Drafting Effective Mediator Messages

• Drafting mediator messages is defined as the systematic construction of neutral, context-aligned interventions that resolve disputes in multi-agent settings.
• The topic covers foundational Bayesian models, LLM-based frameworks, and practical architectures that drive message generation in diverse scenarios.
• Empirical studies show that mediator messages improve communication clarity and dispute resolution metrics, with LLM techniques often outperforming traditional methods.

Drafting mediator messages refers to the systematic construction of messages by a neutral party (the “mediator”) intended to facilitate, steer, or summarize communication between disputing agents or in complex multi-agent settings. Mediator messages are employed in diverse domains, from Bayesian persuasion networks with communication constraints to LLM–based dispute resolution, medical multi-agent systems, pull request platforms, and online flame war de-escalation. Rigorous design, context alignment, and a deep understanding of both the underlying utility structures and interactional dynamics are critical for effective mediator message drafting.

1. Foundational Models and Theoretical Principles

The canonical treatment of mediator message design in economic theory is embodied by the extension of Bayesian persuasion to environments involving uninformed mediators. In the single-mediator model, agents’ utilities are functions $u_S(a, \theta)$, $u_M(a, \theta)$, and $u_R(a, \theta)$, with the sender observing the state $\theta$ and communicating to the receiver only via the mediator. The design of mediator messages is reduced to an equilibrium concept, the fully-revealing-mediator equilibrium, which relies on the concept of affine domination: the set of possible posteriors induced by the mediator must form upper supporting hyperplanes over the mediator’s utility, making garbling unprofitable. The sender's optimal value at prior $\mu$ is given by the constrained concavification of the sender's utility function over the domain of affine-dominating distributions: $V(\mu) = \cav_D[v_S](\mu) = \max_{(p_1, \ldots, p_N) \in D,\, \alpha \in \Delta} \sum_k \alpha_k v_S(p_k)$ subject to $\sum_k \alpha_k p_k = \mu$. Any optimal signaling scheme requires at most $|\Theta|$ distinct mediator messages, determined via Carathéodory’s theorem, and is constructed by probabilistic Bayesian splitting aligned to the maximizing convex decomposition, with each decomposed posterior mapped to a canonical message that is passed ungarbled through the mediator to the receiver (Arieli et al., 2022).

2. LLM-Based Dispute Mediation: Methodologies and Frameworks

Mediators powered by LLMs are increasingly deployed in online dispute resolution (ODR), where drafting intervention messages consists of (a) selecting an appropriate taxonomy-grounded intervention type, and (b) generating a message embodying that move. The process begins with the analysis of the dispute, classification of conversational state (emotional, confusion, deadlock, evidential), and selection of 1–3 intervention types from a validated taxonomy (e.g., the Canadian Department of Justice 13-move schema: information exchange, emotional validation, issue narrowing, solution proposal, etc.). Each chosen intervention type dictates a 1–2 sentence message component, produced either by human or LLM, with explicit anchoring to parties’ concerns and situational facts.

Quality control is based on rubric-scored criteria—understanding (contextualization), neutrality, empathy, and actionable resolution—or by self-review checklists. LLM-drafted mediator messages have empirically matched or outperformed human interventions on these criteria in blind evaluation, with 84% of LLM-generated interventions rated as equal or better than human analogs (Tan et al., 2024). An iterative workflow combining initial LLM generation, rubric review, and possible “refinement” re-prompt, robustly yields principled mediator interventions.

3. Practical Architecture and Prompt Design for Mediator Message Generation

Technical implementations such as LLMediator systematize mediator message drafting as a multi-component pipeline: user messages are monitored and optionally reformulated for tone; recent conversation history is retained in a context manager; and tailored prompts are constructed to instruct the LLM to generate mediator interventions.

Prompt templates must precisely specify the mediator’s neutrality objective, prohibit continuation of the parties’ own dialogue, and emphasize guidance toward amicable settlement. In the ODR context, three core interaction modes are supported:

• F1: Reformulate inflammatory party messages to neutral, solution-oriented tone.
• F2: Draft mediator interventions conditioned on recent dialogue and optional instruction.
• F3: Trigger autonomous mediator interventions per configuration.

Mediator message generation is formalized as a mapping $M: (C_t, I_m) \to m_m$, where $C_t$ is the ordered set of historical messages, $u_M(a, \theta)$0 is an instruction, and $u_M(a, \theta)$1 is the generated mediator message. This design preserves auditability and allows for human-in-the-loop override, ensuring quality control and minimization of risks related to factual error or bias (Westermann et al., 2023).

4. Mediator Messages in Complex Multi-Agent Collaboration

In multi-agent medical decision scenarios, the MedOrch framework operationalizes mediator-led consensus formation. An LLM-based mediator agent aggregates responses from multiple heterogeneous VLM expert agents, detects answer conflicts via a similarity-based conflict score, and initiates targeted Socratic questioning rounds to elicit clarification or revision.

Mediator messages in this context are heavily structured:

• Header with agent role, round, and context.
• Restatement of both the clinical question and the agent’s prior answer.
• Explicit summary of inter-expert conflict.
• A single, focused Socratic question triggering self-reflection on evidence.
• Rigid output format (e.g., JSON or bulletized: Revised_Answer, Key_Evidence, Confidence_Score), facilitating downstream automatic judge-agent assessment.

Decision-making logic, including conflict thresholds and round limits, is logged for clinical auditability (Chen et al., 8 Aug 2025). This design ensures high-fidelity integration of heterogeneous agent responses and systematic convergence on a consensus grounded in explicit evidentiary reasoning.

5. Mediator Message Aggregation and Summarization in Software Engineering

Mediator-style aggregation is also used in collaborative software engineering, where bots serve as mediators between developer teams and arrays of automated feedback tools (CI, linter, coverage, etc.). FunnelBot implements an aggregation pipeline where mediator messages:

• Begin with a one-line summary (TL;DR) listing bots and high-level state.
• Group all feedback by issue category (not bot source), with each group collapsed by default for efficiency and clarity.
• Label each group with explicit, action-oriented headings and issue counts.
• Present a single exemplar comment per group and link to full details.
• End with a concrete, prioritized call to action for the developer.

No advanced clustering or summarization is used; categorical grouping and controlled verbosity are key to reducing cognitive overload for newcomers, maximizing readability, and promoting onboarding efficacy (Ribeiro et al., 2022).

6. Empirical Principles and Best Practices for Mediation-Focused LLM Steering

LLM-generated mediator messages in contentious online environments (e.g., flame wars, public comment threads) are structured via an explicit “steering” pipeline: first, an LLM generates a judgmental summary of escalation and emotional vectors; then, another LLM—conditioned on this analysis and the conversation history—drafts a concise, empathetic mediation message. Prompting recipes are highly prescriptive, mandating that the message:

• Acknowledge both sides’ feelings.
• Summarize the core dispute in neutral terms.
• Avoid direct blame, directives, or moralizing language.
• Use inclusive, calm tone and open-ended questioning.
• Remain concise (20–40 words), free of excess punctuation or all-caps.

Evaluations employ principle-weighted scoring (mean rubric score across empathy, neutrality, etc.), user simulation (toxicity reduction, engagement dynamics), and human–LLM comparative analysis (linguistic complexity, interpersonal stance). Effective mediator messages, by these metrics, consistently outperform baseline or open-source alternatives in de-escalation and constructive engagement (Li et al., 2 Dec 2025).

7. Agency and Granularity in Mediator Message Suggestions

In AI-mediated communication, message-level drafting (where a mediator supplies a full-reply suggestion) produces higher user satisfaction and efficiency but reduces the final author’s sense of agency relative to sentence-level assistance. Sentence-level suggestions support greater content ownership but are slower and less efficient. The optimal unit of suggestion should thus be aligned to the communication context, balancing efficiency, sense of agency, and authenticity. Interfaces that embed mediator message drafting should tune granularity, number of alternatives, and transparency/disclosure practices accordingly (Fu et al., 2023).

---

These converging research strands underscore the centrality of context alignment, neutrality, principled taxonomy selection, and workflow integration in the construction of mediator messages, whether in formal economic games, LLM-driven ODR, collaborative multi-agent AI systems, or practical software engineering platforms. The empirical literature provides direct templates, evaluation rubrics, and architectural designs, forming a rigorous foundation for implementing automated or human-in-the-loop mediator message drafting across technical domains.