Overview
Artificial Social Childhood (ASC) creates a generative, socially rich environment for AI—an artificial childhood—to tackle a known limitation: today’s LLM/RL systems struggle with consistent causal reasoning, counterfactual testing, and extracting causal structure from static datasets. In ASC, agents learn in dynamic, social settings that mirror how children acquire knowledge: active, social, ecological.
What this grant builds
- A dynamic simulator (“social gym”) grounded in an object-oriented POMDP where an “AI child” explores, acts, and asks questions to reduce uncertainty—the “AI of Why.”
- Guidance & explanations from an “AI adult,” the “AI of Because,” which synthesizes knowledge and provides causal feedback in digestible formats.
- Causal learning modules (e.g., a causal graph), meta-cognitive control via ReAct-style planning, and adaptive, multi-source curiosity to drive informative exploration.
- Curriculum of social subtasks to evaluate gains in causal inference, question-asking, perspective-taking, and intention recognition.
Why it matters
Robust decision-making depends on causal knowledge. By training AI in a social, interactive childhood rather than on static corpora, ASC aims to produce agents that generalize better, ask smarter questions, and reason about what would happen if…—key abilities current systems lack.
Methods at a glance
Environment → Object-Oriented POMDP;
Learner → AI Child (active exploration & question-asking);
Teacher → AI Adult (knowledge synthesis & causal explanations);
Control → ReAct-based meta-cognition for planning/actions;
Motivation → Adaptive curiosity from multiple sources.
Programme & team
- Funder: Volkswagen Foundation (VolkswagenStiftung)
- Leads: Prof. Dr. Dr. Azzurra Ruggeri (Central European University & Technical University of Munich) and Prof. Dr. Dimitri Ognibene (University of Milano-Bicocca)