PaperlayerView exact revision on arXiv

Source-linked AI summary

KITA AI: A Multi-Agent LLM System for Pluralistic Policy Deliberation

Arnau Mayoral-Macau, Jiaqi Lai, Manala Tyobeka, Vukosi Marivate, William Chandra Tjhi, Georgina Curto

arXiv:2609.34890v1cs.CY

TL;DR

Public-policy debate systems often resolve disagreement through alignment or convergence, despite the need to examine affected stakeholders’ perspectives and rationales. KITA AI uses stakeholder personas and structured debate to report positions through multiple value systems, while treating non-convergence as an outcome. It is presented as deliberation support, not a decision procedure or population consultation.

  • Problem

    Structured multi-agent debate has largely pursued alignment or majority-vote resolution, while public policymaking calls for examining conflicting stakeholder rationales and impacts.

    Source: Introduction · p. 1

  • Method

    KITA AI builds stakeholder personas grounded in demographic axes and value systems, runs structured debates, and generates reports through value-system lenses.

    Source: System Overview · p. 2

  • Results

    KITA AI combines demographic and ethical grounding, reportable non-convergence, and user-selectable indicators for evaluating deliberation outcomes.

    Source: Introduction · p. 1

  • Takeaways & Limitations

    KITA AI offers decision-makers multiple stakeholder viewpoints and indicators through an open, modular design that supports customization to regional value systems.

    Source: Conclusion · p. 2

    KITA AI outcomes retain underlying model biases, are not population consultations, and require domain experts to interpret their rationales.

    Source: Responsible Use and Limitations · p. 2

Abstract

from arXiv · show

Public policies addressing urgent social and environmental challenges need to explicitly consider the diverse, often conflicting perspectives of the affected stakeholders. Despite computational decision-support approaches increasingly offering recommendations across diverse human value systems, they still tend to deliver a single consensus-driven outcome. We present KITA AI, a modular system in which multiple large language model agents, each grounded in distinct demographic stakeholder personas and conceptual frameworks, deliberate on policy scenarios. The objective of KITA AI is not merely to inform about a preferred policy proposal, but also to automatically surface who is affected by the scenario and provide decision-makers with the rationales and quantitative indicators behind each position. KITA AI treats non-convergence as a first-class explainable output, enabling policymakers to better understand the trade-offs and human impacts of the policies being discussed.

Introduction

Prior multi-agent debate systems have generally aimed to align agents or converge on a recommendation, leaving disagreement underused as support for public decision-making. KITA AI combines demographic and ethical grounding, reportable non-convergence, and user-selectable evaluation indicators.

Introduction

  • Earlier structured debates generally resolve divergence by majority vote, while the Habermas Machine deliberately converges toward majoritarian human views.The paper argues that public policymaking also requires examining stakeholder rationales and quantitative impacts.Source: Introduction · p. 1
  • ADEPT stages ethicists debating a fixed medical-triage scenario, while Steging, Leeuwen, and Zbiegie´n (2026) use courtroom-inspired roles for legal topics.Demographically conditioned personas have also been used to predict legislative behavior, including roll-call and European Parliament voting (Li, Gong, and Jiang 2024; Kreutner, Lutz, and Strohmaier 2026).Source: Introduction · p. 1
  • KITA AI combines demographic and ethical grounding, reportable non-convergence, and user-selectable indicators for evaluating deliberation outcomes.Its pipeline builds a stakeholder roster, runs a structured debate, and reports potential outcomes under several value systems.Source: Introduction · p. 1

System Overview

KITA AI uses sequential modules that pass a shared transcript through the policy-deliberation pipeline. The modules build stakeholder personas, make debate disagreements explicit, and generate reports through multiple value-system lenses, with optional context and simulation modules.

System Overview

  • The mandatory modules run sequentially and incrementally build a shared transcript artifact that carries context through the pipeline.Modules are defined by their inputs and outputs, while specific implementations are installable plugins.Source: System Overview · pp. 1–2

Agent Builder

  • The Agent Builder maps a policy scenario to stakeholder personas, using user-specified or LLM-proposed demographic axes and value systems.The roster should include prominent conflicting voices because missing stakeholders cannot be recovered downstream.Source: System Overview · p. 2

Debate Engine

  • The Debate Engine builds an argument graph from debate moves, aiming to make disagreement explicit and resolve or characterize it.The implemented plugin uses Toulmin-form proposals and formal dialogue-game moves; unanswered objections are conceded, and consensus is derived from the argument graph.Source: System Overview · p. 2

Report Engine

  • The Report Engine evaluates each candidate’s proposals through its own value-system lens, with multiple engines able to report on the same transcript.Published engines include established philosophical traditions and indicators such as the UN Sustainable Development Goals.Source: System Overview · p. 2

Optional Modules

  • Optional modules add context before persona-building through web research or other data sources, while the Simulator evaluates policy impacts using domain-specific external social models.Reliable simulation plugins may need tailoring to the policy domain.Source: System Overview · p. 2

Modularity and Scalability

KITA AI is designed to be replicable and scalable through model-agnostic routing, module independence, and an installable plugin library that can grow through community contributions.

Modularity and Scalability

  • KITA AI routes agent calls through OpenRouter so newer or more capable models can be adopted without redesign.Modules share a transcript instead of calling adjacent modules directly, allowing independent development, testing, and replacement.Source: Modularity and Scalability · p. 2
  • Installable plugins let the library grow with community contributions, including search wrappers, region-specific report engines, and alternative debate engines.Each module includes a template with an agents.md file to guide plugin creation.Source: Modularity and Scalability · p. 2

Use Cases and User Interface

The interface supports default policy analysis, historical backtesting, and perturbation tests, alongside tools for running debates and experimenting with modules.

Use Cases and User Interface

  • The default pipeline analyzes policy proposals, while backtesting restricts searches to information predating a policy debate to reproduce its outcome.Perturbation extends backtesting by varying the initial scenario formulation and checking whether proposals adapt.Source: Use Cases and User Interface · p. 2
  • A full pass through the three pipelines takes several minutes, so precomputed runs are available through the user interface.The Module Lab allows users to test specific modules with manual input.Source: Use Cases and User Interface · p. 2
  • Through the interface, users can propose policies, inspect evaluations and discussion outcomes, run persona-based debates, or build plugins with a coding agent.Source: Use Cases and User Interface · p. 2

Responsible Use and Limitations

KITA AI supports deliberation rather than making decisions. Its outputs retain underlying model biases, are not population consultations, and require expert interpretation.

Responsible Use and Limitations

Conclusion

KITA AI brings diverse stakeholder perspectives and multiple indicators to social challenges, supporting decision-makers. Its open, modular design supports plugins, scalability, and customization to regional value systems.

Conclusion

  • KITA AI leverages diverse stakeholder perspectives and offers multifaceted viewpoints and indicators to support decision-makers.Source: Conclusion · p. 2
  • Its open, modular design allows plugins, scalability, and customization to region-specific value systems.Source: Conclusion · p. 2
Loading 2609.34890v1…