What problem does this paper attempt to address?

The problem that this paper attempts to solve is the limitations of current artificial intelligence systems in abstraction and reasoning abilities, especially the lack of broad generalization ability of these systems when facing new situations outside the training data. Specifically, although neural networks can solve various specific tasks, in abstract visual reasoning tasks (such as Abstraction & Reasoning Corpus, ARC), existing algorithms are still difficult to achieve broad generalization, that is, the ability to learn new concepts from a small number of examples and apply these concepts in different situations. ARC is a data set containing abstract visual reasoning tasks, aiming to test the broad generalization ability of algorithms. Although the international competition has a prize of up to $100,000, the best algorithm still cannot solve most ARC tasks and relies on complex hand - made rules without using machine - learning methods. For this reason, the authors explored two new directions to solve ARC tasks: 1. **Adaptation of the DreamCoder algorithm**: The authors adapted the DreamCoder algorithm, a recent advanced algorithm in the field of program induction, to solve ARC tasks. They designed a pure functional language named Perceptual Abstraction & Reasoning Language (PeARL), enabling DreamCoder to solve ARC tasks, and proposed a new recognition model, which significantly improved the performance of the previous best implementation. 2. **Application of large - language models**: The authors introduced a new framework for using large - language models (LLMs) to solve ARC tasks, transforming these visual tasks into the text domain. Through detailed evaluation of the performance of three model categories, it was found that LLMs can achieve performance competitive with human - made systems after appropriate data augmentation and domain transformation. In addition, the authors also constructed an integration of multiple ARC solvers, considering heterogeneous performance. This integration method performs better than the current state - of - the - art single system. Finally, the authors released the arckit Python library to promote future research in this field. In conclusion, this paper aims to improve the broad generalization ability of computers in abstraction and reasoning tasks by exploring new machine - learning systems, especially methods combined with neural networks, so as to be closer to the human intelligence level.

Neural networks for abstraction and reasoning: Towards broad generalization in machines

Computational principles of intelligence: learning and reasoning with neural networks

A Feature-based Generalizable Prediction Model for Both Perceptual and Abstract Reasoning

Measuring abstract reasoning in neural networks

Learning Representations for Reasoning: Generalizing Across Diverse Structures

Learning Algebraic Representation for Systematic Generalization in Abstract Reasoning

Abstract Reasoning with Distracting Features

Generalization in Neural Networks: A Broad Survey

Untrained neural networks can demonstrate memorization-independent abstract reasoning

Relational reasoning and generalization using non-symbolic neural networks

Representations and generalization in artificial and brain neural networks

Towards Neural Network-based Reasoning

A Mathematical Approach to Constraining Neural Abstraction and the Mechanisms Needed to Scale to Higher-Order Cognition

Abstraction Learning

A neural network model of perception and reasoning

Human-Like Geometric Abstraction in Large Pre-trained Neural Networks

Inference of Abstraction for a Unified Account of Reasoning and Learning

Learning to reason over visual objects

Aligning Generalisation Between Humans and Machines

Evolutionary Training and Abstraction Yields Algorithmic Generalization of Neural Computers

Tunneling Neural Perception and Logic Reasoning through Abductive Learning