The growth of large language models underscores the need for parameter-efficient fine-tuning. Despite its popularity, LoRA encounters storage and computational challenges when deploying multiple task- or user-specific modules. To address this, we introduce LoRA-XS, a novel fine-tuning method backed by a theoretical derivation. LoRA-XS drastically reduces trainable parameters by incorporating a small, trainable weight matrix between frozen low-rank matrices derived from the Singular Value Decomposition of pre-trained weights. This design enables LoRA-XS to reduce storage requirements by over 100x in 7B models compared to LoRA. Additionally, unlike other methods, LoRA-XS imposes no lower bound on trainable parameters – it can scale from a single parameter per module to arbitrarily large values, adapting to any storage or computational constraint. Evaluations on GLUE, GSM8K, MATH, and commonsense reasoning benchmarks across different model scales reveal that LoRA-XS consistently outperforms or matches LoRA and VeRA in accuracy, offering unmatched parameter efficiency. Our ablation studies highlight the significance of singular vectors in transformer weights, establishing LoRA-XS as a powerful, storage-efficient solution for scaling and personalizing large language models.

We propose a novel gradient-based attack against transformer-based language models that searches for an adversarial example in a continuous space of token probabilities.Our algorithm mitigates the gap between adversarial loss for continuous and discrete text representations by performing multi-step quantization in a quantization-compensation loop.Experiments show that our method significantly outperforms other approaches on various natural language processing (NLP) tasks.

The problem of reducing processing time of large deep learning models is a fundamental challenge in many real-world applications. Early exit methods strive towards this goal by attaching additional Internal Classifiers (ICs) to intermediate layers of a neural network. ICs can quickly return predictions for easy examples and, as a result, reduce the average inference time of the whole model. However, if a particular IC does not decide to return an answer early, its predictions are discarded, with its computations effectively being wasted. To solve this issue, we introduce Zero Time Waste (ZTW), a novel approach in which each IC reuses predictions returned by its predecessors by (1) adding direct connections between ICs and (2) combining previous outputs in an ensemble-like manner. We conduct extensive experiments across various multiple modes, datasets, and architectures to demonstrate that ZTW achieves a significantly better accuracy vs. inference time trade-off than other early exit methods. On the ImageNet dataset, it obtains superior results over the best baseline method in 11 out of 16 cases, reaching up to 5 percentage points of improvement on low computational budgets.

The problem of reducing processing time of large deep learning models is a fundamental challenge in many real-world applications. Early exit methods strive towards this goal by attaching additional Internal Classifiers (ICs) to intermediate layers of a neural network. ICs can quickly return predictions for easy examples and, as a result, reduce the average inference time of the whole model. However, if a particular IC does not decide to return an answer early, its predictions are discarded, with its computations effectively being wasted. To solve this issue, we introduce Zero Time Waste (ZTW), a novel approach in which each IC reuses predictions returned by its predecessors by (1) adding direct connections between ICs and (2) combining previous outputs in an ensemble-like manner. We conduct extensive experiments across various datasets and architectures to demonstrate that ZTW achieves a significantly better accuracy vs. inference time trade-off than other recently proposed early exit methods.