From Words to Worlds: Compositionality for Cognitive Architectures

TL;DR

This paper investigates the compositional abilities of large language models, revealing that scaling improves these skills but instruction tuning may reduce them, raising questions about aligning models with human cognition.

Contribution

It provides empirical analysis across multiple LLMs and tasks, highlighting the complex effects of scaling and instruction tuning on compositionality.

Findings

Scaling enhances compositional strategies in LLMs

Instruction tuning can decrease compositional abilities

Open issues in aligning LLMs with human cognition

Abstract

Large language models (LLMs) are very performant connectionist systems, but do they exhibit more compositionality? More importantly, is that part of why they perform so well? We present empirical analyses across four LLM families (12 models) and three task categories, including a novel task introduced below. Our findings reveal a nuanced relationship in learning of compositional strategies by LLMs -- while scaling enhances compositional abilities, instruction tuning often has a reverse effect. Such disparity brings forth some open issues regarding the development and improvement of large language models in alignment with human cognitive capacities.