What question did this study set out to answer?

The research aims to explore how an AI agent can transfer learning without traditional training methods.

February 24, 2026Open Access

Cross-Domain Transfer Learning in Developmental Knowledge Graphs Without Training

Key Points

The research aims to explore how an AI agent can transfer learning without traditional training methods.
Utilized a dual-graph neural architecture for separating emotional and analytical cognition.
Employed graph-structural embeddings for knowledge transfer.
Implemented multi-head attention mechanisms for activation dynamics.
Demonstrated metacognitive self-awareness without prior large language model training.
Achieved 29–70% accuracy in cross-domain transfer learning across four domain pairs.
Agent showed capability to express uncertainty by stating 'I don't know' when evidence is lacking.

Abstract

We present Vyan, a developmental AI agent that achieves 29–70% cross-domain transfer learning across four domain pairs without any gradient-based training, pre-trained embeddings, or large language model reasoning. The agent builds knowledge through a dual-graph neural architecture separating emotional from analytical cognition, and transfers knowledge via graph-structural embeddings and multi-head attention-driven activation dynamics. The agent also demonstrates metacognitive self-awareness, including the ability to say "I don't know" when evidence is insufficient.

Read Full Paperexternally

Perguntar à IA

Bookmark

View Full Paper