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 1,603 records · Page 89Linked to original sources

arXiv2024

The Mysterious Case of Neuron 1512: Injectable Realignment Architectures Reveal Internal Characteristics of Meta's Llama 2 Model

Large Language Models (LLMs) have an unrivaled and invaluable ability to "align" their output to a diverse range of human preferences, by mirroring them in the text they generate. The internal characteristics of such models, however, remain largely opaque. This work presents the Injectable Realignment Model (IRM) as a novel approach to language model interpretability and explainability. Inspired by earlier work on Neural Programming Interfaces, we construct and train a small network -- the IRM -- to induce emotion-based alignments within a 7B parameter LLM architecture. The IRM outputs are injected via layerwise addition at various points during the LLM's forward pass, thus modulating its behavior without changing the weights of the original model. This isolates the alignment behavior from the complex mechanisms of the transformer model. Analysis of the trained IRM's outputs reveals a curious pattern. Across more than 24 training runs and multiple alignment datasets, patterns of IRM activations align themselves in striations associated with a neuron's index within each transformer layer, rather than being associated with the layers themselves. Further, a single neuron index (1512) is strongly correlated with all tested alignments. This result, although initially counterintuitive, is directly attributable to design choices present within almost all commercially available transformer architectures, and highlights a potential weak point in Meta's pretrained Llama 2 models. It also demonstrates the value of the IRM architecture for language model analysis and interpretability. Our code and datasets are available at https://github.com/DRAGNLabs/injectable-alignment-model

cs.CL↗

arXiv2025

Optimality and NP-Hardness of Transformers in Learning Markovian Dynamical Functions

Transformer architectures can solve unseen tasks based on input-output pairs in a given prompt due to in-context learning (ICL). Existing theoretical studies on ICL have mainly focused on linear regression tasks, often with i.i.d. inputs. To understand how transformers express ICL when modeling dynamics-driven functions, we investigate Markovian function learning through a structured ICL setup, where we characterize the loss landscape to reveal underlying optimization behaviors. Specifically, we (1) provide the closed-form expression of the global minimizer (in an enlarged parameter space) for a single-layer linear self-attention (LSA) model; (2) prove that recovering transformer parameters that realize the optimal solution is NP-hard in general, revealing a fundamental limitation of one-layer LSA in representing structured dynamical functions; and (3) supply a novel interpretation of a multilayer LSA as performing preconditioned gradient descent to optimize multiple objectives beyond the square loss. These theoretical results are numerically validated using simplified transformers.

cs.LG↗

arXiv2025

EMTSF:Extraordinary Mixture of SOTA Models for Time Series Forecasting

The immense success of the Transformer architecture in Natural Language Processing has led to its adoption in Time Se ries Forecasting (TSF), where superior performance has been shown. However, a recent important paper questioned their effectiveness by demonstrating that a simple single layer linear model outperforms Transformer-based models. This was soon shown to be not as valid, by a better transformer-based model termed PatchTST. More re cently, TimeLLM demonstrated even better results by repurposing a Large Language Model (LLM) for the TSF domain. Again, a follow up paper challenged this by demonstrating that removing the LLM component or replacing it with a basic attention layer in fact yields better performance. One of the challenges in forecasting is the fact that TSF data favors the more recent past, and is sometimes subject to unpredictable events. Based upon these recent insights in TSF, we propose a strong Mixture of Experts (MoE) framework. Our method combines the state-of-the-art (SOTA) models including xLSTM, en hanced Linear, PatchTST, and minGRU, among others. This set of complimentary and diverse models for TSF are integrated in a Trans former based MoE gating network. Our proposed model outperforms all existing TSF models on standard benchmarks, surpassing even the latest approaches based on MoE frameworks.

cs.CL↗

arXiv2026

Leviathan: Decoupling Input and Output Representations in Language Models

Modern language models use a single matrix for input embedding and output projection. This couples two distinct objectives: token representation and discrimination over a vocabulary. This work introduces Leviathan, a Transformer architecture that replaces the input embedding matrix with learned embedding vectorization (LEV), a compact continuous mapping from token indices to embeddings. Leviathan's output head remains untied for a parameter increase of as low as 0.2%. Under controlled comparisons with identical Transformer backbones, Leviathan consistently improves language modeling performance over standard tied-embedding baselines across a 200M-1.2B parameter regime on The Pile with gains that grow during training. At 1.2B scale, Leviathan reduces validation perplexity by 9%, requires $2.1\times$ fewer training tokens to reach the tied baseline's final loss, and improves on all six downstream benchmarks evaluated, including a 30% reduction in LAMBADA perplexity. Frequency-stratified analysis reveals gains to be concentrated in rare tokens, where continuous parameterization reduces perplexity by 81%, falling to near zero for the most frequent.

cs.CL↗

arXiv2025

Graph Embedding with Mel-spectrograms for Underwater Acoustic Target Recognition

Underwater acoustic target recognition (UATR) is extremely challenging due to the complexity of ship-radiated noise and the variability of ocean environments. Although deep learning (DL) approaches have achieved promising results, most existing models implicitly assume that underwater acoustic data lie in a Euclidean space. This assumption, however, is unsuitable for the inherently complex topology of underwater acoustic signals, which exhibit non-stationary, non-Gaussian, and nonlinear characteristics. To overcome this limitation, this paper proposes the UATR-GTransformer, a non-Euclidean DL model that integrates Transformer architectures with graph neural networks (GNNs). The model comprises three key components: a Mel patchify block, a GTransformer block, and a classification head. The Mel patchify block partitions the Mel-spectrogram into overlapping patches, while the GTransformer block employs a Transformer Encoder to capture mutual information between split patches to generate Mel-graph embeddings. Subsequently, a GNN enhances these embeddings by modeling local neighborhood relationships, and a feed-forward network (FFN) further performs feature transformation. Experiments results based on two widely used benchmark datasets demonstrate that the UATR-GTransformer achieves performance competitive with state-of-the-art methods. In addition, interpretability analysis reveals that the proposed model effectively extracts rich frequency-domain information, highlighting its potential for applications in ocean engineering.

cs.SD↗

arXiv2026

AutoRegressive Generation with B-rep Holistic Token Sequence Representation

Previous representation and generation approaches for the B-rep relied on graph-based representations that disentangle geometric and topological features through decoupled computational pipelines, thereby precluding the application of sequence-based generative frameworks, such as transformer architectures that have demonstrated remarkable performance. In this paper, we propose BrepARG, the first attempt to encode B-rep's geometry and topology into a holistic token sequence representation, enabling sequence-based B-rep generation with an autoregressive architecture. Specifically, BrepARG encodes B-rep into 3 types of tokens: geometry and position tokens representing geometric features, and face index tokens representing topology. Then the holistic token sequence is constructed hierarchically, starting with constructing the geometry blocks (i.e., faces and edges) using the above tokens, followed by geometry block sequencing. Finally, we assemble the holistic sequence representation for the entire B-rep. We also construct a transformer-based autoregressive model that learns the distribution over holistic token sequences via next-token prediction, using a multi-layer decoder-only architecture with causal masking. Experiments demonstrate that BrepARG achieves state-of-the-art (SOTA) performance. BrepARG validates the feasibility of representing B-rep as holistic token sequences, opening new directions for B-rep generation.

cs.CV↗

arXiv2019

Question Generation from Paragraphs: A Tale of Two Hierarchical Models

Automatic question generation from paragraphs is an important and challenging problem, particularly due to the long context from paragraphs. In this paper, we propose and study two hierarchical models for the task of question generation from paragraphs. Specifically, we propose (a) a novel hierarchical BiLSTM model with selective attention and (b) a novel hierarchical Transformer architecture, both of which learn hierarchical representations of paragraphs. We model a paragraph in terms of its constituent sentences, and a sentence in terms of its constituent words. While the introduction of the attention mechanism benefits the hierarchical BiLSTM model, the hierarchical Transformer, with its inherent attention and positional encoding mechanisms also performs better than flat transformer model. We conducted empirical evaluation on the widely used SQuAD and MS MARCO datasets using standard metrics. The results demonstrate the overall effectiveness of the hierarchical models over their flat counterparts. Qualitatively, our hierarchical models are able to generate fluent and relevant questions

cs.CL↗

arXiv2020

Solomon at SemEval-2020 Task 11: Ensemble Architecture for Fine-Tuned Propaganda Detection in News Articles

This paper describes our system (Solomon) details and results of participation in the SemEval 2020 Task 11 "Detection of Propaganda Techniques in News Articles"\cite{DaSanMartinoSemeval20task11}. We participated in Task "Technique Classification" (TC) which is a multi-class classification task. To address the TC task, we used RoBERTa based transformer architecture for fine-tuning on the propaganda dataset. The predictions of RoBERTa were further fine-tuned by class-dependent-minority-class classifiers. A special classifier, which employs dynamically adapted Least Common Sub-sequence algorithm, is used to adapt to the intricacies of repetition class. Compared to the other participating systems, our submission is ranked 4th on the leaderboard.

cs.CL↗

arXiv2023

An Effective Transformer-based Contextual Model and Temporal Gate Pooling for Speaker Identification

Wav2vec2 has achieved success in applying Transformer architecture and self-supervised learning to speech recognition. Recently, these have come to be used not only for speech recognition but also for the entire speech processing. This paper introduces an effective end-to-end speaker identification model applied Transformer-based contextual model. We explored the relationship between the hyper-parameters and the performance in order to discern the structure of an effective model. Furthermore, we propose a pooling method, Temporal Gate Pooling, with powerful learning ability for speaker identification. We applied Conformer as encoder and BEST-RQ for pre-training and conducted an evaluation utilizing the speaker identification of VoxCeleb1. The proposed method has achieved an accuracy of 87.1% with 28.5M parameters, demonstrating comparable precision to wav2vec2 with 317.7M parameters. Code is available at https://github.com/HarunoriKawano/speaker-identification-with-tgp.

cs.SD↗

arXiv2024

Transformers Can Represent $n$-gram Language Models

Existing work has analyzed the representational capacity of the transformer architecture by means of formal models of computation. However, the focus so far has been on analyzing the architecture in terms of language \emph{acceptance}. We contend that this is an ill-suited problem in the study of \emph{language models} (LMs), which are definitionally \emph{probability distributions} over strings. In this paper, we focus on the relationship between transformer LMs and $n$-gram LMs, a simple and historically relevant class of language models. We show that transformer LMs using the hard or sparse attention mechanisms can exactly represent any $n$-gram LM, giving us a concrete lower bound on their probabilistic representational capacity. This provides a first step towards understanding the mechanisms that transformer LMs can use to represent probability distributions over strings.

cs.CL↗

arXiv2025

Revisiting LRP: Positional Attribution as the Missing Ingredient for Transformer Explainability

The development of effective explainability tools for Transformers is a crucial pursuit in deep learning research. One of the most promising approaches in this domain is Layer-wise Relevance Propagation (LRP), which propagates relevance scores backward through the network to the input space by redistributing activation values based on predefined rules. However, existing LRP-based methods for Transformer explainability entirely overlook a critical component of the Transformer architecture: its positional encoding (PE), resulting in violation of the conservation property, and the loss of an important and unique type of relevance, which is also associated with structural and positional features. To address this limitation, we reformulate the input space for Transformer explainability as a set of position-token pairs. This allows us to propose specialized theoretically-grounded LRP rules designed to propagate attributions across various positional encoding methods, including Rotary, Learnable, and Absolute PE. Extensive experiments with both fine-tuned classifiers and zero-shot foundation models, such as LLaMA 3, demonstrate that our method significantly outperforms the state-of-the-art in both vision and NLP explainability tasks. Our code is publicly available.

cs.LG↗

arXiv2025

ReSSFormer: A Recursive Sparse Structured Transformer for Scalable and Long-Context Reasoning

While Transformer architectures have demonstrated impressive scalability across domains, they continue to face challenges in long-context reasoning, computational efficiency, and structural generalization - largely due to rigid layer stacking, dense attention, and reliance on positional encodings. We present ReSSFormer, a Recursive Sparse Structured Transformer that integrates three complementary innovations: Recurrent Reasoning & Memory Unit (R2MU) for iterative reasoning with bounded depth, Adaptive Sparse Attention Module (ASAM) for efficient and focused context selection, and Self-Organizing Encoder Structure (SOES) for position-free structure induction. ReSSFormer replaces conventional depth stacking with recurrent inference, substitutes full attention with token- and expert-level sparsity, and models latent token topology directly from content. Across language modeling, multi-hop QA, and structure-sensitive tasks, ReSSFormer consistently outperforms strong baselines under comparable FLOPs and parameter budgets, highlighting its scalability, efficiency, and structural flexibility.

cs.CL↗

arXiv2025

Towards Unsupervised Causal Representation Learning via Latent Additive Noise Model Causal Autoencoders

Unsupervised representation learning seeks to recover latent generative factors, yet standard methods relying on statistical independence often fail to capture causal dependencies. A central challenge is identifiability: as established in disentangled representation learning and nonlinear ICA literature, disentangling causal variables from observational data is impossible without supervision, auxiliary signals, or strong inductive biases. In this work, we propose the Latent Additive Noise Model Causal Autoencoder (LANCA) to operationalize the Additive Noise Model (ANM) as a strong inductive bias for unsupervised discovery. Theoretically, we prove that while the ANM constraint does not guarantee unique identifiability in the general mixing case, it resolves component-wise indeterminacy by restricting the admissible transformations from arbitrary diffeomorphisms to the affine class. Methodologically, arguing that the stochastic encoding inherent to VAEs obscures the structural residuals required for latent causal discovery, LANCA employs a deterministic Wasserstein Auto-Encoder (WAE) coupled with a differentiable ANM Layer. This architecture transforms residual independence from a passive assumption into an explicit optimization objective. Empirically, LANCA outperforms state-of-the-art baselines on synthetic physics benchmarks (Pendulum, Flow), and on photorealistic environments (CANDLE), where it demonstrates superior robustness to spurious correlations arising from complex background scenes.

cs.LG↗

arXiv2026

TMallGS: Scaling Unified Feature and Sequence Modeling for Generative E-commerce Search

In industrial search and ranking systems, Click-Through Rate (CTR) prediction is shifting from traditional Deep Learning Recommendation Models (DLRM) toward unified, compute-intensive Transformer architectures. This transition is driven by the need to improve Model FLOPs Utilization (MFU) and achieve predictable gains through scaling laws. However, existing approaches such as OneTrans and Climber often adopt an all-in-tokenization strategy when adapting Large Language Model (LLM) architectures, overlooking the heterogeneous nature of ranking features. We propose TmallGS, a scalable ranking architecture for Tmall search. TmallGS includes five key components: (1) Hierarchical Distribution-Calibrated Tokenization, which combines Field-wise Saliency Reweighting (FSR) and Distribution-Calibrated Projection (DCP) to map diverse features into optimized subspaces; (2) a Field-Adaptive Gated Transformer Backbone with per-field QKV projections and noise-adaptive gating for refined semantic interaction; (3) Decoupled FiLM Late Fusion to preserve explicit high-frequency signals; (4) a Context-Aware Bias Net to decouple systemic bias from user intent; and (5) Error-Aware Progressive Training with dynamically weighted losses for robust learning. Extensive offline experiments and online A/B tests on Tmall Search show that TmallGS improves training throughput and achieves substantial gains in UCTCVR and GMV.

cs.IR↗

arXiv2026

Tapered Language Models

Modern language models, including transformer, recurrent, and memory-based variants, share a common chassis: a stack of identical layers in which parameters are allocated uniformly across depth. This is a default inherited from the original transformer and largely unchanged since, yet a growing body of evidence suggests that layers contribute non-uniformly to the final output, with later layers refining the residual stream rather than transforming it. We ask whether parameter capacity should reflect this asymmetry. Our controlled experiment shows that, under a fixed budget, allocating more capacity to earlier layers and less to later layers improves perplexity over a uniform-width baseline, while the reverse allocation hurts. Building on this result, we introduce Tapered Language Models (TLMs), an architectural principle in which a parameter-bearing component is monotonically tapered across depth under a fixed total budget. MLPs are the natural site for this instantiation: they dominate parameter count across all modern LM families and expose width as a single, clean axis of variation. Across three model scales and four architectures (Transformer, Gated Attention, Hope-attention, and Titans), tapering MLP width via a smooth cosine schedule consistently improves perplexity and downstream benchmark performance over uniform baselines, at no additional parameter or compute cost. These findings establish depth-aware capacity allocation as a simple, architecture-agnostic axis of language model design, a free lever hidden in plain sight.

cs.LG↗

arXiv2026

SparseWorld-TC: Trajectory-Conditioned Sparse Occupancy World Model

This paper introduces a novel architecture for trajectory-conditioned forecasting of future 3D scene occupancy. In contrast to methods that rely on variational autoencoders (VAEs) to generate discrete occupancy tokens, which inherently limit representational capacity, our approach predicts multi-frame future occupancy in an end-to-end manner directly from raw image features. Inspired by the success of attention-based transformer architectures in foundational vision and language models such as GPT and VGGT, we employ a sparse occupancy representation that bypasses the intermediate bird's eye view (BEV) projection and its explicit geometric priors. This design allows the transformer to capture spatiotemporal dependencies more effectively. By avoiding both the finite-capacity constraint of discrete tokenization and the structural limitations of BEV representations, our method achieves state-of-the-art performance on the nuScenes benchmark for 1-3 second occupancy forecasting, outperforming existing approaches by a significant margin. Furthermore, it demonstrates robust scene dynamics understanding, consistently delivering high accuracy under arbitrary future trajectory conditioning.

cs.CV↗

arXiv2026

Sortblock: Similarity-Aware Feature Reuse for Diffusion Model

Diffusion Transformers (DiTs) have demonstrated remarkable generative capabilities, particularly benefiting from Transformer architectures that enhance visual and artistic fidelity. However, their inherently sequential denoising process results in high inference latency, limiting their deployment in real-time scenarios. Existing training-free acceleration approaches typically reuse intermediate features at fixed timesteps or layers, overlooking the evolving semantic focus across denoising stages and Transformer blocks.To address this, we propose Sortblock, a training-free inference acceleration framework that dynamically caches block-wise features based on their similarity across adjacent timesteps. By ranking the evolution of residuals, Sortblock adaptively determines a recomputation ratio, selectively skipping redundant computations while preserving generation quality. Furthermore, we incorporate a lightweight linear prediction mechanism to reduce accumulated errors in skipped blocks.Extensive experiments across various tasks and DiT architectures demonstrate that Sortblock achieves over 2$\times$ inference speedup with minimal degradation in output quality, offering an effective and generalizable solution for accelerating diffusion-based generative models.

cs.CV↗

arXiv2026

Comparative Assessment of Deep Learning Architectures for Underwater Subsurface Kelp Forest Segmentation with The Kelp-o-Tron

Submerged kelp forests are vital coastal ecosystems that support marine biodiversity and ecosystem dynamics, yet accurate underwater kelp segmentation remains challenging due to optical degradation, illumination variability, turbidity, overlapping vegetation, and complex benthic backgrounds. We systematically evaluated three deep learning semantic segmentation frameworks, ResNet34-U-Net, ResNet50-DeepLabV3, and a hybrid ResNet50-ASPP-Transformer architecture, for kelp detection using high-resolution underwater RGB imagery collected from northeastern U.S. coastal waters. A dataset of 3,395 SSeg assisted annotated image-mask pairs was developed for model training and validation, while geographically independent sites were used for quantitative and qualitative evaluation. All models used consistent preprocessing, augmentation, and evaluation protocols. On independent test data, ResNet50-DeepLabV3 achieved the highest Dice (0.7120) and Intersection over Union (IoU; 0.6267), followed by ResNet34 U Net (Dice 0.6868; IoU 0.5978). The hybrid ASPP Transformer achieved the highest pixel accuracy (0.8528) but lower Dice (0.6437) and IoU (0.5746). External qualitative evaluation further showed that DeepLabV3 produced more consistent segmentation across varying environmental conditions, image qualities, and benthic habitats. Overall, ResNet50-DeepLabV3, termed Kelp-O-Tron, provided the best balance of segmentation accuracy, robustness, and generalization. The dataset, annotation workflow, and comparative evaluation provide resources for advancing automated underwater habitat mapping and ecological monitoring.

cs.CV↗