Research dataset · Preprint release

Policy-driven Agentic World Simulation

A temporally grounded dataset for tracing how policies, news, institutions, and markets interact.

Tiviatis Sim¹, Woon Jia Hui¹, Xinming Gao², Chen Gao², Fengbin Zhu¹, Zheng Huanhuan³, Chua Tat Seng¹, Kenji Kawaguchi¹

¹National University of Singapore²Tsinghua University³City University of Hong Kong

The Dataset download contains the sanitized public subset used by this website. For access to the full dataset, contact the corresponding author listed in the Information section.

PAWS dataset generation pipeline, from verified policy selection through news and market collection, action extraction, validation, and multi-agent simulation — Pipeline overview
PAWS connects policy interventions to dated evidence, structured stakeholder actions, and replay-ready world states.

01 · Introduction

From policy shock
to agent action.

PAWS is a policy-centred dataset for studying multi-agent behaviour in financial markets. It aligns verified policy episodes with policy-window news, normalized organizations, extracted actions, and daily market-return context.

Instead of treating news or prices as isolated signals, PAWS preserves the sequence: what changed, who responded, how they acted, and what the surrounding market looked like.

02 · Dataset overview

One auditable view of a policy episode.

Current database snapshot reported in the paper.

Verified policy episodes: 36
Policy query keys: 301
Policy-linked news rows: 12,727
Structured action rows: 65,291

CoverageFinancial & economic policy

Primary unitPolicy · actor · action · day

Dataset licenseGNU GPLv3

Release statusPreprint preview

03 · Visual overview

See the dataset from three angles.

Actions for policies P5 through P43 plotted by relative position in each policy's observed action window; color encodes linked origin-news count — Figure 04Policy action timeline

Relative action timing across verified policy episodes; color encodes each action's linked origin-news count.
Open full resolution ↗

04 · Interactive playground

Inspect the chain of evidence.

Explore a sanitized public subset derived from the PAWS SQLite tables. Article bodies and raw event-frame payloads are not published here.

… published News + Action rows

Loading public database subset…The full SQLite database is never sent to the browser.

Interactive evidence map

Policy-normalized timeline

Every point keeps its real date. Horizontal position shows where it falls within the selected policy’s observed evidence window or documented policy window.

Policy Normalize by

News rows Action rows Position: 0% beginning · 100% end. Dense rows are evenly sampled here; every published row is searchable below.

Select a news or action point to inspect its real date, source row, and linked records.

Organizations & involvement

Who appears across policy episodes?

Ranked from all action rows linked to the policies in the public subset. Select an organization to compare its policy involvement and most frequent actions.

Policy Show

Select an organization to see its policy involvement and recurring action families.

SQLite table browser

Search the published rows

Loading table manifest…

05 · Citation

Build on PAWS.

The formal paper citation will replace this preview entry when publication details are available.

BibTeX · preview

Download .bib

@misc{sim2026pawspolicydrivenagenticworld,
      title={PAWS: Policy-driven Agentic World Simulation}, 
      author={Tiviatis Sim and Jia Hui Woon and Xinming Gao and Chen Gao and Fengbin Zhu and Zheng Huanhuan and Chua Tat Seng and Kenji Kawaguchi},
      year={2026},
      eprint={2609.28547},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2609.28547}, 
}