Recent advances in Parameter-Efficient Fine-Tuning (PEFT) have enabled effective adaptation of large-scale pre-trained models, particularly vision–language architectures such as CLIP, with substantially reduced training cost. In this paper, we propose Spectral \ (₅ₓ, \) a novel PEFT framework that explicitly exploits the spectral structure of pre-trained weight matrices for efficient multimodal adaptation. Our method performs a truncated singular value decomposition of each linear weight matrix and introduces trainable low-rank perturbations on the principal singular subspaces, while keeping the remaining spectral components frozen. The adapted weights are reconstructed in the spectral domain, ensuring that the initial model is strictly identical to the pre-trained model and that no additional inference-time latency is introduced. From a theoretical perspective, we show that spectral-domain perturbations enable a richer family of admissible weight updates than standard low-rank additive methods, achieving a higher rank capacity under an equal parameter budget. We evaluate Spectral \ (₅ₓ\) on multiple multimodal sentiment analysis benchmarks, including MVSA-S, MVSA-M, and HMF. Extensive experiments demonstrate that our approach consistently outperforms strong PEFT baselines such as Prompt-tuning, Adapter, LoRA, and DoRA in terms of classification accuracy and F1 score, while preserving the same parameter, computational, and memory efficiency as LoRA.
Qin et al. (Tue,) studied this question.