Do LLMs Use Compositional Mechanisms to Solve Compositional Tasks?

Determine whether large language models produce compositional behavior by invoking compositional processing mechanisms—rather than non-compositional or heuristic mechanisms—when solving tasks that are compositionally structured.

Background

The paper examines how feedforward LLMs handle two-hop factual retrieval tasks that can be formalized as y = g(f(x)). Although such tasks are compositional by construction, prior work and the authors’ experiments reveal a persistent compositionality gap: success on each individual hop does not guarantee success on the composed mapping.

Using logit-lens analyses, the authors identify two distinct processing modes: a compositional mechanism that surfaces the intermediate variable f(x), and a direct mechanism that maps x to y without a detectable intermediate representation. They further report that the choice of mechanism correlates with linearity properties in the model’s embedding/unembedding spaces, suggesting representational geometry influences processing strategies.

Despite empirical insights and correlations, the overarching question remains unresolved at a general level: whether LLMs’ apparent compositional behaviors are underpinned by truly compositional mechanisms across tasks, models, and settings.