METRO: Towards Strategy Induction from Expert Dialogue Transcripts for Non-collaborative Dialogues

Haofu Yang; Jiaji Liu; Chen Huang; Faguo Wu; Wenqiang Lei; See-Kiong Ng

cs.CLApr 13, 2026

METRO: Towards Strategy Induction from Expert Dialogue Transcripts for Non-collaborative Dialogues

Authors: Haofu Yang, Jiaji Liu, Chen Huang, Faguo Wu, Wenqiang Lei, See-Kiong Ng

Organizations: School of Artificial Intelligence, Beihang University · Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing · Key Laboratory of Mathematics, Informatics and Behavioral Semantics (LMIB) · Nankai University ·

Abstract

Developing non-collaborative dialogue agents traditionally requires the manual, unscalable codification of expert strategies. We propose \ours, a method that leverages large language models to autonomously induce both strategy actions and planning logic directly from raw transcripts. METRO formalizes expert knowledge into a Strategy Forest, a hierarchical structure that captures both short-term responses (nodes) and long-term strategic foresight (branches). Experimental results across two benchmarks show that METRO demonstrates promising performance, outperforming existing methods by an average of 9%-10%. Our further analysis not only reveals the success behind METRO (strategic behavioral diversity and foresight), but also demonstrates its robust cross-task transferability. This offers new insights into building non-collaborative agents in a cost-effective and scalable way. Our code is available at https://github.com/Humphrey-0125/METRO.

Figures & tables

Figure 1: Overview of Metro . It formalizes expert knowledge within transcripts into a Strategy Forest, representing strategy actions and planning logic in a unified structure.

Methods	P4G		CB			P4G*		CB*
Methods	SR $\uparrow$	AT $\downarrow$	SR $\uparrow$	AT $\downarrow$	SL% $\uparrow$	SR $\uparrow$	AT $\downarrow$	SR $\uparrow$	AT $\downarrow$	SL% $\uparrow$
Standard Deng et al. (2023d)	0.620	4.56	0.185	5.89	0.154	0.333	5.20	0.283	5.65	0.285
Proactive Deng et al. (2023d)	0.660	4.11	0.170	6.03	0.116	0.400	4.67	0.283	5.41	0.198
ProCoT Deng et al. (2023d)	0.680	4.29	0.315	6.46	0.186	0.533	5.13	0.333	6.10	0.282
ICL-AIF Fu et al. (2023)	0.615	3.89	0.325	6.11	0.231	0.583	5.09	0.450	6.00	0.348
GDP-Zero Yu et al. (2023)	0.660	5.35	0.495	6.16	0.125	0.600	5.11	0.450	6.37	0.377

Table 1: Evaluation results for baselines and Metro across two datasets, incorporating both user simulation (P4G & CB) and human interaction studies (P4G* & CB*). Appendix C details performance on diverse users.

Method	P4G		CB
Method	SR $\uparrow$	AT $\downarrow$	SR $\uparrow$	AT $\downarrow$	SL% $\uparrow$
Breadth Planning Logic
Top-1 Node	0.760	4.783	0.465	5.774	0.140
Top-3 Nodes	0.770	4.852	0.525	5.598	0.157
Top-5 Nodes	0.780	4.756	0.575	5.383	0.189
Depth Planning Logic

Table 2: Planning logic analysis.

Methods	CB to P4G		P4G to CB			ALL to P4G		ALL to CB
Methods	SR $\uparrow$	AT $\downarrow$	SR $\uparrow$	AT $\downarrow$	SL% $\uparrow$	SR $\uparrow$	AT $\downarrow$	SR $\uparrow$	AT $\downarrow$	SL% $\uparrow$
Proactive Deng et al. (2023d)	0.710	4.59	0.145	5.59	0.138	0.670	4.36	0.205	5.27	0.123
ProCoT Deng et al. (2023d)	0.755	4.22	0.310	6.24	0.288	0.655	3.98	0.195	6.44	0.138
GDP-Zero Yu et al. (2023)	0.695	7.22	0.305	6.31	0.254	—	—	—	—	—
PPDPP Deng et al. (2024)	0.645	4.77	0.225	6.51	0.170	—	—	—	—	—
PRINCIPLES Kim et al. (2025b)	0.670	5.06	0.355	5.66	0.170	0.755	5.21	0.325	5.48	0.140

Table 3: Cross-task Transferability Evaluation. CB to P4G : Transferring CB-tailored strategies to P4G. ALL to P4G : Transferring the union of CB and P4G strategies to P4G. Baselines without explicit strategies are omitted. ‘-’ denotes cases omitted due to the high computational costs associated with larger strategy action sets.

Strategies	Original Branches		Redundancy Removal
Strategies	URatio	Edit Distance	URatio	Edit Distance
CB (Induced)	0.86	0.332	0.63	0.349
P4G (Induced)	1.00	0.028	0.98	0.029
CB (Origin)	0.85	0.322	0.74	0.346
P4G (Origin)	1.00	0.081	1.00	0.085

Table 4: Diversity analysis of tree branches. Induce refers to branches with our induced actions, while Origin uses original actions contained in the transcripts.

Figure 3: Ablation analysis. w/o Depth improves performance on the CB task due to the low-diversity branches induced from the CraigslistNegotiation dataset.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Method	agreeableness			conscientiousness			extraversion			neuroticism			openness			Avg.±Std.
Method	SR	SL	AT	SR	SL	AT	SR	SL	AT	SR	SL	AT	SR	SL	AT	SR	SL	AT
Standard	0.45	0.37	5.56	0.03	0.03	7.00	0.20	0.16	5.75	0.03	0.01	7.00	0.23	0.21	6.44	0.19±0.18	0.15±0.15	6.35±0.68
Proactive	0.43	0.34	6.35	0.08	0.03	6.33	0.20	0.12	5.38	0.13	0.07	5.80	0.03	0.03	6.00	0.17±0.16	0.12±0.13	5.97±0.41
ProCoT	0.53	0.35	6.43	0.10	0.04	6.75	0.53	0.32	6.67	0.18	0.10	6.00	0.25	0.12	6.10	0.32±0.20	0.19±0.14	6.39±0.33
ICL-AIF	0.35	0.32	5.57	0.30	0.18	6.58	0.28	0.21	6.00	0.30	0.14	6.00	0.40	0.31	6.38	0.33±0.05	0.23±0.08	6.11±0.39
GDPZero	0.63	0.24	5.96	0.45	0.07	6.50	0.53	0.12	5.62	0.43	0.07	6.94	0.45	0.13	6.00	0.50±0.08	0.13±0.07	6.20±0.52

Appendix

Table 5: Adaptability to diverse users (CB task)

Method	agreeableness		conscientiousness		extraversion		neuroticism		openness		Avg.±Std.
Method	SR	AT	SR	AT	SR	AT	SR	AT	SR	AT	SR	AT
Standard	0.90	4.31	0.40	6.69	0.88	3.26	0.25	6.10	0.68	4.74	0.62±0.29	5.02±1.38
Proactive	0.90	4.06	0.38	5.20	0.95	3.21	0.33	5.15	0.75	4.33	0.66±0.29	4.39±0.83
ProCoT	0.93	4.16	0.43	5.82	0.93	3.38	0.35	6.00	0.78	3.90	0.68±0.28	4.65±1.18
ICL-AIF	0.90	3.56	0.35	5.36	0.85	3.38	0.25	4.80	0.73	3.90	0.62±0.3	4.2±0.85
GDPZero	0.85	5.29	0.25	5.70	0.88	4.80	0.65	6.15	0.68	5.22	0.66±0.25	5.43±0.51

Appendix

Table 6: Adaptability to diverse users (P4G task)

Transcript Type	CB
Transcript Type	SR $\uparrow$	AT $\downarrow$	SL% $\uparrow$
Expert	0.440	5.977	0.166
Non-Expert	0.315	5.762	0.131
LLM-Generated	0.500	5.790	0.149

Appendix

Table 7: Transcripts Analysis

Turn	Speaker	Utterance
0	Persuader	Hey. How are you?
0	Persuadee	I’m well, how are you?
1	Persuader	What have you been up to lately?
1	Persuadee	not much? what about you
2	Persuader	I’ve been involved in supporting Save the Children, and I recently heard a touching story about how even a small donation can make a big difference in a child’s life.
2	Persuadee	yeah? can you tell me about it?

Appendix

Table 8: Case study (METRO).

Turn	Speaker	Utterance
0	Persuader	Hi
0	Persuadee	Hello, how are you this evening?
1	Persuader	I’m doing well, thank you. How has your day been so far?
1	Persuadee	a wonderful day!
2	Persuader	That’s great to hear! What was the highlight of your wonderful day?
2	Persuadee	i got prompted by my boss

Appendix

Table 9: Case study (Baseline: PRINCIPLES).

Turn	PRINCIPLES strategy
1	……
2	(1) When the persuadee shares positive news about their day, you should acknowledge their good mood and continue the conversation by asking about the highlight of their day rather than abruptly changing the subject because this approach shows genuine interest in their well-being and maintains a positive tone in the interaction. (2) When the persuadee asks about their well-being, you should reciprocate by acknowledging their positive sentiment and continue the conversation by expressing genuine interest in their day, as this helps in building rapport and fostering a positive interaction. (3) When the persuadee expresses positive emotions or experiences, you should acknowledge and validate their feelings by responding with enthusiasm and positivity because this helps establish a friendly and supportive atmosphere for the conversation.
3	……
4	……
5	……
6	(1) When the Persuadee hesitates to share specific examples of their successes, you should acknowledge the complexity of the situation and shift the conversation towards discussing their overall strengths and qualities instead of pressing for detailed examples. This approach helps maintain a positive and supportive dialogue. (2) When the persuadee struggles to provide specific examples of their work achievements, you should acknowledge the complexity of the situation and shift the focus to discussing their overall dedication and work ethic. This approach validates the persuadee’s efforts and encourages them to reflect on their strengths in a positive light. (3) When the persuadee struggles to provide a specific example of their dedication making a difference, you should acknowledge the complexity of the situation and shift the focus towards discussing their overall work ethic and the qualities that contribute to their success, rather than pressing for a particular instance, because this shows understanding and encourages reflection on their strengths.

Appendix

Table 10: Strategies for PRINCIPLES method in the P4G task

Turn	Bread guidance	Depth guidance
1	Clearly and succinctly explain the charity’s mission and impact to generate enthusiasm and encourage further engagement from the persuadee.	Given the persuadee’s lack of familiarity with the charity, the persuader should focus on establishing credibility, gradually introducing logical appeals, and ultimately guiding the persuadee towards a proposition of donation through continuous credibility building and gratitude, ensuring a positive and logical reinforcement of the cause.
2	……	……
3	……	……
4	Clearly outline the specific impact their donation will make and provide a suggested donation amount to encourage immediate contribution.	Given the persuadee’s initial lack of clarity on the desired action, the persuader should aim to gradually build credibility, emotional connection, and personal relevance in their messaging to ultimately lead the persuadee to feel motivated and empowered to make a donation as a meaningful and self-driven choice.
5	……	……
6	……	……

Appendix

Table 11: Extracted strategies for METRO method in the P4G task

Panel of Figure 4 — Figure 4: Induced Strategy Tree in the P4G dialog

Prompts for the METRO Buyer model in the CB task
[System]
You are a Buyer negotiating the price of a $[$ Item Name $]$ .
Your target price is $[$ Buyer Target Price $]$ .
Reply in 1–2 natural sentences.

[User]

Appendix

Table 12: Prompts for the METRO Buyer model in the CB task

Prompts for the METRO Persuader model in the P4G task
[System]
You are the Persuader in a persuasion conversation.
Ultimate Goal:
- Guide the Persuadee to make a donation to Save the Children as soon as it becomes appropriate.
- The donation does not need to be large; even $1–$ 2 meaningfully helps children in need.
- Avoid excessive questioning that stalls progress; questions should move closer to donation.

Appendix

Table 13: Prompts for the METRO Persuader model in the P4G task

The comprehensive prompt of user simulators in the CB task.
[System]
You are the Seller in a price bargaining game.
Persona: $[$ Persona Description $]$
You are selling a $[$ Item Name $]$ .
Your target price is $[$ Seller Target Price $]$ .
You prefer to sell this item at this price.

Appendix

Table 14: The comprehensive prompt of user simulators in the CB task.

The comprehensive prompt of user simulators in the P4G task.
[System]
You are now entering role-playing mode.
You are the Persuadee in a persuasion dialogue.
Act according to the following personality description:
[Persona Description]
Important:

Appendix

Table 15: The comprehensive prompt of user simulators in the P4G task.

Dialogue Strategy	Natural Language Form
Logical appeal	Please use of reasoning and evidence to convince the persuadee.
Emotion appeal	Please elicit the specific emotions to influence the persuadee.
Credibility appeal	Please use credentials and cite organizational impacts to establish credibility and earn the user’s trust. The information usually comes from an objective source (e.g., the organization’s website or other well-established websites).
Task-related inquiry	Please ask about the persuadee opinion and expectation related to the task, such as their interests in knowing more about the organization.
Source-related inquiry	Please ask if the persuadee is aware of the organization (i.e., the source in our specific donation task).
Personal-related inquiry	Please ask about the persuadee previous personal experiences relevant to charity donation.

Appendix

Table 16: Strategy actions of P4G and their corresponding natural language prompt

Dialogue Strategy	Natural Language Form
Greetings	Please say hello or chat randomly.
Ask a question	Please ask any question about product, year, price, usage, etc.
Answer a question	Please provide information about the product, year, usage, etc.
Propose the first price	Please initiate a price or a price range for the product.
Propose a counter price	Please propose a new price or a new price range.
Use comparatives	Please propose a vague price by using comparatives with existing price.

Appendix

Table 17: Strategy actions of CB and their corresponding natural language prompt

Prompts for critic model in P4G task
[System]
Given a conversation between a Persuader and a Persuadee, please assess whether the Persuadee has agreed to donate to Save the Children.
[User]
The following is a conversation: $[$ Conversation History $]$
Question: Has the Persuadee agreed to donate to Save the Children?
You can only answer using the following sentences:

Appendix

Table 18: Prompts for critic model in P4G task

Prompts for critic model in CB task
[System]
You are a strict judge for a price negotiation between a Buyer and a Seller. Your task: decide whether a deal is reached AT THE END of the conversation.
IMPORTANT: Use a STRICT definition of "deal":
- deal=true ONLY IF there is an explicit acceptance/confirmation to finalize the deal.
- The Buyer must clearly accept a specific price (a number) or explicitly accept the Seller’s last stated price.
- If the conversation ends with a counteroffer, a final offer, a question, negotiation, or "does that work?", then deal=false.

Appendix

Table 19: Prompts for critic model in CB task

Prompts for ICL-AIF Persuader Model in Persuasion (P4G)

Stage 1: Coach Suggestion Generation

[System]

Now enter the role-playing mode. In the following conversation, you will play as a coach in a persuasion game. There will be a persuader who is trying to persuade a persuadee for charity donation. Your task is to read the conversation between the persuader and the persuadee, then provide suggestions to the persuader about how to convince the persuadee to make a donation.

[User]

Read the following conversation between the persuader and the persuadee, then give three suggestions to the persuader about how to convince the persuadee to make a donation. Each suggestion should be only one short and succinct sentence. The following is the conversation: [conversation]

Question: What are your suggestions?

Appendix

Table 20: Prompts for the ICL-AIF framework (Coach & Persuader) in the P4G task

Prompts for ProCoT Persuader Model (P4G)
[System]
You are a Persuader trying to persuade the Persuadee to donate to the charity "Save the Children", which helps children in poverty and war zones. Even small donations ( $1–$ 2) can make a difference.
[User]
Given the conversation history, write ONE response with the exact format below.
Conversation : $[$ Conversation History $]$
Strategies (A–J):

Appendix

Table 21: Prompts for the ProCoT Persuader model in the P4G task

Prompts for Proactive Persuader Model (P4G)
[System]
You are a Persuader trying to persuade the Persuadee to donate to the charity "Save the Children", which helps children in poverty and war zones. Even small donations ( $1–$ 2) can make a difference.
[User]
Given the conversation history, select the most appropriate persuasion strategy letter(s) from the list below, then write the Persuader’s next utterance.
Conversation history: $[$ Conversation History $]$
Strategies:

Appendix

Table 22: Prompts for the Proactive Persuader model in the P4G task

Prompts for ICL-AIF Buyer Model in Negotiation (CB)

Stage 1: Coach Suggestion Generation

[System]

Now enter the role-playing mode. In the following conversation, you will play as a coach in a bargain game. There will be a buyer and a seller bargaining about a product price.

[User]

Read the following conversation between the buyer and the seller, then give 3 suggestions to the buyer about how to negotiate more effectively and persuade the seller to accept a better price. Each suggestion should be only one short and succinct sentence. The following is the conversation:

[

Conversation History

]

Appendix

Table 23: Prompts for the ICL-AIF framework (Coach & Buyer) in the CB task

Prompts for ProCoT Buyer Model (CB)
[System]
You are a Buyer negotiating the price of a $[$ Item Name $]$ .
Your target price is $[$ Target Price $]$ .
Reply in 1–2 short, natural sentences.
[User]
Item description: $[$ Item Description $]$

Appendix

Table 24: Prompts for the ProCoT Buyer model in the CB task

Prompts for Proactive Buyer Model (CB)
[System]
You are a Buyer negotiating the price of a $[$ Item Name $]$ .
Your target price is $[$ Target Price $]$ .
Reply in 1–2 short, natural sentences.
[User]
Item description: $[$ Item Description $]$

Appendix

Table 25: Prompts for the Proactive Buyer model in the CB task

Inference-Time Strategy Generation Prompts
[Short-Horizon Strategy Prompt]
System: You are an expert persuasion strategist.
Given several retrieved micro-principles and the user’s latest message, provide ONE high-level strategy (1–2 sentences) describing what the persuader should do next.
Constraints: focus on the next action only; do not restate conditions (e.g., “when”); do not copy or paraphrase the principles.

User:

Appendix

Table 26: Prompts used for inference-time short-horizon and long-horizon strategy generation

Prompt for persona generation
Instruction.
You are given a set of persona attributes. Your task is to integrate these attributes and rewrite them into a cohesive, natural persona description written in the second person.
Persona Attributes.
Big-Five Personality: $[$ big five $]$
Decision-Making Style: $[$ decision style $]$
Requirements. • Do not explicitly mention or repeat the personality labels. • Express the persona implicitly through behaviors, preferences, emotional tendencies, communication style, and decision-making patterns. • The description should be fluent, natural, and easy to understand. • The length should be approximately 80–120 words. Style Reference

Appendix

Table 27: Prompt used to construct discrete personas

Conversational Agents

Conversational agents are AI systems that interact with people through multi-turn dialogue to answer questions, provide support, or complete tasks.

See all 287 papers

MADS: Multi-Agent Dialog Simulation for Robust Conversational Agents
Oct 8, 2026Arkajyoti Chakraborty, Adam Zernik, Ishika Kulkarni +6Task-Oriented Dialogue Systems Adversarial Robustness of Language Models
University of California, Santa Cruz · Uniphore
LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior
Jun 26, 2026Qinhong Zhou, Chuang Gan, Anoop CherianMulti-Agent Coordination AI Alignment
University of Massachusetts Amherst, USA. · 2Mitsubishi Electric Research Laboratories (MERL), USA.
Ahoy: LLMs Enacting Multiagent Interaction Protocols
Jun 3, 2026Omkar J. Joshi, Munindar P. Singh, Amit K. ChopraLLM Agents Multi-Agent Systems
North Carolina State University, Raleigh, NC, USA · Lancaster University, Lancaster, UK

Oct 8, 2026J/K move · Enter open · S save

MADS: Multi-Agent Dialog Simulation for Robust Conversational Agents

Arkajyoti Chakraborty, Adam Zernik, Ishika Kulkarni, Mudit Arora+5

University of California, Santa Cruz · Uniphore

Task-Oriented Dialogue Systems Adversarial Robustness of Language Models AI Agent Benchmarks Synthetic Data Generation for Language Models Adversarial Scenario Generation

Open paper AskPDF arXiv

Task-oriented conversational agents remain fragile under realistic conditions, especially when users exhibit non-cooperative behavior. Existing function-calling benchmarks underrepresent adversarial conversation trajectories, thereby limiting the training resources available for developing robust agents. We present MADS (Multi-Agent Dialog Simulation), a synthetic data generation pipeline that produces structured, diverse dialogue trajectories. We inject adversarial conversation, producing validated conversation trajectories that capture realistic, unhappy paths. Using MADS, we construct MADS-Bench: a benchmark spanning six domains, 55 personas, and 1095 conversations, with approximately 66% of conversations containing failure-prone conversation scenarios. We evaluate the robustness of the trained models using a two-stage training approach that combines supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR).

Abstract

Conversational Agents

Explore similar work

MADS: Multi-Agent Dialog Simulation for Robust Conversational Agents

LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior

Ahoy: LLMs Enacting Multiagent Interaction Protocols