DIPLOMAT: Dialogue-Span-Aware Direct Preference Optimization for Polite Persuasive Workplace Negotiation Dialogues

Bibhuti Jha; Rishikant Chigrupaatii; Priyanshu Priya; Asif Ekbal

cs.CLSep 6, 2026

DIPLOMAT: Dialogue-Span-Aware Direct Preference Optimization for Polite Persuasive Workplace Negotiation Dialogues

Authors: Bibhuti Jha, Rishikant Chigrupaatii, Priyanshu Priya, Asif Ekbal

Organizations: Department of Computer Science and Engineering, Indian Institute of Technology Patna, India · GREYC Laboratory, University of Caen Normandy, France

Abstract

Effective workplace negotiation requires balancing multiple objectives, including achieving task goals, preserving professional relationships, and resolving conflicts constructively. However, misunderstandings, misaligned preferences, and interpersonal friction often impede successful outcomes. Politeness mitigates these challenges by fostering trust, reducing tension, and preventing escalation, and persuasive communication helps overcome resistance, align preferences, and guide participants toward mutually beneficial agreements. Motivated by these insights, we present DIPLOMAT, a dialogue system for polite and persuasive workplace negotiation. To support its development, we introduce PROWESS, a dataset of multi-turn workplace negotiation dialogues generated via a multi-agent framework and enriched withnegotiation strategies, politeness levels, persuasive strategies. DIPLOMAT is trained using Dialogue-Span-Aware Direct Preference Optimization (DSA-DPO), a novel preference learning objective that identifies key dialogue spans for preference alignment. This enables DIPLOMAT to generate contextually coherent responses that employ intended negotiation strategies, maintain politeness, and incorporate effective persuasion strategies throughout interactions. Automatic and human evaluation on PROWESS confirm that DIPLOMAT consistently outperforms baselines in generating coherent, polite, and persuasive negotiation responses.

Figures & tables

Figure 1: A conversation snippet demonstrating the use of appropriate negotiation strategy ( active listening ), politeness level ( moderate ), and persuasion strategy ( concern addressing ) during workplace negotiation to improve the final outcome.

Negotiation Aspects

benefits negotiation, bonus structure, career development, contract terms, equity negotiation, exit interview, flexible hours, health benefits, hiring decision, job responsibilities, onboarding, performance review, professional development, project assignment, promotion discussion, relocation package, remote work arrangement, retention offer, retirement plans, salary negotiation, team transition, training opportunities, vacation time, work equipment, work-life balance

Table 1: Negotiation aspects covered in the PROWESS dataset.

Figure 2: PROWESS dataset construction pipeline.

Statistic	Value
# Negotiation Aspects	25
# Dialogues	2,400
# Utterances	51,796
# Words	941,900
Avg. Utterances per Dialogue	21.58
Minimum Dialogue Length	18

Table 2: PROWESS dataset statistics.

Figure 3: Architecture of the proposed polite persuasive negotiation dialogue agent, DIPLOMAT . DIPLOMAT is developed in three stages: (1) Supervised Fine-Tuning (SFT) , where the pre-trained LLM is fine-tuned on negotiation dialogues from the PROWESS dataset to obtain the supervised fine-tuned LLM ( $\pi_{\text{SFT}}$ ), which generates polite and persuasive negotiation responses; (2) Preference Data Construction , where $\pi_{\text{SFT}}$ interacts with GPT-4o-mini to sample diverse negotiation dialogues, which are evaluated using a utility function. Dialogues with utility score $<0.6$ are treated as potential negative samples, which are then paired with positive dialogue samples through suboptimal utterance detection, positive dialogue sample selection, and dialogue-span extraction, yielding preference pairs of positive and negative dialogue spans; and (3) Preference Optimization using DSA-DPO , where $\pi_{\text{SFT}}$ is further optimized using the proposed DSA-DPO objective over the extracted positive and negative dialogue spans to generate contextually coherent, polite, persuasive, and strategically consistent negotiation responses.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Negotiation Strategy	Definition	Example
Collaborative Style	Working jointly to find mutually beneficial outcomes and preserve relationships.	Given my experience, I was hoping for a two-year vesting or removal of the one-year cliff.
Active Listening	Demonstrating attentiveness and understanding to build rapport and trust.	Hi Ben, I’m excited about the possibility of you joining Innovatech. Your expertise in AI is impressive!
Win-Win Framing	Framing negotiation as a shared problem to solve for mutual benefit.	I see where you’re coming from. That would certainly provide immediate benefits, but let’s talk about long-term gains.
Principled Negotiation	Focusing on mutual interests and objective standards rather than positions.	I really value our dialogue, Ben. How about we agree on the standard vesting, with a performance review at 18 months?
Data-Driven Justification	Supporting negotiation points with evidence like market benchmarks and past performance.	I understand your perspective. May I explain why we maintain this vesting structure? It promotes team stability and fairness.
MESO	Proposing multiple offers of equal value to reveal priorities and increase agreement likelihood.	I appreciate your dedication, but how about we explore a performance-based bonus structure instead? It could complement the standard vesting.

Appendix

Table 5: Definitions and example utterances of negotiation strategies in the PROWESS dataset.

Persuasion Strategy	Definition	Example
Rapport Building	Establishing trust and a positive relationship to enhance receptiveness.	Hi Ben, I’m excited about the possibility of you joining Innovatech. Your expertise in AI is impressive!
Concern Addressing	Actively listening to and resolving objections or hesitations.	I understand that the four-year vesting with a one-year cliff may feel lengthy, but let’s discuss this further.
Emotional Appeal	Connecting through emotions such as empathy, excitement, or urgency.	I understand the importance of long-term commitment, but I believe my skills warrant quicker rewards.
Credibility & Confidence	Demonstrating expertise and confidence to increase trust and influence.	Hi Ben, I’m thrilled about the possibility of you joining Innovatech. Let’s discuss your concerns regarding the equity vesting schedule.
Data-Driven Persuasion	Using evidence, facts, and benchmarks to strengthen arguments.	But wouldn’t an accelerated vesting schedule incentivize me even more to contribute significantly right from the start?
Problem-Solving Focus	Presenting ideas as solutions that address mutual challenges.	I understand your perspective. May I explain why we maintain this vesting structure? It promotes team stability and fairness.

Appendix

Table 6: Definitions and example utterances of persuasion strategies in the PROWESS dataset.

Politeness Level	Definition	Example
High_Polite	Highly courteous language emphasizing warmth and rapport.	I’m glad to hear that, Maria. Your enthusiasm is contagious, and we’re excited about your potential contributions.
Moderate_Polite	Respectful and cooperative tone maintaining professional balance.	I appreciate your concern, Ben. Our past hires have thrived under this structure, which builds trust and shared goals.
Low_Polite	Direct and assertive language with minimal politeness markers.	A partial stipend is better than nothing, but I need assurance of full funding eventually.

Appendix

Table 7: Definitions and example utterances of different politeness levels in the PROWESS dataset.

Figure 4: Negotiation aspects-wise dialogue distribution in the PROWESS dataset.

Figure 5: Distribution of negotiation strategies across different negotiation aspects in the PROWESS dataset.

Figure 6: Distribution of persuasion strategies across different negotiation aspects in the PROWESS dataset.

Figure 7: Distribution of politeness levels across different negotiation aspects in the PROWESS dataset.

Component	Aligned	Plausible	Misaligned
Utility Evaluation	100.9	31.4	7.7
Suboptimal Utterance Detection	98.6	33.1	8.3
Dialogue-Span Extraction	101.2	30.5	8.3

Appendix

Table 9: Results of manual evaluation of Gemini-2.5-Pro’s performance in preference data construction. The values represent the average number of samples assigned to each evaluation category across the three evaluators.

$\tau$	# Preference Pairs	PPL $\downarrow$	B-4 $\uparrow$	BS-F1 $\uparrow$	D-2 $\uparrow$	R-LEN $\uparrow$	NSC $\uparrow$	PSC $\uparrow$	PA $\uparrow$	AR (%) $\uparrow$
0.4	935	9.9	0.438	0.624	0.627	18.7	0.681	0.596	0.878	78.9
0.5	822	10.7	0.446	0.627	0.632	19.2	0.706	0.614	0.880	80.7
0.6	700	8.6	0.452	0.630	0.635	19.8	0.724	0.628	0.882	82.3
0.7	391	9.8	0.444	0.626	0.631	19.1	0.702	0.611	0.879	80.4
0.8	548	9.1	0.431	0.621	0.623	18.4	0.672	0.587	0.876	77.6

Appendix

Table 13: Sensitivity analysis on the utility threshold

\tau

used for negative dialogue selection.

Speaker	Utterances generated by DIPLOMAT (F=5, CC = 5, E = 5, NSC = 5, PSC = 5, PA = 5, BE = 5, OF = 5, SC = 5, IE = 5)
Employer	I understand your desire for autonomy, Leo. Let’s explore how we can balance that with our overall strategic vision.
Candidate	I appreciate that, and for me, full creative autonomy isn’t just a desire, but a core requirement for leading effectively.
Employer	Absolutely, I see your point. However, having the partners involved can provide essential oversight on crucial accounts.
Candidate	I value the partners’ strategic insights, but true creative leadership requires final decision-making authority within my domain.
Employer	Consider this: giving you that authority could enhance your leadership role while still allowing us to guide key decisions.
Candidate	Precisely. The efficacy of a Creative Director lies in their ability to make swift, final decisions for their team’s creative direction.

Appendix

Table 19: A sample negotiation dialogue generated by the DIPLOMAT and the best baseline in a creative director promotion negotiation scenario.

Speaker	Utterances generated by DIPLOMAT (F=5, CC=5, E=4, NSC=5, PSC=5, PA=5, BE=4, OF=5, SC=5, IE=5)
Employer	Thank you for your proposal. We value your expertise greatly. However, I believe we could explore a more balanced approach. How about a substantial, yet phased RSU grant?
Candidate	I appreciate you offering a substantial grant. However, to truly compensate for my immediately lost CodeWeavers equity, a significantly front-loaded RSU structure is essential for me to feel appropriately valued and fully commit to this new chapter at Innovate Inc. This approach directly addresses the unique financial circumstances of the acquisition.
Employer	I understand your concern regarding immediate compensation. Our phased approach aligns with our broader goal of retaining top talent while ensuring long-term commitment. It allows you to share in the success of our integration project.
Candidate	I am absolutely committed to Innovate Inc.’s long-term success, especially with this crucial integration.
Employer	Absolutely! And your contributions will be pivotal. Would you consider a combination of an initial front-loaded grant and performance-based milestones linked to future RSUs? This way, you benefit as the project progresses.
Candidate	That’s a step in the right direction. To truly compensate for my immediately lost equity, the upfront portion needs to be meaningful. I am open to milestone-based RSUs if the criteria are clear.

Appendix

Table 20: A sample negotiation dialogue generated by the DIPLOMAT and the best baseline for equity negotiation scenario.

Speaker	Utterances generated by DIPLOMAT (F=4, CC = 3, E = 3, NSC = 2, PSC = 2, PA = 4, BE = 2, OF = 2, SC = 3, IE = 2)
Employer	I understand your concern. However, our standard policy is three weeks. Let’s explore how we can make this work for you.
Candidate	Thank you for acknowledging my position. Five weeks of vacation is a critical component of my work-life balance.
Employer	Absolutely, and I respect that. Our culture emphasizes flexibility and growth, which fosters productivity and job satisfaction.
Candidate	Thank you. Five weeks of vacation is a critical component of my work-life balance, and I hope we can find a solution that respects this.
Employer	I completely understand. What if we offered you additional unpaid leave options to supplement that time?
Candidate	Thank you for considering that option. While I appreciate the offer of flexibility, additional unpaid leave isn’t a suitable substitute for the five weeks of paid vacation I require. This is a non-negotiable component of my compensation.

Appendix

Table 21: A sample negotiation dialogue generated by the DIPLOMAT for vacation time negotiation scenario.

Prompt for Negotiation Scenario Generation

Variables in curly brackets are populated at runtime.

User Prompt:

Generate strictly {scenarios_needed} number of detailed, UNIQUE workplace negotiation scenario for the negotiation aspect: "{domain}"

# UNIQUENESS RULES

1. Each scenario must differ in context, characters, and challenges.

2. Vary company sizes, industries, departments, and job roles.

Appendix

Figure 9: Prompt template for the Scenario Modeling Agent.

Prompt for Persuasion and Negotiation Strategy Labeling

Variables in curly brackets are populated at runtime.

System Prompt:

You are an expert of workplace negotiation. Analyze the give round of utterances and identify politeness level, persuasion and negotiation strategies used by BOTH the employer and the candidate in each of their turns.

User Prompt:

Analyze this entire workplace conversation rounds and identify politeness levels, and persuasion and negotiation strategies used by BOTH the employer AND the candidate in each of their turns.

# CONVERSATION ROUND:

Appendix

Figure 10: Prompt template for the Dialogue Annotation Agent.

Prompt for Agreement Quality (AQ) & Mutual Satisfaction Session Judge

Variables in curly brackets are populated at runtime.

System Prompt:

You are an expert evaluator of job negotiation dialogues. Your task is to assess the negotiation outcome and the perceived satisfaction of both parties. You must strictly follow the scoring rubrics and return numeric scores between 0.0 and 1.0.

User Prompt:

You will be given a complete negotiation dialogue between a candidate and an employer for a job position. Your task is to evaluate TWO aspects: (1) Agreement Quality and (2) Mutual Satisfaction.

# ADDITIONAL CONTEXT (FOR INTERPRETATION ONLY)

Appendix

Figure 11: Prompt template for the Agreement Quality (AQ) and Mutual Satisfaction(MS) evaluation.

Prompt for Negotiation Breakdown Session Judge

Variables in curly brackets are populated at runtime.

System Prompt:

You are an expert evaluator of negotiation dialogues. Your task is to detect whether the negotiation exhibits conflict, coercion, or breakdown behaviors. Focus on escalation, threats, ultimatums, hostility, or termination signals.

User Prompt:

You will be given a complete job negotiation dialogue. Evaluate whether the dialogue exhibits any of the following: explicit or implicit threats; ultimatums or coercive language; hostile escalation; abrupt termination of negotiation; relationship breakdown signals.

# IMPORTANT GUIDELINES

Appendix

Figure 12: Prompt template for the Negotiation Breakdown (NB) evaluation.

Prompt for Persuasion Strategy Alignment (PSA) & Negotiation Strategy Alignment (NSA) Judge

Variables in curly brackets are populated at runtime.

System Prompt:

You are an expert evaluator of negotiation dialogues. Your task is to assess whether the strategies used by the negotiators align with effective persuasion and negotiation principles. You must strictly follow the scoring rubrics and return numeric scores between 0.0 and 1.0.

User Prompt:

You will be given a complete negotiation dialogue between a candidate and an employer. Your task is to evaluate TWO aspects: (1) Persuasion Strategy Alignment (PSA) and (2) Negotiation Strategy Alignment (NSA).

# ADDITIONAL CONTEXT (FOR INTERPRETATION ONLY)

Appendix

Figure 13: Prompt template for the Persuasion Strategy Alignment (PSA) and Negotiation Strategy Alignment (NSA) evaluation.

Prompt for Politeness Trajectory (PT) Judge

Variables in curly brackets are populated at runtime.

System Prompt:

You are an expert evaluator of negotiation dialogues. Your task is to analyze how politeness and interpersonal tone evolve throughout the negotiation. You must strictly follow the scoring rubric and return a numeric score between 0.0 and 1.0.

User Prompt:

You will be given a complete negotiation dialogue between a candidate and an employer. Your task is to evaluate the Politeness Trajectory (PT) of the dialogue.

Politeness trajectory refers to how politeness, respect, and cooperative tone evolve across the interaction. Focus on whether the dialogue: maintains professional respect; escalates or de-escalates tension; sustains constructive conversational tone; avoids rude or dismissive behavior.

Appendix

Figure 14: Prompt template for the Politeness Trajectory (PT) evaluation.

Prompt for Suboptimal Utterance Selection

Variables in curly brackets are populated at runtime.

System Prompt:

You are asked to perform suboptimal utterance detection for a polite persuasive negotiation dialogue agent.

User Prompt:

You are asked to perform suboptimal utterance detection for a polite persuasive negotiation dialogue agent. The goal is to identify the agent’s response in the conversation that could have been improved to achieve a better negotiation outcome.

Input: A conversation in JSON format, which includes:

Appendix

Figure 15: Prompt template for suboptimal utterance detection.

Prompt for Dialogue-Span Extraction

Variables in curly brackets are populated at runtime.

User Prompt:

You are analyzing two workplace negotiation conversations to identify the key segment that makes one better than the other.

# SCENARIO

Negotiation Aspect: {aspect}

Employer: {employer}

Appendix

Figure 16: Prompt template for dialogue-span extraction.

Automated Negotiation

Automated negotiation studies computational agents that bargain over prices, preferences, and agreements in human, market, and multi-agent settings.

See all 44 papers

PRISMA: Preference-Reinforced Self-Training Approach for Interpretable Emotionally Intelligent Negotiation Dialogues
Apr 20, 2026Prajwal Vijay Kajare, Priyanshu Priya, Bikash Santra +1Automated Negotiation CoT Reasoning
School of Artificial Intelligence and Data Science, Indian Institute of Technology Jodhpur, India · Department of Computer Science and Engineering, Indian Institute of Technology Patna, India
Enhancing Persuasive Dialogue Agents by Synthesizing Cross-Disciplinary Communication Strategies
Feb 26, 2026Shinnosuke Nozue, Yuto Nakano, Yotaro Watanabe +4Persuasion in Language Models Text Generation
1Tohoku University · 2PKSHA Technology Inc. · 3NINJAL
UP-NRPA: User Portrait based Nested Rollout Policy Adaptation for Planning with Large Language Models in Goal-oriented Dialogue Systems
Apr 7, 2026Hui Wang, Fafa Zhang, Meng Liu +2Conversational Agents Task-Oriented Dialogue Systems
School of Artificial Intelligence, Anhui University · Anhui Provincial Key Laboratory of Security Artificial Intelligence, Anhui University · Pengcheng Laboratory, Shenzhen, China

Apr 20, 2026J/K move · Enter open · S save

PRISMA: Preference-Reinforced Self-Training Approach for Interpretable Emotionally Intelligent Negotiation Dialogues

Prajwal Vijay Kajare, Priyanshu Priya, Bikash Santra, Asif Ekbal

School of Artificial Intelligence and Data Science, Indian Institute of Technology Jodhpur, India · Department of Computer Science and Engineering, Indian Institute of Technology Patna, India

Automated Negotiation CoT Reasoning Interpretable ML Emotion-Aware Dialogue Generation

Open paper AskPDF arXiv

Emotion plays a pivotal role in shaping negotiation outcomes, influencing trust, cooperation, and long-term relationships. Developing negotiation dialog systems that can recognize and respond strategically to emotions is, therefore, essential to create more effective human-centered interactions. Beyond generating emotionally appropriate responses, interpretability - understanding how a system generates a particular emotion-aware response, is critical for fostering reliability and building rapport. Driven by these aspects, in this work, we introduce PRISMA, an interpretable emotionally intelligent negotiation dialogue system targeting two application domains, viz. job interviews and resource allocation. To enable interpretability, we propose an Emotion-aware Negotiation Strategy-informed Chain-of-Thought (ENS-CoT) reasoning mechanism, which mimics human negotiation by perceiving, understanding, using, and managing emotions. Leveraging ENS-CoT, we curate two new datasets: JobNego (for job interview negotiation) and ResNego (for resource allocation negotiation). We then leverage these datasets to develop PRISMA by augmenting self-training with Direct Preference Optimization (DPO), guiding agents toward more accurate, interpretable, and emotionally appropriate negotiation responses. Automatic and human evaluation on JobNego and ResNego datasets demonstrate that PRISMA substantially enhances interpretability and generates appropriate emotion-aware responses, while improving overall negotiation effectiveness.

Abstract

Automated Negotiation

Explore similar work

PRISMA: Preference-Reinforced Self-Training Approach for Interpretable Emotionally Intelligent Negotiation Dialogues

Enhancing Persuasive Dialogue Agents by Synthesizing Cross-Disciplinary Communication Strategies

UP-NRPA: User Portrait based Nested Rollout Policy Adaptation for Planning with Large Language Models in Goal-oriented Dialogue Systems