Does Combining Parameter‐Efficient Modules Improve Few‐Shot Transfer Accuracy?

ABSTRACT Parameter‐efficient fine‐tuning has become a key technique for adapting large language and vision models to diverse downstream tasks, including emerging applications in engineering, such as design optimisation, fault detection in complex systems, material property prediction and autonomous robotics control. Specifically, the efficiency of low‐rank adaptation has facilitated the creation and sharing of hundreds of custom low‐rank adaptation (LoRA) modules, each trained on distinct data from various downstream tasks. In this paper, we explore the composability of LoRA modules, examining if combining these pre‐trained modules enhances generalisation to unseen downstream tasks. Our investigation involves evaluating two approaches: (a) uniform composition , involving averaging upstream LoRA modules with equal weights and (b) learned composition , where we learn the weights for each upstream module and perform weighted averaging. Through experiments on both vision and language models, we demonstrate that in few‐shot settings—common in many engineering workflows such as structural design, material discovery and system diagnostics—both composition methods consistently outperform full fine‐tuning and training LoRA from scratch. Moreover, in full‐shot settings, learned composition matches the performance of standard LoRA training while requiring significantly fewer trainable parameters. These results highlight the potential of compositional parameter‐efficient adaptation as a scalable and flexible approach for AI‐driven engineering tasks, enabling faster deployment, reduced data requirements and improved model reusability across diverse domains.