A Neuro-Symbolic Framework for Enhancing Generalization and Skill Acquisition in Abstract Reasoning Tasks

Ramalingam Sathishkumar, Harris Narayanen (2025) A Neuro-Symbolic Framework for Enhancing Generalization and Skill Acquisition in Abstract Reasoning Tasks. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (2MB) | Preview
Preview	PDF (Configuration Manual) Download (578kB) | Preview

Abstract

The current state of the art AI systems score less than 5% on the Abstraction and Reasoning Corpus (ARC-AGI-2) benchmark Chollet (2019), whereas humans achieve near-perfect scores of 95%. Large Language Models excel at pattern matching on seen data but struggle with genuine generalization and skill acquisition from minimal examples. Developing a system that can infer abstract, compositional rules from just a few examples remains a fundamental challenge for artificial intelligence. This research investigates a neuro-symbolic framework designed to address this abstract reasoning gap through parameter-efficient fine-tuning of compact language models with direct grid tokenization.

The ARC-AGI-2 dataset is used to develop and evaluate the approach, consisting of unique visual puzzles requiring few-shot pattern recognition and rule generalization. The methodology explores three architectural versions with progressive enhancements: Version 1 employs an LLM-driven approach with natural language intermediation, Version 2 adapts the ARChitects baseline from the 2024 ARC-AGI-1 competition (53.5% accuracy) ARC Prize (2025) using Mistral-NeMo-Minitron with LoRA rank-32 adapters Hu et al. (2022), and Version 3 incorporates systematic optimizations including increased LoRA rank, Unsloth framework integration, Turbo DFS inference, and enhanced ensemble scoring targeting 58- 63% accuracy. The preliminary implementation achieved 0% accuracy even though training appeared successful with decreasing loss values. This reveals a critical insight: reduced training loss does not automatically translate to correct predictions. The failure stemmed from three main issues: grid tokenization bugs that produced malformed outputs (affecting approximately 15% of predictions), lack of systematic end-to-end testing during development, and integration problems between the training and inference components. This research provides a comprehensive methodology framework, systematic failure analysis, and honest assessment of implementation challenges, demonstrating that theoretical architectural sophistication requires rigorous incremental validation to achieve practical results.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Fajemisin, Ade UNSPECIFIED
Uncontrolled Keywords:	Artificial intelligence; Domain-Specific Language (DSL); neuro-symbolic AI; abstract reasoning; generalization; Abstraction and Reasoning Corpus (ARC); Artificial General Intelligence (AGI)
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	02 Sep 2026 10:37
Last Modified:	02 Sep 2026 10:37
URI:	https://norma.ncirl.ie/id/eprint/9766

Actions (login required)

View Item