Introduction

Welcome to our research showcase on Meta-Causal Graphs, a novel approach to understanding how causal relationships change across different environmental contexts. Traditional causal modeling often assumes fixed relationships between variables, but in real-world scenarios, these relationships can shift dramatically based on underlying latent states.

Motivation

The Challenge of Context-Dependent Causality

Imagine a simple scenario: you push a door, expecting it to open. Most of the time, this works perfectly—push leads to open. But what happens when the door is locked? Suddenly, the same action produces no effect. This everyday example highlights a fundamental challenge in causal reasoning that traditional models struggle to capture.

Moving Beyond Static Causal Models

As illustrated in Figure 1, traditional approaches to causality face two critical limitations when modeling real-world scenarios:

Problem 1: Conventional causal graphs treat relationships as fixed and uniform, neglecting how context affects causality. When analyzing data from both locked and unlocked doors, these models either incorrectly generalize the "push → open" relationship to all situations or fail to identify any consistent pattern.

Problem 2: Even domain-specific models that acknowledge different contexts often require predefined labels (like "locked" vs. "unlocked"), limiting their ability to generalize to novel scenarios or discover hidden contextual factors autonomously.

Our Research Focus

Our work introduces the Meta-Causal Graph framework and Curiosity-Seeking Agents that can actively explore environments to discover how causal structures evolve across different meta states. By combining targeted interventions with a curiosity-driven exploration policy, our approach adaptively reveals global causal rules governing complex environments.

Key Innovations

• Meta-Causal Graph: A unified representation that efficiently encodes how causal structures shift across different latent world states

• Curiosity-Driven Exploration: A novel reward function that guides agents to explore states where causal structure remains uncertain

• Interventional Verification: Direct testing of causal relationships under different state conditions to establish robust causal links

• Adaptive Learning: Continuous refinement of causal knowledge through ongoing exploration and experience

Experiment

We rigorously evaluated our Meta-Causal Graph framework and Curiosity-Seeking Agents across multiple experimental settings designed to test different aspects of causal discovery and generalization.