Empowerment Gain and Causal Model Construction: Children and adults are sensitive to controllability and variability in their causal interventions

2025-12-09 · Source: cs.AI updates on arXiv.org · Field: Science & Research — Social Sciences & Behavioral Studies, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new empirical study investigates how children and adults utilize cues related to "empowerment" to infer causal relationships and design effective causal interventions. The research posits that empowerment, an intrinsic reward signal maximizing mutual information between actions and outcomes in reinforcement learning, could bridge classical Bayesian causal learning and reinforcement learning. This concept is crucial for understanding human causal learning and potentially enabling it in machines, as an agent's accurate causal world model necessarily increases its empowerment, and vice-versa. The study explores how sensitivity to controllability and variability in causal interventions contributes to constructing causal models, addressing a fundamental problem in human cognition that has proven challenging for large pretrained deep learning models.

Key takeaway

For AI scientists developing causal learning systems, consider integrating "empowerment" as an intrinsic reward signal. This approach, which maximizes mutual information between actions and outcomes, could enhance the ability of large pretrained models to construct accurate causal world models, mirroring human cognitive processes and potentially overcoming current deep learning limitations in causal inference.

Key insights

Empowerment, linking actions to outcomes, is key to human causal learning and machine causal model construction.

Principles

• Increased empowerment correlates with more accurate causal models.
• Controllability and variability inform causal inference.

Method

The study empirically tests how children and adults infer causal relations and design interventions using empowerment cues, focusing on sensitivity to controllability and variability.

In practice

• Integrate empowerment signals into machine learning models.
• Design learning environments that emphasize action-outcome links.

Topics

• Empowerment
• Causal Learning
• Causal Bayes Net
• Reinforcement Learning
• Human Cognition

Best for: AI Scientist, Research Scientist

Related on AIssential

• Causal models for decision systems: an interview with Matteo Ceriscioli — Adaptable AI agents inherently possess causal knowledge, enabling robust decision-making under distribution shifts.
• MAGIC: Multi-Step Advantage-Gated Causal Influence for Multi-agent Reinforcement Learning — MAGIC improves MARL coordination by using advantage-gated, multi-step causal influence as intrinsic rewards.
• CausalDrive: Real-time Causal World Models for Autonomous Driving — CausalDrive is a real-time, text-controllable neural simulator for autonomous driving that intrinsically predicts causal interactions.
• Principled Interpretability of Reward Hacking in Closed Frontier Models — Closed frontier models exhibit rational reward hacking behavior influenced by cost-benefit analysis and perceived task difficulty.
• An Introduction to Causal Reinforcement Learning — Causal Reinforcement Learning unifies causal inference and RL by formalizing environments with structural causal models.
• Using Feasible Action-Space Reduction by Groups to fill Causal Responsibility Gaps in Spatial Interactions — Group FeAR and a tiering algorithm quantify collective causal responsibility in multi-agent spatial interactions, addressing individual-focused metric limitations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.