CARMA: Enhanced Compositionality in LLMs via Advanced Regularisation and Mutual Information Alignment

TL;DR

CARMA is a scalable regularisation method that enhances compositional reasoning in large language models by improving stability and robustness of internal representations without sacrificing task performance.

Contribution

We introduce CARMA, a novel regularisation technique using mutual information and stability constraints to improve compositional generalisation in LLMs.

Findings

CARMA reduces variability in fine-tuned models.

It stabilises token representations across layers.

It improves compositional reasoning and robustness.

Abstract

Large language models (LLMs) struggle with compositional generalisation, limiting their ability to systematically combine learned components to interpret novel inputs. While architectural modifications, fine-tuning, and data augmentation improve compositionality, they often have limited adaptability, face scalability constraints, or yield diminishing returns on real data. To address this, we propose CARMA, an intervention that enhances the stability and robustness of compositional reasoning in LLMs while preserving fine-tuned performance. CARMA employs mutual information regularisation and layer-wise stability constraints to mitigate feature fragmentation, ensuring structured representations persist across and within layers. We evaluate CARMA on inverse dictionary modelling and sentiment classification, measuring its impact on semantic consistency, performance stability, and robustness to lexical perturbations. Results show that CARMA reduces the variability introduced by fine-tuning, stabilises token representations, and improves compositional reasoning. While its effectiveness varies across architectures, CARMA's key strength lies in reinforcing learned structures rather than introducing new capabilities, making it a scalable auxiliary method. These findings suggest that integrating CARMA with fine-tuning can improve compositional generalisation while maintaining task-specific performance in LLMs.