Lightweight · Multi-Modal · Topology-Aware

MicroWorld A Lightweight Multi-Modal Social Simulation Engine

MicroWorld turns documents, images, videos, and graph signals into a runnable simulation pipeline. It builds event context, prepares simulation agents, runs cross-platform interactions, and outputs an analysis report together with detailed simulation records.

🔭 View Example 📖 Read User Guide

Text + Image + Video + Graph
Incremental Memory
PPR-based Influence

TopoSim paper overview figure about structure-coupled agent behavior and influence

Paper figure: topology changes how agents update and whom they follow.

Text + video inputs One event package can combine documents and media clips

Researchers to public audience The simulation can cover experts, media, and ordinary users together

Hype to verification The public demo follows how a story changes after evidence accumulates

Readable evidence trails Each run keeps an analysis report and the full simulation details

Highlights

Key Features

MicroWorld combines multimodal grounding with lightweight execution and topology-aware simulation mechanisms in both clustering and influence estimation.

🧩 Multi-modal Inputs

Text, image, video, and graph signals enter one simulation workflow.

🪶 Lightweight Runtime

Incremental memory and compact runtime artifacts reduce unnecessary overhead.

🧮 Dual Cluster Modes

Topology-aware simulation supports both threshold and LLM-keyword cluster drivers.

📈 PPR-driven Influence

Directional influence is derived from topology instead of uniform pairwise assumptions.

Dual Cluster Modes Efficiency

Cluster-based token reduction

Topology-aware coordination removes redundant updates and cuts token usage sharply as the workload grows.

OASIS Saved 91.7%

376.7K used from 4.53M baseline

ECS-500 Saved 77.3%

743.8K used from 3.28M baseline

FDE-LLM Saved 52.23%

136.7K used from 286.5K baseline

ECS-50 Saved 43.0%

185.6K used from 320.2K baseline

PPR-driven Influence Accuracy

PPR-guided trajectory fidelity

PPR-based influence weighting keeps the simulation trajectory closer to the reference trend than the baseline run.

Baseline 82.5%

PPR-guided 93.1%

Gain +10.6 pts

PPR-guided influence 93.1%

Baseline influence 82.5%

Closer alignment across the full simulated trajectory

Paper evidence

Trajectory consistency in ECS-50

In the ECS-50 simulation, agents with similar opinion trajectories also begin from highly consistent states. Intra-group variance stays far below population-level variance, which exposes stable structural regularities in the update process.

Trajectory Analysis ECS-50

Consistency analysis of opinion update trajectories in ECS-50 simulation

Group-level trends, individual trajectories inside a representative group, and group proportions all indicate strong regularities in opinion updates.

Consistency Reading Implication

Key reading

Agents with similar opinion trajectories also start from similar initial states, so the grouping reflects real structure rather than random overlap.

Trajectory groups Opinion updates cluster into stable evolution patterns.

Variance signal Intra-group variance stays far below the full population.

Implication These regularities support topology-aware updates for better fidelity and efficiency.

Project overview

End-to-end workflow

The repository covers the full path from multimodal input to graph construction, topology-aware simulation, and report output.

🛤️ Core workflow

MicroWorld parses event materials into graph structure and agent profiles, runs topology-aware simulation, and produces both a report and detailed run records.

Input Graph Build Simulation Report

🎯 Best-fit scenarios

📣 Public event and opinion analysis
🏷️ Brand reputation experiments
🕸️ Influence propagation studies
📡 Multi-source social signal simulation

Signal snapshot

Current demo coverage

The current site uses the LK-99 case to show how one workflow links multimodal input, simulation, and report generation.

Complete example The LK-99 demo covers input, simulation, and report output

Multiple input types Long-form text and several videos enter the same workflow

Multi-platform discussion The same event can be simulated as different styles of online discussion

Report and details Each run keeps the final report and the detailed simulation records

LK-99 demo

LK-99 demo at a glance

This example combines article-style text, videos, and scientific discussion, making the full workflow easy to understand.

Visual Entry Point LK-99

Example overview

The demo starts from high-profile superconductivity claims and follows how discussion shifts as replication results and technical details appear.

Research figure strip with synthesis chart and sample photos — Attention-grabbing clips drive the first wave of discussion.

Annotated LK-99 explanation frame with magnet and pellet labels — Public explanations translate a technical claim into shareable language.

Conceptual levitation image used as a widely shared visual anchor — The topic naturally attracts imagination, hype, and speculative framing.

LK-99 video frame used in public discussion — As more research groups joined the discussion, skepticism gradually began to surface.

Narrative Arc

What the workflow captures

One event can move through attention, debate, replication, and correction. The workflow keeps both the narrative outcome and the process behind it.

💡 Claim 📣 Hype 🔬 Replication ✅ Correction

Case Fit

LK-99 as a good fit

A multimodal event with fast amplification, visible role differences, and a clear correction cycle.

System view

How it works

The project can be summarized in three connected layers.

📥 Ingestion

Parse PDFs, text files, images, and videos, then extract usable event context for graph construction.

⚡ Runtime

Build ontology, graph, entity prompts, social relation graph, and platform-ready simulation configuration.

📑 Outputs

Produce an analysis report, runtime records, memory traces, and topology snapshots.

The site focuses on the end-to-end path from input package to simulation artifacts.

Next steps

Explore more

Architecture deep-dive, setup guide, and the LK-99 walkthrough.