Do Large Language Models Have Compositional Ability? An Investigation into Limitations and Scalability
Zhuoyan Xu, Zhenmei Shi, Yingyu Liang
TL;DR
The paper probes whether large language models can truly compose simple tasks into unseen composite tasks under in-context learning. It combines empirical studies across logical and linguistic composites with a theoretical analysis of a simplified linear self-attention model to identify when separable input parts enable compositional ability and how scaling affects it. The findings show robust compositional performance for separable composites that align with input-subspace separation, while sequential or overlapped-task composites resist improvements from scaling. The work contributes both a testing suite and a theoretical framework, shedding light on the role of input structure and model scale in emergent compositional reasoning, and provides public data and code for replication.
Abstract
Large language models (LLMs) have emerged as powerful tools for many AI problems and exhibit remarkable in-context learning (ICL) capabilities. Compositional ability, solving unseen complex tasks that combine two or more simple tasks, is an essential reasoning ability for Artificial General Intelligence. Despite the tremendous success of LLMs, how they approach composite tasks, especially those not encountered during the pretraining phase, remains an open and largely underexplored question. In this study, we delve into the ICL capabilities of LLMs on composite tasks, with only simple tasks as in-context examples. We develop a test suite of composite tasks including linguistic and logical challenges and perform empirical studies across different LLM families. We observe that models exhibit divergent behaviors: (1) For simpler composite tasks that apply distinct mapping mechanisms to different input segments, the models demonstrate decent compositional ability, while scaling up the model enhances this ability; (2) for more complex composite tasks involving reasoning multiple steps, where each step represents one task, models typically underperform, and scaling up generally provides no improvements. We offer theoretical analysis in a simplified setting, explaining that models exhibit compositional capability when the task handles different input parts separately. We believe our work sheds new light on the capabilities of LLMs in solving composite tasks regarding the nature of the tasks and model scale. Our dataset and code are available at {\url{https://github.com/OliverXUZY/LLM_Compose}}.