Our Team

This is an academic research project by Dr. Mika Letonsaari. He holds an M.Sc. in Engineering Physics from the Helsinki University of Technology and a Ph.D. in Educational Sciences (Computational Thinking) from the University of Helsinki, Finland. He currently serves as a lecturer at the Institute of Intelligent and Interactive Technologies at the University of Economics Ho Chi Minh City, Vietnam.

Project details

This project is inspired by the ARC-AGI challenge, which tasks AI systems with solving abstract visual reasoning problems using only a few example input-output grid pairs. While many ARC tasks involve relatively simple basic shapes and transformations, such as rotations, symmetries, and color groupings, the deeper challenge lies not in recognizing these primitives, but in how an agent can learn, abstract, and reuse them across tasks in a human-like way. The process by which reusable abstractions emerge (concepts like “hollow rectangle” or “checkerboard symmetry”) and how they facilitate transfer between tasks remains poorly understood.

This project investigates concept blending as a mechanism for this kind of abstraction. Implemented within the MeTTa language of the Hyperon framework, our approach treats symbolic visual structures as hypergraph-based representations. Concept blending is used to generate novel mid-level concepts by combining and generalizing elements from past reasoning episodes.

Many low-level constructs in ARC-style tasks, such as adjacency, connectivity, and simple geometric forms, are well understood and widely implemented in current solvers. We expect these to reappear naturally in any symbolic reasoning system, including ours. However, mid-level abstractions (which involve relational configurations, symmetry groupings, or multi-shape constraints) are rarely explicitly represented. Instead, they are often rediscovered through brute-force search in each new task. This project aims to replace such inefficiency with learned, interpretable abstractions produced by concept blending, offering a more principled path toward generalization.

The resulting system will be evaluated on its ability to (1) construct coherent, generalizable abstractions, (2) apply them to new visual tasks, and (3) do so in a transparent, inspectable form. This work aims to contribute toward cognitively plausible, adaptive reasoning mechanisms within symbolic AI systems. While this project does not aim to actually participate to the ARC-AGI challenge, it seeks to demonstrate the viability of concept blending within hypergraph-based frameworks and to encourage the broader research community to explore symbolic, MeTTa-based approaches where appropriate.

Open Source Licensing