Organizations: University of Maryland, Department of Computer Science and UMIACS · Johns Hopkins University, Department of Electrical and Computer Engineering
Multimodal Transformers serve as the backbone for state-of-the-art vision-language models, yet their quadratic attention complexity remains a critical barrier to scalability. In this work, we investigate the viability of Linear Attention (LA) as a high-efficiency alternative within multimodal frameworks. By integrating LA, we reduce the computational overhead from quadratic to linear relative to sequence length while preserving competitive performance. We evaluate our approach across ViT-S/16, ViT-B/16, and ViT-L/16 architectures trained on the LAION-400M dataset, with validation focused on ImageNet-21K zero-shot accuracy. Our systematic evaluation demonstrates that Linear Attention not only yields significant computational savings but also adheres to the same scaling laws as standard softmax attention. These findings position Linear Attention as a robust, scalable solution for next-generation multimodal Transformers tasked with processing increasingly large and complex datasets.
Figures & tables
Figure 1: Heatmap visualization of attention scores aij of Eq. 4 and 7 . As i increases, the attention scores of Eq. 4 smooth-out, making it difficult for the model to distinguish relevant query and key pairs. The query and key are random matrices
Figure 2: Forward and backward pass time for a single attention layer with 8 heads, 64 dimensions per head, and batch size of 4 on an H200 GPU.
Figure 3: Training metrics for ViT-S/16, ViT-B/16, and ViT-L/16 models trained on LAION-400M using four NVIDIA A5500 GPUs. Linear attention with our adjustments achieves performance comparable to standard attention.
Model
Param
Layers
Width
Heads
ViT-S-16
22M
12
384
6
ViT-B-16
86M
12
512
8
ViT-L-16
304M
24
768
12
Table 1: Specifications of the models used in our study.
Figure 4: Training loss scaling with model size for ViTs on LAION-400M. Our adjusted LA follows a trend similar to standard attention.
This paper introduces Exact Linear Attention (ELA), a mechanism that achieves linear computational complexity for Transformer attention by exploiting the exact decomposition property of kernel functions, thereby eliminating approximation error. We identify and address two key limitations of prior linear attention -- gradient explosion and token attention dilution -- by imposing kernel constraints that ensure non-negativity, discriminability, and geometric interpretability. Several kernel functions are proposed, including the Hadamard Exp Kernel, Summation Squared Euclidean Distance Kernel, and Subtraction Squared Euclidean Distance Kernel, each tailored for specific attention behaviors. Beyond the core attention formulation, the paper presents three engineering innovations: (1) a Hyper-Link structure that replaces traditional residual connections to mitigate gradient degradation; (2) a Memory Lobe module based on bidirectional linear attention, which captures "transformation flow" across layers to implement qualitative memory and an implicit reinforcement learning paradigm; and (3) a routing-score-based bias mechanism for Mixture-of-Experts (MoE) to improve interpretability and semantic alignment. Experimental results demonstrate that ELA achieves up to 6x faster decoding speed and 75% reduction in KV cache memory usage compared to full attention, while maintaining comparable or superior training performance. The proposed memory module accelerates convergence and enhances generalization. Furthermore, we extend the linear attention principle to vision models, yielding YOLO-LAT, which attains up to 4.3x GPU inference speedup and 7.9x parameter reduction with competitive detection accuracy. These results underline the broad applicability of exact linear attention for scaling Transformer models to ultra-long sequences and efficient visual tasks.
Although large language models (LLMs) exhibit remarkable reasoning capabilities, their reliance on text-only pre-training restricts the perception of the multimodal physical world. Native multimodal pre-training avoids this limitation by training models from scratch on multimodal inputs, thereby achieving deep cross-modal integration and mitigating optimization asymmetries inherent to traditional late-fusion architectures. Despite these advantages, the scaling properties of this paradigm remain incompletely characterized. To address this gap, we investigate the optimal model size and token count for training a Transformer-based vision-language model under a fixed computational budget. Our study demonstrates that minimal objective loss adheres to a predictable compute law, whereas compute-optimal model sizes and token counts scale as power laws. Notably, language and multimodal objectives manifest distinct allocation trends. The language allocation exponents lie in a similar range across the different data mixtures. The multimodal model-allocation exponent decreases modestly with the multimodal token ratio, indicating a relative shift toward token allocation. Additionally, our scaling analysis yields a budget-compensation rule. Specifically, an additional multimodal-token budget can offset the text-efficiency penalty caused by incorporating visual information into native multimodal pre-training under a fixed compute budget. Downstream evaluations further reveal that native multimodal pre-training is associated with improved spatial reasoning and multimodal few-shot learning. Generally, this empirical research establishes the essential groundwork for predictably scaling multimodal foundation models.
Haoyuan Wu, Aoqi Wu, Hai Wang +3
The Chinese University of Hong Kong · LLM Department, Tencent
Looped Transformers (LT) have emerged as a powerful architecture by iterating their layers multiple times before decoding the final token. However, pairing them with full attention retains quadratic complexity, making them computationally expensive and slow. We introduce LT2 (Linear-Time Looped Transformers), a family of looped architectures that replace quadratic softmax attention with subquadratic, linear-time attention. We study two variants: LT2-linear with linear attention and LT2-sparse with sparse attention. We find that looping uniquely synergizes with these variants: it enables iterative memory refinement in linear attention and progressively expands the effective receptive field in sparse attention. We formalize these benefits theoretically and demonstrate consistent empirical gains across controlled recall, state-tracking, and language modeling tasks. We then explore LT2-hybrid, which combines different attention variants in a looped setting. Two variants are especially promising: LT2-hybrid (GDN+DSA), which interleaves linear and sparse attention to maximize efficiency and matches the standard looped transformer's quality at fully linear-time cost; and LT2-hybrid (Full+GDN), which interleaves GDN with a small fraction of full attention layers to maximize quality, surpassing the standard looped transformer in both performance and efficiency. We also show how to convert a pre-trained LT into an LT2-hybrid model. With about 1B tokens of training, our converted model, Ouro-hybrid-1.4B, outperforms industry-level 1B models and is competitive with industry-level 4B models while retaining the speed benefits of linear-time attention. Together, these results show a clear path toward making looped transformers more scalable and advancing efficient, capable small language models.