Evidence-grounded reasoning
Use multilingual retrieval and RAG to connect every interpretation to relevant policy clauses, guidance and administrative documents.
VinUniversity AREP Research Project
PolicyVerse
Multilingual Multi-Agent LLMs for Policy Reasoning
PolicyVerse develops an evidence-grounded framework in which specialised AI agents retrieve, interpret and reason across complex policy documents, exposing contradictions, dependencies and cross-domain effects across multiple languages.
About PolicyVerse
Turning fragmented policy documents into grounded, interpretable reasoning
Government policy is distributed across ministries, document types, languages and administrative contexts. PolicyVerse investigates how large language models can reason over that complexity while remaining grounded in source material and transparent about the relationships they infer.
Policy reasoning as coordinated dialogue
Instead of relying on one model response, PolicyVerse treats interpretation as an interaction between specialised agents, for example labour, taxation or immigration agents, that can question one another's interpretations, retrieve additional evidence, surface contradictions and map dependencies between policy domains.
Interpretable policy logic
Represent rules, exceptions and dependencies through Policy World Models rather than leaving reasoning buried inside free-form generations.
Cross-lingual transfer
Develop the framework in English and Portuguese, then test its robustness in Vietnamese and Welsh policy environments with different linguistic and institutional characteristics.
Research programme
Four connected research pillars
PolicyVerse combines multilingual NLP, retrieval, multi-agent coordination and structured policy modelling into one reusable scientific framework.
R1
Multilingual retrieval and aligned policy data
Build clause-level, semantically aligned policy resources and retrieval pipelines that ground reasoning in source evidence across languages and document types.
R2
Collaborative multi-agent policy reasoning
Coordinate specialised domain agents so that interpretations can be challenged, refined and reconciled rather than produced as isolated single-model answers.
R3
Policy World Models
Extract structured representations of rules, exceptions, conditions and dependencies from unstructured policy text to support interpretable reasoning and cross-document analysis.
R4
Multilingual evaluation and low-resource transfer
Evaluate factual grounding, consistency, interpretability, contradiction detection and complex question answering, then test whether the framework transfers reliably to Vietnamese and Welsh.
PolicyVerse architecture
Policy sources → Retrieval → Specialist agents → Policy World Models → Evaluation
The architecture is designed so that policy reasoning remains connected to evidence, while complex cross-domain relationships are made explicit and testable.
Policy SourcesDecrees, circulars, guidance, parliamentary texts and administrative manuals
Retrieve & GroundClause-level multilingual retrieval and evidence selection through RAG
Specialist AgentsDomain agents compare interpretations and interrogate one another's evidence
Policy World ModelsStructured rules, dependencies, exceptions and interpretable policy logic
Evaluate & ExplainQuestion answering, contradiction analysis, dependency mapping and auditability
Contradiction Detection
Identify policy statements that conflict, overlap or impose incompatible constraints across documents or domains.
Dependency Modelling
Map how one policy rule depends on another, including conditions, exceptions and cross-ministry relationships.
Complex Policy QA
Answer questions that require evidence from multiple documents, policy areas or languages rather than simple passage retrieval.
Transparent Reasoning
Expose the supporting evidence and structured policy logic behind an answer so interpretations can be inspected and evaluated.
Two-phase multilingual design
Develop with high-resource languages, then test real cross-lingual transfer
The project deliberately separates framework construction from low-resource adaptation, making it possible to test which components transfer and which need language-specific redesign.
Phase 1 · Core development
English and Portuguese
Establish the PolicyVerse research infrastructure in higher-resource settings before moving to more heterogeneous multilingual policy environments.
- Build the aligned PolicyVerse Dataset v1.0 with clause segmentation and semantic annotation.
- Develop the multilingual retrieval pipeline and initial cross-lingual RAG prototype.
- Design the Policy World Model extraction pipeline for interpretable policy rule structures.
- Develop and validate the first multi-agent coordination mechanisms for policy reasoning.
EN · EnglishPT · Portuguese
Phase 2 · Low-resource transfer
Vietnamese and Welsh
Adapt and stress-test the framework in linguistically and administratively different settings using authentic local policy materials.
- Integrate Vietnamese and Welsh decrees, circulars, guidance, parliamentary texts and administrative manuals.
- Produce Policy World Models across all four project languages and compare their structural consistency.
- Evaluate contradiction detection, dependency modelling and multi-document policy question answering.
- Measure cross-lingual transfer and identify where language or governance context changes model behaviour.
VI · VietnameseCY · Welsh
Project deliverables
A reusable research infrastructure for multilingual policy analysis
PolicyVerse is designed to produce datasets, models, evaluation resources and open tools that can support research beyond the lifetime of the project.
PolicyVerse Dataset v1.0
An aligned, clause-segmented and semantically annotated policy corpus supporting multilingual retrieval, reasoning and evaluation.
Cross-lingual RAG prototype
A multilingual retrieval and grounding pipeline that connects model reasoning to relevant policy evidence across languages.
Policy World Models v1.0
Structured representations of policy rules, exceptions, conditions and dependencies across English, Portuguese, Vietnamese and Welsh.
Multi-agent reasoning system
An integrated framework in which specialised agents coordinate, challenge interpretations and reason across policy domains and documents.
GlobalPolicyQA benchmark
A multilingual evaluation benchmark for factual grounding, consistency, interpretability and complex governance scenarios.
PolicyVerse open resources
Reusable code, evaluation components, documentation and research artefacts released through the project where licences permit.
Research team
NLP, AI and policy expertise across Vietnam and the UK
PolicyVerse brings together expertise in multilingual NLP, multi-agent systems, corpus and evaluation methods, Welsh and Vietnamese language technologies, and policy and regulatory analysis.
Mo El-Haj
Principal Investigator
Multilingual NLP, low-resource language technology, large language models and project leadership.
VinUniversity
Vietnam
Vietnam
Leandro S. Marcolino
Co-PI
Multi-agent AI, Portuguese policy data and evaluation.
VinUniversity
Vietnam
Vietnam
Paul Rayson
Co-Investigator
Corpus design, multilingual NLP resources and evaluation frameworks.
Lancaster University
United Kingdom
United Kingdom
Dawn Knight
Co-Investigator
Welsh language resources, corpus linguistics, discourse analysis and multilingual validation.
Cardiff University
United Kingdom
United Kingdom
Nguyen Thi Mai Lan
Project Collaborator
Vietnamese policy and regulatory expertise supporting domain grounding and evaluation.
VinUniversity
Vietnam
Vietnam
Nguyen Ha Thanh
Project Collaborator
Research collaboration supporting the Vietnamese policy and AI components of PolicyVerse.
VinUniversity
Vietnam
Vietnam
Nguyen Minh Chi
Research Assistant
Technical research and implementation across data, retrieval, evaluation and multi-agent experimentation.
VinUniversity
Vietnam
Vietnam
Project collaboration
VinUniversity CECSVinUniversity CBMNLP @ VinUniUCRELLancaster UniversityCardiff University
Expected impact
More transparent, accessible and evidence-grounded policy intelligence
By combining multilingual NLP, computational social science and public policy, PolicyVerse aims to advance foundational research in cross-lingual retrieval, LLM interpretability and multi-agent coordination while creating practical infrastructure for understanding complex policy environments.
ScientificNew methods for multilingual policy retrieval, structured policy modelling, multi-agent reasoning and interpretable evaluation.
Vietnam and low-resource policy AIA dedicated pathway from high-resource development to Vietnamese policy analysis, with direct attention to local documents and administrative contexts.
Public and organisational valueResearch infrastructure that can support clearer policy interpretation, accessibility and evidence-based decision-making for policymakers, organisations and citizens.
Contact
Interested in PolicyVerse?
For research collaboration, technical discussion or information about the project, contact the PolicyVerse Principal Investigator at VinUniversity.