Towards Compositional Generalization of LLMs via Skill Taxonomy Guided Data Synthesis

TL;DR

This paper introduces STEPS, a data synthesis framework guided by skill taxonomies to enhance compositional generalization in large language models and agents, addressing data scarcity in complex skill combinations.

Contribution

The paper proposes a novel taxonomy-guided data synthesis method that explicitly targets compositional generalization in LLMs and agents, leveraging structural information theory.

Findings

STEPS outperforms existing data synthesis methods on instruction-following benchmarks.

Improves compositional generalization in downstream agent-based tasks.

Effectively uncovers latent skill relationships for better data generation.

Abstract

Large Language Models (LLMs) and agent-based systems often struggle with compositional generalization due to a data bottleneck in which complex skill combinations follow a long-tailed, power-law distribution, limiting both instruction-following performance and generalization in agent-centric tasks. To address this challenge, we propose STEPS, a Skill Taxonomy guided Entropy-based Post-training data Synthesis framework for generating compositionally challenging data. STEPS explicitly targets compositional generalization by uncovering latent relationships among skills and organizing them into an interpretable, hierarchical skill taxonomy using structural information theory. Building on this taxonomy, we formulate data synthesis as a constrained information maximization problem, selecting skill combinations that maximize marginal structural information within the hierarchy while preserving semantic coherence. Experiments on challenging instruction-following benchmarks show that STEPS outperforms existing data synthesis baselines, while also yielding improved compositional generalization in downstream agent-based evaluations.