Skills-in-Context Prompting: Unlocking Compositionality in Large Language Models

We investigate how to elicit compositional generalization capabilities in large language models (LLMs). Compositional generalization empowers LLMs to solve complex problems by combining foundational skills, a critical reasoning ability akin to human intelligence. However, even the most advanced LLMs currently struggle with this form of reasoning. We examine this problem within the framework of in-context learning and find that demonstrating both foundational skills and compositional examples grounded in these skills within the same prompt context is crucial. We refer to this prompt structure as skills-in-context (SKiC). With as few as two exemplars, this in-context learning structure enables LLMs to tackle more challenging problems requiring innovative skill combinations, achieving near-perfect systematic generalization across a broad range of tasks. Intriguingly, SKiC also unlocks the latent potential of LLMs, allowing them to more actively utilize pre-existing internal skills acquired during earlier pretraining stages to solve complex reasoning problems. The SKiC structure is robust across different skill constructions and exemplar choices and demonstrates strong transferability to new tasks.

LLMs still struggle with compositional generalization, i.e., the ability to use existing skills to solve more complex unseen problems

chain-of-thought (CoT) prompting (Wei et al., 2022b) significantly improves the reasoning performance of LLMs by demonstrating how to approach a complex problem through a sequence of basic steps. Follow-ups such as Least-to-Most prompting (Zhou et al., 2022a) and decomposed prompting (Khot et al., 2022) propose a two-stage strategy, which first decomposes the problem into sub-problems, and then solve and combine them sequentially.

We find that the key insight is to teach the LLM to explicitly ground each of its reasoning steps on the (more foundational) skills. To this end, it is crucial to demonstrate both the foundational skills and the compositional examples grounded in these skills within the same prompt context. We refer to this (one-stage) prompting structure as SKills-in-Context (SKiC). Specifically, the SKiC prompt is constructed from three main blocks (Figure 1). The first block contains a short (non-exhaustive) list of skills that LLMs may need to use in order to solve a more complex problem, which include the instructions of the skills.

Compared to recent prompting methods for handling compositional problems such as Least-to-Most (LtM) (Zhou et al., 2022a) and Decomp (Khot et al., 2022), our SKiC is superior in several dimension: (i) Our SKiC is more general to solve extended sets of problems. Previous decomposing-based approaches like LtM and Decomp usually solve complex problems in a two-stage fashion by first decomposing the problem into a linear sequence of subproblems and then solving them sequentially. However, many of the tasks that have complex computation graphs such as multiplication and dynamic programming problems (Dziri et al., 2023) cannot be decomposed in a simple manner, which makes these decomposition-based approaches less applicable. (ii) The decomposition operation can also be viewed as one basic skill in SKiC (see Figure 16 for an example in a question-answer task). (iii) SKiC solves the complex problems in a single stage, which could alleviate the error propagation compared to decomposition-based approaches that require multiple distinct stages.