Organizations: Computer Vision and Biometrics Lab, Department of Information Technology, Indian Institute of Information Technology Allahabad, Prayagraj, Uttar Pradesh-211015, India
Vision Transformers (ViTs) achieve strong performance on image recognition tasks, yet their quadratic self-attention complexity over input tokens creates prohibitive computational costs for practical deployment. Token reduction methods offer a practical solution by progressively reducing the token sequence during the forward pass. We propose Saliency-Weighted Asymmetric Merging (SWAM), a training-based token merging framework that extends the Bipartite Soft Matching (BSM) with two saliency-aware mechanisms. We introduce a two-step matching formulation in which cosine similarity decides candidate pairs, while a multiplicative maximum saliency penalty regulates their selection for merging. We then further use asymmetric exponential saliency-weighted aggregation, which allows more salient tokens to contribute more strongly to the merged representation. A lightweight MLP-based saliency prediction module (75k parameters per stage, 1% of backbone parameters) supervised via attention rollout is trained end-to-end jointly with a DeiT-Small backbone using a composite cross-entropy, knowledge-distillation, and saliency-alignment objective. On CIFAR-100, TinyImageNet, and ImageNet-100, under a conservative schedule that reduces tokens from 196 to 100 patch tokens, with the CLS token retained throughout, 27.5% FLOPs reduction and 1.35 times speedup, SWAM matches or exceeds baseline across the reported results. Under an aggressive schedule (196 to 70 tokens with the CLS token kept throughout, 33.8% FLOPs reduction, and 1.46 times speedup), accuracy gains are achieved on ImageNet-100 (+0.23%) with minimal losses on CIFAR-100 and TinyImageNet.
Figures & tables
Fig. 1: SWAM pipeline. Top : DeiT-Small backbone with three SWAM reduction stages inserted after blocks 4, 7, and 10. Patch-token counts decrease from 196 to 100 (conservative, C) or 70 (aggressive, A) while the CLS token is retained throughout all reduction stages. Bottom : Detail of a single stage. The saliency head predicts per-token importance scores that feed into multiplicative max-saliency pair scoring (Eq. 6 ) and asymmetric saliency-weighted merging (Eq. 9 ).
Token merging combines redundant transformer tokens to reduce sequence length, attention costs, and inference time while preserving task-relevant information.
The quadratic cost of self-attention in Vision Transformers (ViTs) constitutes a fundamental bottleneck for practical deployment, motivating a vibrant line of research on token reduction. Among existing approaches, token merging (ToMe) has emerged as an elegant training-free solution; yet its design rests on an unspoken premise of token equality, which contravenes the well-documented non-uniformity of self-attention and leads to information loss in high-salience tokens under aggressive compression. We address this limitation with AdaMerge, a token-merging framework based on two complementary mechanisms. First, salience-weighted similarity leverages column-wise feature-affinity centrality as a token-importance proxy and incorporates the resulting salience scores into the bipartite matching score, ensuring that pivotal tokens contribute more strongly to the merged representation. Second, adaptive merging intensity uses pre-computed layer-wise similarity statistics to dynamically modulate the per-layer reduction count in accordance with input-specific redundancy. On ImageNet-1k with ViT-B/16, AdaMerge consistently outperforms ToMe, PiToMe, and DSM across all FLOPs-matched regimes. The accuracy gap widens monotonically with compression: at the 13.4G FLOPs operating point, AdaMerge sustains a Top-1 degradation of only -1.06%, compared to -1.45% for PiToMe and -4.62% for DSM. To our knowledge, AdaMerge is the first to combine salience-weighted similarity and adaptive per-layer reduction into a single training-free token merging framework, advancing the accuracy-FLOPs Pareto frontier of ViT acceleration.
Semi Lee, Hyejin Go, Hyesong Choi
Electronic Engineering · Soongsil University · Seoul, South Korea
Vision Transformers achieve strong performance but incur high computational costs from quadratic token interactions. Training-free token reduction methods improve efficiency, but their robustness remains underexplored. We propose Prune-and Merge ViT, a training-free framework for off-the-shelf models. It combines class attention-based pruning with similarity-based merging to reduce redundancy while preserving semantic information. To evaluate robustness we introduce two metrics. Relative Accuracy Drop (RAD) measures relative accuracy degradation under transformations, and Prediction Flip Rate (PFR) quantifies prediction consistency. On pretrained ViT and DeiT, Prune-and-Merge reduces computation competitively while maintaining robustness and better accuracy than existing token reduction methods. Our analysis suggests that hybrid token reduction does not introduce hallucination-like behavior and preserves prediction stability for common image transformations.
Vatsal, Soumil Malhotra, Priyanshu Chahal +1
Department of Mathematics, Indian Institute of Technology Roorkee
Vision Transformers (ViTs) face severe computational bottlenecks due to the quadratic complexity of self-attention at high resolutions. Existing token reduction methods rely on local metrics -- such as single-layer attention scores -- that are inherently vulnerable to the attention sink phenomenon, where uninformative tokens are paradoxically preserved over salient foreground objects. We propose ASAP (Attention Sink Anchored Pruning), a training-free framework that recasts this sink as a feature. Modeling ViT information flow as a Lazy Random Walk, ASAP identifies the sink as a dominant accumulator of probability mass. By computing the diffusion distance to the sink within the cumulative transition matrix, ASAP partitions tokens via Radial Diffusion Clustering and compresses background redundancy through Transition Weight Pooling in a single shot. Extensive experiments across image, video, and vision-language tasks demonstrate ASAP outperforms state-of-the-art methods, accelerating throughput by up to 48% while maintaining or even exceeding baseline accuracy.
Jaehyuk Lee, Hanyoung Kim, Yanggee Kim +1
Department of Mathematics, Korea University, Seoul, Republic of Korea · Korea University, Seoul, Republic of Korea