SEARCH · Search arXiv

Results for “transformer architecture”

Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 451 records · Page 25Linked to original sources

arXiv2025

LM2: Large Memory Models

This paper introduces the Large Memory Model (LM2), a decoder-only Transformer architecture enhanced with an auxiliary memory module that aims to address the limitations of standard Transformers in multi-step reasoning, relational argumentation, and synthesizing information distributed over long contexts. The proposed LM2 incorporates a memory module that acts as a contextual representation repository, interacting with input tokens via cross attention and updating through gating mechanisms. To preserve the Transformers general-purpose capabilities, LM2 maintains the original information flow while integrating a complementary memory pathway. Experimental results on the BABILong benchmark demonstrate that the LM2model outperforms both the memory-augmented RMT model by 37.1% and the baseline Llama-3.2 model by 86.3% on average across tasks. LM2 exhibits exceptional capabilities in multi-hop inference, numerical reasoning, and large-context question-answering. On the MMLU dataset, it achieves a 5.0% improvement over a pre-trained vanilla model, demonstrating that its memory module does not degrade performance on general tasks. Further, in our analysis, we explore the memory interpretability, effectiveness of memory modules, and test-time behavior. Our findings emphasize the importance of explicit memory in enhancing Transformer architectures.

cs.CL↗

arXiv2025

Adaptive Rank Allocation: Speeding Up Modern Transformers with RaNA Adapters

Large Language Models (LLMs) are computationally intensive, particularly during inference. Neuron-adaptive techniques, which selectively activate neurons in Multi-Layer Perceptron (MLP) layers, offer some speedups but suffer from limitations in modern Transformers. These include reliance on sparse activations, incompatibility with attention layers, and the use of costly neuron masking techniques. To address these issues, we propose the Adaptive Rank Allocation framework and introduce the Rank and Neuron Allocator (RaNA) adapter. RaNA adapters leverage rank adapters, which operate on linear layers by applying both low-rank matrix decompositions and adaptive masking to efficiently allocate compute without depending on activation sparsity. This enables RaNA to be generally applied to MLPs and linear components of attention modules, while eliminating the need for expensive maskers found in neuron-adaptive methods. Notably, when compared to neuron adapters, RaNA improves perplexity by up to 7 points and increases accuracy by up to 8 percentage-points when reducing FLOPs by $\sim$44% in state-of-the-art Transformer architectures. These results position RaNA as a robust solution for improving inference efficiency in modern Transformer architectures.

cs.LG↗

arXiv2025

Tracr-Injection: Distilling Algorithms into Pre-trained Language Models

Motivated by the surge of large language models, there has been a push to formally characterize the symbolic abilities intrinsic to the transformer architecture. A programming language, called RASP, has been proposed, which can be directly compiled into transformer weights to implement these algorithms. However, the tasks that can be implemented in RASP are often uncommon to learn from natural unsupervised data, showing a mismatch between theoretical capabilities of the transformer architecture, and the practical learnability of these capabilities from unsupervised data. We propose tracr-injection, a method that allows us to distill algorithms written in RASP directly into a pre-trained language model. We showcase our method by injecting 3 different algorithms into a language model. We show how our method creates an interpretable subspace within the model's residual stream, which can be decoded into the variables present in the code of the RASP algorithm. Additionally, we found that the proposed method can improve out-of-distribution performance compared to our baseline, indicating that indeed a more symbolic mechanism is taking place in the inner workings of the model. We release the code used to run our experiments.

cs.CL↗

arXiv2022

TransMorph: Transformer for unsupervised medical image registration

In the last decade, convolutional neural networks (ConvNets) have been a major focus of research in medical image analysis. However, the performances of ConvNets may be limited by a lack of explicit consideration of the long-range spatial relationships in an image. Recently Vision Transformer architectures have been proposed to address the shortcomings of ConvNets and have produced state-of-the-art performances in many medical imaging applications. Transformers may be a strong candidate for image registration because their substantially larger receptive field enables a more precise comprehension of the spatial correspondence between moving and fixed images. Here, we present TransMorph, a hybrid Transformer-ConvNet model for volumetric medical image registration. This paper also presents diffeomorphic and Bayesian variants of TransMorph: the diffeomorphic variants ensure the topology-preserving deformations, and the Bayesian variant produces a well-calibrated registration uncertainty estimate. We extensively validated the proposed models using 3D medical images from three applications: inter-patient and atlas-to-patient brain MRI registration and phantom-to-CT registration. The proposed models are evaluated in comparison to a variety of existing registration methods and Transformer architectures. Qualitative and quantitative results demonstrate that the proposed Transformer-based model leads to a substantial performance improvement over the baseline methods, confirming the effectiveness of Transformers for medical image registration.

eess.IV↗

arXiv2022

Swin-Pose: Swin Transformer Based Human Pose Estimation

Convolutional neural networks (CNNs) have been widely utilized in many computer vision tasks. However, CNNs have a fixed reception field and lack the ability of long-range perception, which is crucial to human pose estimation. Due to its capability to capture long-range dependencies between pixels, transformer architecture has been adopted to computer vision applications recently and is proven to be a highly effective architecture. We are interested in exploring its capability in human pose estimation, and thus propose a novel model based on transformer architecture, enhanced with a feature pyramid fusion structure. More specifically, we use pre-trained Swin Transformer as our backbone and extract features from input images, we leverage a feature pyramid structure to extract feature maps from different stages. By fusing the features together, our model predicts the keypoint heatmap. The experiment results of our study have demonstrated that the proposed transformer-based model can achieve better performance compared to the state-of-the-art CNN-based models.

cs.CV↗

arXiv2025

Theoretical Guarantees for LT-TTD: A Unified Transformer-based Architecture for Two-Level Ranking Systems

Modern recommendation and search systems typically employ multi-stage ranking architectures to efficiently handle billions of candidates. The conventional approach uses distinct L1 (candidate retrieval) and L2 (re-ranking) models with different optimization objectives, introducing critical limitations including irreversible error propagation and suboptimal ranking. This paper identifies and analyzes the fundamental limitations of this decoupled paradigm and proposes LT-TTD (Listwise Transformer with Two-Tower Distillation), a novel unified architecture that bridges retrieval and ranking phases. Our approach combines the computational efficiency of two-tower models with the expressivity of transformers in a unified listwise learning framework. We provide a comprehensive theoretical analysis of our architecture and establish formal guarantees regarding error propagation mitigation, ranking quality improvements, and optimization convergence. We derive theoretical bounds showing that LT-TTD reduces the upper limit on irretrievable relevant items by a factor that depends on the knowledge distillation strength, and prove that our multi-objective optimization framework achieves a provably better global optimum than disjoint training. Additionally, we analyze the computational complexity of our approach, demonstrating that the asymptotic complexity remains within practical bounds for real-world applications. We also introduce UPQE, a novel evaluation metric specifically designed for unified ranking architectures that holistically captures retrieval quality, ranking performance, and computational efficiency.

cs.IR↗

arXiv2023

TRAMS: Training-free Memory Selection for Long-range Language Modeling

The Transformer architecture is crucial for numerous AI models, but it still faces challenges in long-range language modeling. Though several specific transformer architectures have been designed to tackle issues of long-range dependencies, existing methods like Transformer-XL are plagued by a high percentage of ineffective memories. In this study, we present a plug-and-play strategy, known as TRAining-free Memory Selection (TRAMS), that selects tokens participating in attention calculation based on one simple metric. This strategy allows us to keep tokens that are likely to have a high attention score with the current queries and ignore the other ones. We have tested our approach on the word-level benchmark (WikiText-103) and the character-level benchmark (enwik8), and the results indicate an improvement without having additional training or adding additional parameters.

cs.CL↗

arXiv2025

SST: Multi-Scale Hybrid Mamba-Transformer Experts for Time Series Forecasting

Time series forecasting has made significant advances, including with Transformer-based models. The attention mechanism in Transformer effectively captures temporal dependencies by attending to all past inputs simultaneously. However, its quadratic complexity with respect to sequence length limits the scalability for long-range modeling. Recent state space models (SSMs) such as Mamba offer a promising alternative by achieving linear complexity without attention. Yet, Mamba compresses historical information into a fixed-size latent state, potentially causing information loss and limiting representational effectiveness. This raises a key research question: Can we design a hybrid Mamba-Transformer architecture that is both effective and efficient for time series forecasting? To address it, we adapt a hybrid Mamba-Transformer architecture Mambaformer, originally proposed for language modeling, to the time series domain. Preliminary experiments reveal that naively stacking Mamba and Transformer layers in Mambaformer is suboptimal for time series forecasting, due to an information interference problem. To mitigate this issue, we introduce a new time series decomposition strategy that separates time series into long-range patterns and short-range variations. Then we show that Mamba excels at capturing long-term structures, while Transformer is more effective at modeling short-term dynamics. Building on this insight, we propose State Space Transformer (SST), a multi-scale hybrid model with expert modules: a Mamba expert for long-range patterns and a Transformer expert for short-term variations. SST also employs a multi-scale patching mechanism to adaptively adjust time series resolution: low resolution for long-term patterns and high resolution for short-term variations. Experiments show that SST obtains SOTA performance with linear scalability. The code is at https://github.com/XiongxiaoXu/SST.

cs.LG↗

arXiv2024

OTCE: Hybrid SSM and Attention with Cross Domain Mixture of Experts to construct Observer-Thinker-Conceiver-Expresser

Recent research has shown that combining Mamba with Transformer architecture, which has selective state space and quadratic self-attention mechanism, outperforms using Mamba or Transformer architecture alone in language modeling tasks. The quadratic self-attention mechanism effectively alleviates the shortcomings of selective state space in handling long-term dependencies of any element in the sequence. We propose a position information injection method that connects the selective state space model with the quadratic attention, and integrates these two architectures with hybrid experts with cross-sharing domains, so that we can enjoy the advantages of both. We design a new architecture with a more biomimetic idea: Observer-Thinker-Conceiver-Expresser (OTCE), which can compete with well-known medium-scale open-source language models on a small scale in language modeling tasks.

cs.CL↗

arXiv2026

Retraining-Free Depth Pruning of Vision Transformers: A Cross Domain Study of Layer Redundancy

Transformer architectures have achieved strong results in natural language processing and computer vision, but their deployment is limited by high computational, memory and storage costs, motivating research into model compression techniques, such as pruning, while preserving performance. Prior work has shown substantial redundancy in large language models (LLMs), and redundancy in vision transformers has also been investigated, but it remains less systematically analyzed than in LLMs. In this work, we extend similarity-based redundancy analysis to vision models by measuring cosine similarity between layer inputs and outputs to identify transformations with minimal impact. We study redundancy patterns across several transformer architectures and find that deep vision transformers such as DINOv2 and SwinV2, like language models, exhibit significant redundancy. We then evaluate how this redundancy can be exploited to improve efficiency, analyzing the trade-off between inference speed and performance degradation under different pruning strategies. Our results show that moderate pruning can yield substantial acceleration with limited impact on accuracy, and that 20% of the DINOv2 blocks can be removed without any retraining, for a 1.24x throughput gain and a 20% lower energy per image at a relative accuracy loss below 1% on ImageNet-1K. We further propose a model selection pipeline that automatically identifies the best pruned configuration for a given dataset, reducing FLOPs by 13% to 27% on average across architectures with an average accuracy change between -0.3 and +1.0 points.

cs.CV↗

arXiv2026

Towards Tight Bounds for Streaming Attention

The attention mechanism is a cornerstone of modern transformer architectures. However, its expressive power comes at the cost of quadratic runtime and linear space usage. In particular, the classical transformer architecture explicitly stores all previously seen input elements (tokens) in order to generate the next one. The problem of implementing a transformer in limited space, known as KV cache compression, has received much interest over the past few years, spurring the development of powerful heuristics. Recent works of Haris et al, COLT'25 and Kochetkova et al, NeurIPS'25, formalized KV cache compression as the streaming attention approximation problem, providing both upper bounds (based on discrepancy theory) and information theoretic lower bounds. However, those papers left open a significant gap between the upper and lower bounds. For example, the space usage of their algorithms increases with the precision parameter, but the lower bound does not get stronger. In this work, we revisit the streaming attention approximation problem and provide nearly tight bounds on its space complexity. On the algorithmic side, we achieve the result through a surprisingly tight interplay between three distinct methods for kernel density estimation: discrepancy-based coreset constructions (e.g., Charikar-Kapralov-Waingarten'24), the polynomial method (e.g., Greengard-Rokhlin'87, Alman-Song'23), and space partitioning (e.g., Andoni-Laarhoven-Razenshteyn-Waingarten'17, Charikar-Kapralov-Nouri-Siminelakis'20). On the lower bound side, our main technical contribution is a new technique for using the INDEX problem with a large amount of side information that we hope will prove useful in other high dimensional geometric estimation problems.

cs.DS↗

arXiv2025

Future-Aware Interaction Network For Motion Forecasting

Motion forecasting is a crucial component of autonomous driving systems, enabling the generation of accurate and smooth future trajectories to ensure safe navigation to the destination. In previous methods, potential future trajectories are often absent in the scene encoding stage, which may lead to suboptimal outcomes. Additionally, prior approaches typically employ transformer architectures for spatiotemporal modeling of trajectories and map information, which suffer from the quadratic scaling complexity of the transformer architecture. In this work, we propose an interaction-based method, named Future-Aware Interaction Network, that introduces potential future trajectories into scene encoding for a comprehensive traffic representation. Furthermore, a State Space Model (SSM), specifically Mamba, is introduced for both spatial and temporal modeling. To adapt Mamba for spatial interaction modeling, we propose an adaptive reordering strategy that transforms unordered data into a structured sequence. Additionally, Mamba is employed to refine generated future trajectories temporally, ensuring more consistent predictions. These enhancements not only improve model efficiency but also enhance the accuracy and diversity of predictions. We conduct comprehensive experiments on the widely used Argoverse 1 and Argoverse 2 datasets, demonstrating that the proposed method achieves superior performance compared to previous approaches in a more efficient way. The code will be released according to the acceptance.

cs.CV↗

arXiv2026

Next-Token Prediction and Regret Minimization

We consider the question of how to employ next-token prediction algorithms in adversarial online decision-making environments. Specifically, if we train a next-token prediction model on a distribution $\mathcal{D}$ over sequences of opponent actions, when is it the case that the induced online decision-making algorithm (by approximately best responding to the model's predictions) has low adversarial regret (i.e., when is $\mathcal{D}$ a \emph{low-regret distribution})? For unbounded context windows (where the prediction made by the model can depend on all the actions taken by the adversary thus far), we show that although not every distribution $\mathcal{D}$ is a low-regret distribution, every distribution $\mathcal{D}$ is exponentially close (in TV distance) to one low-regret distribution, and hence sublinear regret can always be achieved at negligible cost to the accuracy of the original next-token prediction model. In contrast to this, for bounded context windows (where the prediction made by the model can depend only on the past $w$ actions taken by the adversary, as may be the case in modern transformer architectures), we show that there are some distributions $\mathcal{D}$ of opponent play that are $Θ(1)$-far from any low-regret distribution $\mathcal{D'}$ (even when $w = Ω(T)$ and such distributions exist). Finally, we complement these results by showing that the unbounded context robustification procedure can be implemented by layers of a standard transformer architecture, and provide empirical evidence that transformer models can be efficiently trained to represent these new low-regret distributions.

cs.LG↗

arXiv2025

Neural ODE Transformers: Analyzing Internal Dynamics and Adaptive Fine-tuning

Recent advancements in large language models (LLMs) based on transformer architectures have sparked significant interest in understanding their inner workings. In this paper, we introduce a novel approach to modeling transformer architectures using highly flexible non-autonomous neural ordinary differential equations (ODEs). Our proposed model parameterizes all weights of attention and feed-forward blocks through neural networks, expressing these weights as functions of a continuous layer index. Through spectral analysis of the model's dynamics, we uncover an increase in eigenvalue magnitude that challenges the weight-sharing assumption prevalent in existing theoretical studies. We also leverage the Lyapunov exponent to examine token-level sensitivity, enhancing model interpretability. Our neural ODE transformer demonstrates performance comparable to or better than vanilla transformers across various configurations and datasets, while offering flexible fine-tuning capabilities that can adapt to different architectural constraints.

cs.LG↗

arXiv2022

Analysis of GraphSum's Attention Weights to Improve the Explainability of Multi-Document Summarization

Modern multi-document summarization (MDS) methods are based on transformer architectures. They generate state of the art summaries, but lack explainability. We focus on graph-based transformer models for MDS as they gained recent popularity. We aim to improve the explainability of the graph-based MDS by analyzing their attention weights. In a graph-based MDS such as GraphSum, vertices represent the textual units, while the edges form some similarity graph over the units. We compare GraphSum's performance utilizing different textual units, i. e., sentences versus paragraphs, on two news benchmark datasets, namely WikiSum and MultiNews. Our experiments show that paragraph-level representations provide the best summarization performance. Thus, we subsequently focus oAnalysisn analyzing the paragraph-level attention weights of GraphSum's multi-heads and decoding layers in order to improve the explainability of a transformer-based MDS model. As a reference metric, we calculate the ROUGE scores between the input paragraphs and each sentence in the generated summary, which indicate source origin information via text similarity. We observe a high correlation between the attention weights and this reference metric, especially on the the later decoding layers of the transformer architecture. Finally, we investigate if the generated summaries follow a pattern of positional bias by extracting which paragraph provided the most information for each generated summary. Our results show that there is a high correlation between the position in the summary and the source origin.

cs.CL↗

arXiv2024

GVT2RPM: An Empirical Study for General Video Transformer Adaptation to Remote Physiological Measurement

Remote physiological measurement (RPM) is an essential tool for healthcare monitoring as it enables the measurement of physiological signs, e.g., heart rate, in a remote setting via physical wearables. Recently, with facial videos, we have seen rapid advancements in video-based RPMs. However, adopting facial videos for RPM in the clinical setting largely depends on the accuracy and robustness (work across patient populations). Fortunately, the capability of the state-of-the-art transformer architecture in general (natural) video understanding has resulted in marked improvements and has been translated to facial understanding, including RPM. However, existing RPM methods usually need RPM-specific modules, e.g., temporal difference convolution and handcrafted feature maps. Although these customized modules can increase accuracy, they are not demonstrated for their robustness across datasets. Further, due to their customization of the transformer architecture, they cannot use the advancements made in general video transformers (GVT). In this study, we interrogate the GVT architecture and empirically analyze how the training designs, i.e., data pre-processing and network configurations, affect the model performance applied to RPM. Based on the structure of video transformers, we propose to configure its spatiotemporal hierarchy to align with the dense temporal information needed in RPM for signal feature extraction. We define several practical guidelines and gradually adapt GVTs for RPM without introducing RPM-specific modules. Our experiments demonstrate favorable results to existing RPM-specific module counterparts. We conducted extensive experiments with five datasets using intra-dataset and cross-dataset settings. We highlight that the proposed guidelines GVT2RPM can be generalized to any video transformers and is robust to various datasets.

cs.CV↗

arXiv2026

Improving Prostate Gland Segmentation Using Transformer based Architectures

Inter reader variability and cross site domain shift challenge the automatic segmentation of prostate anatomy using T2 weighted MRI images. This study investigates whether transformer models can retain precision amid such heterogeneity. We compare the performance of UNETR and SwinUNETR in prostate gland segmentation against our previous 3D UNet model [1], based on 546 MRI (T2weighted) volumes annotated by two independent experts. Three training strategies were analyzed: single cohort dataset, 5 fold cross validated mixed cohort, and gland size based dataset. Hyperparameters were tuned by Optuna. The test set, from an independent population of readers, served as the evaluation endpoint (Dice Similarity Coefficient). In single reader training, SwinUNETR achieved an average dice score of 0.816 for Reader#1 and 0.860 for Reader#2, while UNETR scored 0.8 and 0.833 for Readers #1 and #2, respectively, compared to the baseline UNets 0.825 for Reader #1 and 0.851 for Reader #2. SwinUNETR had an average dice score of 0.8583 for Reader#1 and 0.867 for Reader#2 in cross-validated mixed training. For the gland size-based dataset, SwinUNETR achieved an average dice score of 0.902 for Reader#1 subset and 0.894 for Reader#2, using the five-fold mixed training strategy (Reader#1, n=53; Reader#2, n=87) at larger gland size-based subsets, where UNETR performed poorly. Our findings demonstrate that global and shifted-window self-attention effectively reduces label noise and class imbalance sensitivity, resulting in improvements in the Dice score over CNNs by up to five points while maintaining computational efficiency. This contributes to the high robustness of SwinUNETR for clinical deployment.

eess.IV↗

arXiv2025

FinchGPT: a Transformer based language model for birdsong analysis

The long-range dependencies among the tokens, which originate from hierarchical structures, are a defining hallmark of human language. However, whether similar dependencies exist within the sequential vocalization of non-human animals remains a topic of investigation. Transformer architectures, known for their ability to model long-range dependencies among tokens, provide a powerful tool for investigating this phenomenon. In this study, we employed the Transformer architecture to analyze the songs of Bengalese finch (Lonchura striata domestica), which are characterized by their highly variable and complex syllable sequences. To this end, we developed FinchGPT, a Transformer-based model trained on a textualized corpus of birdsongs, which outperformed other architecture models in this domain. Attention weight analysis revealed that FinchGPT effectively captures long-range dependencies within syllables sequences. Furthermore, reverse engineering approaches demonstrated the impact of computational and biological manipulations on its performance: restricting FinchGPT's attention span and disrupting birdsong syntax through the ablation of specific brain nuclei markedly influenced the model's outputs. Our study highlights the transformative potential of large language models (LLMs) in deciphering the complexities of animal vocalizations, offering a novel framework for exploring the structural properties of non-human communication systems while shedding light on the computational distinctions between biological brains and artificial neural networks.

cs.CL↗