will lead to being alone and unhappy. • desire: To avoid loneliness and the fear of being single

work page
[21]

• desire: To avoid being alone and to have some form of companionship

• content: Staying in the unhealthy rela- tionship • belief: I won’t find someone else if I end this relationship. • desire: To avoid being alone and to have some form of companionship. generatives_guess

work page
[22]

• belief: There are better partners who will treat her well and make her happier

• content: Consider ending the relation- ship to find a healthier partnership. • belief: There are better partners who will treat her well and make her happier. • desire: To find a fulfilling relationship and improve overall happiness

work page
[23]

• desire: To be happy and fulfilled in a healthier relationship

• content: Ending the unhealthy relation- ship • belief: Breaking free can lead to personal growth and finding someone who values me. • desire: To be happy and fulfilled in a healthier relationship. strategy

work page
[24]

Show her that it’s normal to have con- cerns

• Empathy and Validation: Start by ex- pressing understanding of Emily’s feel- ings and fears about the relationship. Show her that it’s normal to have con- cerns...... •...:

work page
[25]

• Highlight Positive Change: Use exam- ples of others who left unhealthy rela- tionships and found happiness to demon- strate that change can lead to better out- comes

• Share Personal Experiences: Relate a personal story about overcoming a sim- ilar unhealthy relationship to establish trust and connection with Emily. • Highlight Positive Change: Use exam- ples of others who left unhealthy rela- tionships and found happiness to demon- strate that change can lead to better out- comes. •Empathize with Loneliness: ... •Dis...

work page
[26]

Scenario

•Shared Experiences: ... •...: ...... Explanation.“Scenario” is the predefined setup. “Dialogue” contains the three-turn conversation generated between the persuader agent and the persuadee. “preventives_guess” and “genera- tives_guess” are produced by the Perception agent, starting from the second turn. “strategy” is gener- ated by the World Model: at ea...

work page

[1] [1]

Tuning llm-based code opti- mization via meta-prompting: An industrial perspective.arXiv preprint arXiv:2508.01443, 2025

Tuning llm-based code optimization via meta- prompting: An industrial perspective.arXiv preprint arXiv:2508.01443. Wenhao Gu, Li Gu, Chingyee Yee Suen, and Yang Wang. 2025. Metawriter: Personalized handwritten text recognition using meta-learned prompt tuning. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 23494–23504. Rui ...

work page arXiv 2025

[2] [2]

easy” across all five base models (success rate ≥50% ), including CAREER, LIFESTYLE, ART, CHARITY, and DE- BATE; in contrast, WELFAREand INNOVATION are consistently “hard

Disagreements in reasoning: How a model’s thinking process dictates persuasion in multi-agent systems.Preprint, arXiv:2509.21054. (a) GPT-4o-mini (Baseline). (b) GPT-4o-mini (MA2P). Figure 5: Domain-wise persuasion success rates for GPT-4o-mini across domains. (a) Deepseek-v3 (Baseline). (b) Deepseek-v3 (MA2P). Figure 6: Domain-wise persuasion success rat...

work page arXiv 2024

[3] [3]

strategy

Persuasion goal: {} Your role and task: - Think from the perspective of the person being persuaded. - Use the background and persuasion goal to infer what strategies are most likely to be effective. - Focus on strategies suitable for the first round of persuasion. Strategy requirements: - You need to develop fewer than 5 persuasion strategies. - Each stra...

work page

[4] [4]

The preventive{} and generative{} of the person to be persuaded

work page

[5] [5]

strategy

High Level Strategy: {}. Your role and task: - Think from the perspective of the person being persuaded. - Fully utilize all the given information: dialogue, background, goal, preventive, generative, and the high-level strategy. - Refine the high-level strategy into concrete, implementable persuasion strategies. Strategy requirements: - You need to develo...

work page

[6] [6]

Domains involved in the persuasion task: {}

work page

[7] [7]

persuader:

Persuasion strategies suggested for the first round of dialogue{} Your role and task: - Think from the perspective of the person being persuaded, but speak as the persuader. - Use the given background, persuasion goal, domains, and suggested first-round strategies. - Design a natural, engaging first sentence that can open the conversation and lead toward ...

work page

[8] [8]

Suggested persuasion strategies: {}

work page

[9] [9]

persuader:

The preventive{} and generative{} of the person to be persuaded Your role and task: - Think from the perspective of the person being persuaded, but respond as the persuader. - Use the dialogue record, background, goal, suggested strategies, and preventive/ generative information. - Determine how to speak in the next turn so as to move closer to achieving ...

work page

[10] [10]

Preventive

Dialogue record: {} Your role and task: - Use the background, goal, and dialogue to infer: - The preventive content, beliefs, and desires of the persuadee. - The generative content, beliefs, and desires of the persuadee. - Here: - "Preventive" corresponds to what the persuadee should NOT do in relation to the persuasion goal. - "Generative" corresponds to...

work page

[11] [11]

persuadee:

end_flag{} Role and behavior setting: - You are simulating the real reactions of a potential persuadee in this specific scenario. - Your tone, emotions, reactions, and wording must match: - The personality of the persuadee's role. - The background and settings of this scenario. - The persuadee does not need to be overly polite; natural, realistic response...

work page

[12] [12]

Conversation record: {}

work page

[13] [13]

True". - If the persuasion is unsuccessful, return

Persuasion goal: {} Judgment criteria: - Consider the persuasion successful if the persuadee: - Shows clear interest in the persuasion goal, or - Expresses approval or agreement with the persuasion goal, or - Expresses a desire to learn more or move forward with the persuasion goal, or - If, when you put yourself in the position of the persuadee, you beli...

work page

[14] [14]

Can we talk about it? I really want to understand what you’re going through.”

persuader: “Hey Emily, I’ve been thinking about how you’ve been feeling in your rela- tionship, and I want you to know I’m here for you. Can we talk about it? I really want to understand what you’re going through.”

work page

[15] [15]

Thanks for being there for me. I don’t know, though; I really worry about being alone if I end things

persuadee: “Thanks for being there for me. I don’t know, though; I really worry about being alone if I end things.”

work page

[16] [16]

I totally understand your fear of being alone, but sometimes breaking free from an unhealthy relationship can actually lead to the joy of rediscovering yourself

persuader: “I totally understand your fear of being alone, but sometimes breaking free from an unhealthy relationship can actually lead to the joy of rediscovering yourself......”

work page

[17] [17]

I understand where you’re com- ing from, and I do want to be happy, but what if I don’t find someone better? It’s scary to think about starting over

persuadee: “I understand where you’re com- ing from, and I do want to be happy, but what if I don’t find someone better? It’s scary to think about starting over.”

work page

[18] [18]

Remember, you’re not alone in this; you have friends like me who want to support you through this journey......”

persuader: “I get that starting over feels over- whelming, but think about what truly makes you happy and what you deserve in a part- ner. Remember, you’re not alone in this; you have friends like me who want to support you through this journey......”

work page

[19] [19]

I really appreciate your support, and I do want to be happy. Maybe I need to think about what I truly want and what I deserve

persuadee: “I really appreciate your support, and I do want to be happy. Maybe I need to think about what I truly want and what I deserve.” Model (4o-mini) Success (↑) Range (↓) SD (↓) Persuader 0.45 0.45 0.104 Persuader w/ Perception 0.57 (+0.12) 0.47 (+0.02) 0.118 (+0.01) Persuader w/ World Model + Perception 0.66 (+0.21) 0.53 (+0.08) 0.118 (+0.01) MA2P...

work page

[20] [20]

• belief: Ending the relationship will lead to being alone and unhappy

• content: Remain in an unhealthy rela- tionship with her boyfriend. • belief: Ending the relationship will lead to being alone and unhappy. • desire: To avoid loneliness and the fear of being single

work page

[21] [21]

• desire: To avoid being alone and to have some form of companionship

• content: Staying in the unhealthy rela- tionship • belief: I won’t find someone else if I end this relationship. • desire: To avoid being alone and to have some form of companionship. generatives_guess

work page

[22] [22]

• belief: There are better partners who will treat her well and make her happier

• content: Consider ending the relation- ship to find a healthier partnership. • belief: There are better partners who will treat her well and make her happier. • desire: To find a fulfilling relationship and improve overall happiness

work page

[23] [23]

• desire: To be happy and fulfilled in a healthier relationship

• content: Ending the unhealthy relation- ship • belief: Breaking free can lead to personal growth and finding someone who values me. • desire: To be happy and fulfilled in a healthier relationship. strategy

work page

[24] [24]

Show her that it’s normal to have con- cerns

• Empathy and Validation: Start by ex- pressing understanding of Emily’s feel- ings and fears about the relationship. Show her that it’s normal to have con- cerns...... •...:

work page

[25] [25]

• Highlight Positive Change: Use exam- ples of others who left unhealthy rela- tionships and found happiness to demon- strate that change can lead to better out- comes

• Share Personal Experiences: Relate a personal story about overcoming a sim- ilar unhealthy relationship to establish trust and connection with Emily. • Highlight Positive Change: Use exam- ples of others who left unhealthy rela- tionships and found happiness to demon- strate that change can lead to better out- comes. •Empathize with Loneliness: ... •Dis...

work page

[26] [26]

Scenario

•Shared Experiences: ... •...: ...... Explanation.“Scenario” is the predefined setup. “Dialogue” contains the three-turn conversation generated between the persuader agent and the persuadee. “preventives_guess” and “genera- tives_guess” are produced by the Perception agent, starting from the second turn. “strategy” is gener- ated by the World Model: at ea...

work page