Transformers

Federated learning with heterogeneous clients remains a significant challenge for deep learning, primarily due to client drift arising from inconsistent local updates. Existing federated optimization methods typically address this issue through objective-level regularization or update-correction mechanisms. Recent studies, however, suggest that Transformer-based architectures may be inherently more robust than conventional models under heterogeneous federated training. Motivated by this observation, we investigate how different parameter components within the attention mechanism influence federated optimization. Specifically, we decompose the attention module into a query/key block, which determines the attention kernel, and a value block, which performs semantic transformation under the induced kernel. Based on this perspective, we propose FedFrozen, a two-stage federated optimization framework that first performs full-model warm-up training and then freezes the query/key block while continuing to optimize the value block. Under a linear-attention formulation, we show that the warm-up stage can be interpreted as an inexact descent procedure on a regularized kernel-profile objective, while the frozen stage reduces to a restricted value-block optimization problem under a fixed attention kernel. Our analysis further reveals an explicit trade-off that governs the choice of warm-up length. Simulations validate the predicted bias-drift behavior, and real-data experiments demonstrate that FedFrozen improves both the stability and effectiveness of Transformer models in heterogeneous federated learning.

As large language models (LLMs) continue to advance rapidly, they are becoming increasingly capable while simultaneously demanding ever-longer context lengths. To improve the inference efficiency of long-context processing, several novel low-complexity hybrid architectures have recently been proposed, effectively alleviating the computational burden of long-context inference. However, existing research on long-context prefill acceleration remains predominantly focused on sparse attention mechanisms, which achieve their maximum speedup only on full-attention models. When transferred to emerging architectures--such as linear/full attention hybrids or sliding window/full attention hybrids--these prefill acceleration approaches suffer significant performance degradation. Furthermore, such methods are generally incompatible with continuous batching, making them difficult to integrate into modern inference engines such as vLLM. To this end, we propose UniPrefill, a prefill acceleration framework applicable to virtually any model architecture, which directly accelerates the model's computation at the token level. We further implement UniPrefill as a continuous batching operator and extend vLLM's scheduling strategy to natively support prefill-decode co-processing and tensor parallel for UniPrefill, enabling its seamless integration into vLLM. UniPrefill achieves up to 2.1x speedup in Time-To-First-Token (TTFT), with the acceleration becoming increasingly pronounced as the number of concurrent requests grows.

Mixture of experts has emerged as the primary mechanism for making Large Language Models (LLMs) computationally efficient. However, in distributed settings, communicating token embeddings between experts is a significant bottleneck. We present the novel Federation of Experts (FoE) architecture. FoE restructures the MoE block of a transformer layer into multiple MoE clusters. Each cluster is responsible for only one of the KV heads and expert parallelism is applied between those experts. Between clusters, a sum synchronizes the post-attention residuals, which then drives routing and dispatch for the next MoE block. In a single-node setting, FoE completely eliminates all-to-all communication as all experts within a group are contained on the same GPU. In multi-node settings, FoE confines all-to-all communication to the intra-node fabric, thus significantly reducing communication overhead. An implementation of FoE finds that on LongBench, FoE significantly improves inference throughput and latency in both single-node and multi-node settings, reducing the end-to-end forward-pass latency by up to 5.2x, TTFT by 3.62x, and TBT by 1.95x. It does so while achieving comparable generation quality to a mixture of experts model of the same size and training configuration.

Despite significant progress, RGB-based trackers remain vulnerable to challenging imaging conditions, such as low illumination and fast motion. Event cameras offer a promising alternative by asynchronously capturing pixel-wise brightness changes, providing high dynamic range and high temporal resolution. However, existing event-based trackers often neglect the intrinsic spatial sparsity and temporal density of event data, while relying on a single fixed temporal-window sampling strategy that is suboptimal under varying motion dynamics. In this paper, we propose an event sparsity-aware tracking framework that explicitly models event-density variations across multiple temporal scales. Specifically, the proposed framework progressively injects sparse, medium-density, and dense event search regions into a three-stage Vision Transformer backbone, enabling hierarchical multi-density feature learning. Furthermore, we introduce a sparsity-aware Mixture-of-Experts module to encourage expert specialization under different sparsity patterns, and design a dynamic pondering strategy to adaptively adjust the inference depth according to tracking difficulty. Extensive experiments on FE240hz, COESOT, and EventVOT demonstrate that the proposed approach achieves a favorable trade-off between tracking accuracy and computational efficiency. The source code will be released on https://github.com/Event-AHU/OpenEvTracking.

Decision Transformer (DT) formulates offline reinforcement learning as autoregressive sequence modeling, achieving promising results by predicting actions from a sequence of Return-to-Go (RTG), state, and action tokens. However, RTG is a scalar that summarizes future rewards, containing far less information than typical state or action vectors, yet it consumes the same computational budget per token. Worse, the self-attention cost of Transformers grows quadratically with sequence length, so including RTG as a separate token adds unnecessary overhead. We propose SlimDT, which removes RTG from the autoregressive sequence. Instead, we inject RTG information into the state representations before the sequential modeling step, allowing the Transformer to process only a compact (state, action) sequence. This reduces the sequence length by one-third, directly improving inference efficiency. On the D4RL benchmark, SlimDT surpasses standard DT across various tasks and achieves performance comparable to existing state-of-the-art methods. Decoupling a sparse conditioning signal from an information-rich sequence thus yields both computational gains and higher task performance.

Part-based reasoning is a classical strategy to make a computer vision model directly focus on the object parts that are relevant to the downstream task. In the context of deep learning, this also serves to improve by-design interpretability, often by using part-centric attention mechanisms on top of a latent image representation provided by a standard, black-box model. This approach is based on a locality assumption: that the latent representation of an object part encodes primarily information about the corresponding image region. In this work, we test this basic assumption, measuring intra-object leakage in vision models using part-based attribute annotations. Through a comprehensive experimental evaluation, we show that modern pretrained vision transformers violate the locality assumption and exhibit a strong intra-object leakage, in which each part encodes information from the whole object, a visual metonymy that compromises the faithfulness of attention-based interpretable-by-design methods for part-based reasoning, ultimately rendering them uninterpretable. In addition, we establish an upper bound using a two-stage approach that prevents leakage by design. We then show that this inherently disentangled feature extraction improves attribute-driven part discovery on a variety of tasks, confirming the practical impact of intra-object leakage. Our results uncover a neglected issue affecting the interpretability of part-based representations, such as those in CBMs relying on part-centric concepts, highlighting that two-stage approaches offer a promising way to mitigate it.

We present the E$Δ$-MHC-Geo Transformer, a novel architecture that unifies Manifold-Constrained Hyper-Connections (mHC), Deep Delta Learning (DDL), and the Cayley transform to obtain input-adaptive, unconditionally orthogonal residual connections. Unlike DDL, whose Householder operator is orthogonal only at $β\in \{0,2\}$, our Data-Dependent Cayley rotation $Q(x)=(I+(β/2)A(x))^{-1}(I-(β/2)A(x))$ preserves orthogonality for all $β$ and all inputs. To handle negation, an eigenvalue $-1$ case that Cayley provably excludes, we introduce the E$Δ$-MHC-Geo Hybrid, which combines Cayley rotation with Householder reflection via a learned operator-selection gate $X'=γ(X)Q(X)X+(1-γ(X))H_2(X)X$. A midpoint-collapse regularizer, $4γ(1-γ)$, encourages boundary gate decisions, where each selected component is orthogonal. In matched-parameter comparisons, with approximately 1.79M parameters per model and mean +/- standard deviation over 3 seeds, against four baselines including the concurrent JPmHC, E$Δ$-MHC-Geo achieves the best long-horizon stability, 1.9x over JPmHC and 3.8x over GPT; the best near-$π$ rotation loss, 4.5x over JPmHC on single-plane; strong norm preservation, with 0.001 mean deviation; and 0.96 negation cosine alignment in a diagnostic reflection probe, all with 33% fewer layers. While JPmHC's wider representation excels on pure rotation, its finite Cayley residual mixer excludes an exact $λ=-1$ operator and has no reflection branch, motivating our hybrid approach for accessing both connected components of $O(n)$.

Linear Attention (LA) offers a promising paradigm for scaling large language models (LLMs) to long sequences by avoiding the quadratic complexity of self-attention. Recent LA models such as Mamba2 and GDN interpret linear recurrences as closed-form online stochastic gradient descent (SGD), but naive SGD updates suffer from rapid information decay and suboptimal convergence in optimization. While momentum-based optimizers provide a natural remedy, they pose challenges in simultaneously achieving training efficiency and effectiveness. To address this, we develop a chunkwise parallel algorithm for LA with a stepwise momentum rule by geometrically reordering the update coefficients. Further, from a dynamical systems perspective, we analyze the momentum-based recurrence as a second-order system that introduces complex conjugate eigenvalues. This analysis guides the design of stable gating constraints. The resulting model, Momentum DeltaNet (MDN), leverages Triton kernels to achieve comparable training throughput with competitive linear models such as Mamba2 and KDA. Extensive experiments on the 400M and 1.3B parameter models demonstrate consistent performance improvements over strong baselines, including Transformers, Mamba2 and GDN, across diverse downstream evaluation benchmarks. Code: https://github.com/HuuYuLong/MomentumDeltaNet .

We investigate the ability of transformers to perform in-context reinforcement learning (ICRL), where a model must infer and execute learning algorithms from trajectory data without parameter updates. We show that a linear self-attention transformer block can provably implement policy-improvement methods, including semi-gradient SARSA and actor-critic, via explicit parameter constructions. Beyond existence, we design a teacher-mimicking training procedure, analyze its gradient-flow dynamics, and establish the first convergence guarantee in the ICRL literature: under suitable richness conditions on the training MDP distribution, gradient flow converges locally and exponentially to an optimal parameter manifold corresponding to the desired RL update. Empirically, training transformers on randomly generated tabular MDPs confirms these predictions: the learned models recover the parameter structure of our explicit constructions and, when deployed on unseen MDPs, deliver strong in-context control performance. Together, these results illuminate how transformer architectures internalize and execute classical reinforcement learning algorithms in context, bridging mechanistic understanding and training dynamics in ICRL.

While Large Language Models (LLMs) achieve strong performance across diverse tasks, their inference dynamics remain poorly understood because of the limited resolution of existing analysis tools. In this work, we identify an intrinsic magnification mechanism in transformer architectures: deeper layers inherently magnify the small changes of layer-wise confidence, providing a fine-grained confidence trajectory. Building on this insight, we introduce HyperLens, a high-resolution probe designed to trace confidence trajectories and quantify the cognitive effort during inference. Across LLMs and datasets, HyperLens reveals a consistent divergence in confidence trajectories that separates complex from simple tasks. We abstract this pattern into a quantitative cognitive effort metric. Our analysis reveals a fundamental principle: complex tasks consistently require higher cognitive effort. Finally, we provide a mechanistic diagnosis of a common side effect of standard Supervised Fine-Tuning (SFT): it can reduce cognitive effort and consequently degrade performance on in-domain tasks.

Transformers usually expose one inference cost per trained model, while deployed systems often need multiple cost-quality operating points. We study Budgeted Attention Allocation, a monotone head-gating mechanism conditioned on a requested attention budget. Dense warm-starting is important for stability: on a robust synthetic sequence task, one budgeted model reaches 99.7% accuracy at 0.303 estimated attention cost and 100.0% accuracy at 0.504 cost. On held-out AG News with a custom word-level transformer, hard-gate adaptation turns soft cost control into measured single-thread CPU speed, reaching 82.1% accuracy with 1.28x speedup at budget 0.50. In pretrained BERT-Mini AG News, budgeted structural pruning reaches 87.6% accuracy with 1.20x speedup at budget 0.50; a validation-ranked zero-shot dense post-hoc structural baseline reaches 86.1%, and one recovery epoch raises that per-budget specialist to 87.9%. On DBpedia14, BERT-Mini budgeted gates reach 97.4% at exact budget 0.50 versus 96.6% for dense full attention. Static fixed-budget gates and recovered dense specialists remain strong. The contribution is therefore not universal dominance, but a reproducible feasibility study of one controllable checkpoint across budgets that can trade attention cost for accuracy and be converted into measured structural speedups on small CPU benchmarks.

Despite the rapid evolution of training paradigms, the decoder backbone of large vision--language models (LVLMs) remains fundamentally rooted in the residual-connection Transformer architecture. Therefore, deciphering the distinct roles of internal modules is critical for understanding model mechanics and guiding architectural optimization. While prior statistical approaches have provided valuable attribution-based insights, they often lack a unified theoretical basis. To bridge this gap, we propose a unified framework grounded in information theory and geometry to quantify the geometric and entropic nature of residual updates. Applying this unified framework reveals a fundamental functional decoupling: Attention acts as a subspace-preserving operator focused on reconfiguration, whereas FFNs serve as subspace-expanding operators driving semantic innovation. Strikingly, further experiments demonstrate that replacing learned attention weights with predefined values (e.g., Gaussian noise) yields comparable or even superior performance across a majority of datasets relative to vanilla models. These results expose severe misallocation and redundancy in current mechanisms, suggesting that state-of-the-art LVLMs effectively ``get lost in attention'' rather than efficiently leveraging visual context.

In grokking, a model first fits the training data while test accuracy remains low, and only later begins to generalize. We ask whether this transition can be localized from observed training trajectories before the test accuracy rises, and formulate grokking transition localization as a diagnostic problem with an explicit threshold/FPR/lead-time trade-off. Task-dependent observables are summarized as empirical distributions, mapped to Wasserstein/quantile coordinates, and analyzed by Hankel dynamic mode decomposition (DMD); the resulting reconstruction residual, together with spectrum and effective rank, forms the diagnostic output. On held-out modular-addition Transformer runs, the residual achieves AUROC $\approx$ 0.93 for grokking-vs-non-grokking discrimination at the run level; under a fixed sustained-threshold operating rule, true-positive alarms can precede onset, with lead time reported jointly with false-alarm rate and uncertainty intervals. Perturbation experiments show that, in the tested $wd=1$ pool, high-residual windows exhibit about $3\times$ larger short-horizon perturbation deviation than low-residual windows. In a same-data norm-window control, perturbation sensitivity aligns with the residual ordering rather than total-parameter-norm ordering, suggesting that the residual is not merely a total-norm proxy at the window level in the studied $wd=1$ dynamics. Norm signals remain strong run-level regime indicators, and log-probability performs best among the observables tested under the current protocol. We position the residual as a window-level monitoring and localization signal in the studied modular-arithmetic Transformer settings, not a universal early-warning predictor or an intervention rule.

Identifying and exploiting common features across domains is at the heart of the human ability to make analogies, and is believed to be crucial for the ability to continually learn. To do this successfully, general and flexible computational strategies must be developed. While the extent to which Transformer neural network models can perform compositional reasoning has been the subject of intensive recent investigation, little work has been done to systematically understand how well these models can leverage their representations to learn new, related experiences. To address this gap, we expand the previously developed Learning Equality and Group Operations (LEGO) framework to a continual learning (CL) setting ("continual LEGO"). Using this continual LEGO experimental paradigm, we study the capability of feedforward and recurrent Transformer models to perform CL. We find that BERT, a canonical feedforward Transformer model, learns shortcut solutions that limits its ability to generalize and prevents strong forward transfer to new experiences. In contrast, we find evidence supporting the hypothesis that ALBERT, a recurrent version of BERT, learns a For loop-esque solution, which leads to better CL performance. When applying BERT and ALBERT models to a CL setting that requires composition across experiences, we find that both model families fail. Our investigation suggests that ALBERT models can have their performance drop rescued by use of training strategies that combine data across experiences, but this is not true for BERT models, where a detrimental shortcut solution becomes entrenched with initial training. Our results demonstrate that the recurrent ALBERT model may have an inductive bias better suited for CL and motivate future investigation of the interplay between Transformer architecture and computational solutions that emerge in modern models and tasks.

Standard fine-tuning of transformer models on causal reasoning tasks leads to catastrophic model collapse, where models learn trivial solutions such as always predicting "Yes" or "No" regardless of input structure. We demonstrate that fine-tuning Gemma 270M on transitivity and d-separation tasks without semantic loss results in 100% collapse rate, with models achieving misleadingly high accuracy (73.9%) while learning no causal reasoning. We propose a semantic loss function with graph-based logical constraints and dynamic lambda scheduling that prevents this collapse. Our approach achieves 70.4% accuracy on transitivity tasks and 68.6% on d-separation tasks with stable, context-dependent predictions, representing a 42.7% improvement over collapsed baselines. Adversarial evaluation on 1,000 structural reasoning samples shows semantic models achieve 67-70% accuracy while collapsed models fail catastrophically at 43-71%. We validate our findings through comprehensive benchmarking on 200,000+ evaluation samples across five model variants, demonstrating that semantic loss is essential and not optional, for stable causal reasoning in transformers.

Service level agreements (SLAs) in data center colocation contracts define precise thresholds for power, temperature, and humidity, with tiered violation penalties expressed as credits against monthly recurring charges. Traditional reactive monitoring detects breaches only after they occur, limiting remediation opportunities. We present a framework that encodes SLA rules as structured JSON objects to generate training data without manual annotation. We train a per-customer multi-head transformer model in which each attention head specializes in one SLA rule, learning temporal dependencies that precede violations by 30 minutes. Post-training, the inference service emits structured prediction events transformed into three role-specific views: finance schemas exposing credit liability, operations schemas surfacing risk scores and recommended interventions, and compliance schemas bundling predictions with immutable telemetry signatures for audit. By aligning model architecture directly with contractual obligations, this framework enables operators to anticipate SLA breaches, prioritize corrective actions, and minimize financial penalties.

We study outlier tokens in Diffusion Transformers (DiTs) for image generation. Prior work has shown that Vision Transformers (ViTs) can produce a small number of high-norm tokens that attract disproportionate attention while carrying limited local information, but their role in generative models remains underexplored. We show that this phenomenon appears in both the encoder and denoiser of modern Representation Autoencoder (RAE)-DiT pipelines: pretrained ViT encoders can produce outlier representations, and DiTs themselves can develop internal outlier tokens, especially in intermediate layers. Moreover, simply masking high-norm tokens does not improve performance, indicating that the problem is not only caused by a few extreme values, but is more closely related to corrupted local patch semantics. To address this issue, we introduce Dual-Stage Registers (DSR), a register-based intervention for both components: trained registers when available, recursive test-time registers otherwise, and diffusion registers for the denoiser. Across ImageNet and large-scale text-to-image generation, these interventions consistently reduce outlier artifacts and improve generation quality. Our results highlight outlier-token control as an important ingredient in building stronger DiTs.

Pre-trained transformers are able to learn from examples provided as part of the prompt without any weight updates, a remarkable ability known as in-context learning (ICL). Despite its demonstrated efficacy across various domains, the theoretical understanding of ICL is still developing. Whereas most existing theory has focused on linear models, we study ICL in the nonlinear regression setting. Through the interaction mechanism in attention, we explicitly construct transformer networks to realize nonlinear features, such as polynomial or spline bases, which span a wide class of functions. Based on this construction, we establish a framework to analyze end-to-end in-context nonlinear regression with the constructed features. Our theory provides finite-sample generalization error bounds in terms of context length and training set size. We numerically validate the theory on synthetic regression tasks.

Transformer architectures have been widely adopted for time series forecasting, yet whether the representational mechanisms that make them powerful in NLP actually engage on time series data remains unexplored. The persistent competitiveness of simple linear models such as DLinear has fueled ongoing debate, but no mechanistic explanation for this phenomenon has been offered. We address this gap by applying sparse autoencoders (SAEs), a tool from mechanistic interpretability, to probe the internal representations of PatchTST. We first establish that a single-layer, narrow-dimensional transformer matches the forecasting performance of deeper configurations across commonly used benchmarks. We then train SAEs on the post-GELU intermediate FFN activations with dictionary sizes ranging from 0.5x to 4.0x the native dimensionality. Expanding the dictionary yields negligible downstream performance change (average 0.214%), with large portions of overcomplete dictionaries remaining inactive. Targeted causal interventions on dominant latent features produce minimal forecast perturbation. Across all evaluated settings, we observe no empirical evidence that the analyzed FFN representations rely on strong superposition. Instead, the representations remain sparse, stable under aggressive dictionary expansion, and largely insensitive to latent interventions. These results demonstrate that superposition is not necessary for competitive performance on standard forecasting benchmarks, suggesting they may not demand the rich compositional representations that drive transformer success in language modeling, and helping explain the persistent competitiveness of simple linear models

Efficient CUDA implementations of attention mechanisms are critical to modern deep learning systems, yet supporting diverse and evolving attention variants remains challenging. Existing frameworks and compilers trade performance for flexibility, while expert-written kernels achieve high efficiency but are difficult to adapt. Recent work explores large language models (LLMs) for GPU kernel generation, but prior studies report unstable correctness and significant performance gaps for complex operators such as attention. We present CuBridge, an LLM-based framework that adapts expert-written attention kernels through a structured lift-transfer-lower workflow. CuBridge starts from expert-written CUDA attention kernels and lifts them into an executable intermediate representation that makes execution orchestration explicit while abstracting low-level CUDA syntax. Given a user-provided PyTorch specification, CuBridge generates and verifies a target IR program, then reconstructs optimized CUDA code via reference-guided lowering. Across diverse attention variants and GPU platforms, CuBridge consistently produces correct kernels and substantially outperforms general frameworks, compiler-based approaches, and prior LLM-based methods.

Neural operators generally demonstrate strong predictive performance on in-distribution (ID) problems. However, a critical limitation of existing methods is their significant performance degradation when encountering out-of-distribution (OOD) data. To address this issue, this work introduces continual learning into physics-informed neural operators, with particular emphasis on neural operators built upon the Transolver architecture, and proposes a simple yet effective replay-based continual learning strategy. The proposed method is fully physics-informed and does not require labeled data, relying solely on input fields together with physical constraints for training. When new OOD data become available, a small number of past data are incorporated through a distillation-based constraint to preserve previously acquired knowledge and alleviate catastrophic forgetting. Meanwhile, a transfer learning LoRA is employed to enable rapid adaptation to the new data. The proposed framework is systematically validated on three representative physical problems, including the Darcy flow problem in fluid mechanics, a two-dimensional hyperelastic brain tumor problem in biomechanics, and a three-dimensional linear elastic Triply Periodic Minimal Surfaces problem in solid mechanics. The results demonstrate that the proposed method effectively mitigates catastrophic forgetting on previously learned data while maintaining fast adaptability to new data. Compared with conventional joint training strategies, the proposed method significantly improves training efficiency while reducing additional memory usage and computational cost.

We analyse the computational power of transformer encoders as sequence-to-sequence functions on vectors. We show that average hard attention can be used to simulate arithmetic circuits if they are given as an input to an encoder. The circuit families that can be simulated this way have constant depth while using unbounded addition, binary multiplication and sign gates. The transformers we use have arithmetic circuits instead of feed-forward networks. With typical average attention the functions they compute are also computed by the same class of circuit families. Our results hold for transformers over the reals, rationals and any ring in between the two.

Channel-gain maps provide the channel gain between any two locations in a geographical region. They find numerous applications, from resource allocation and interference control to path planning for autonomous vehicles. Channel-gain map estimation (CGME) is considerably more challenging than conventional radio map estimation (RME) because channel-gain maps are functions over a 6-dimensional input space. This calls for specialized methods, which currently rely on the (inaccurate) radio tomographic model or require a prohibitively large number of measurements since they do not exploit any spatial structure. This paper overcomes this issue by leveraging spatial patterns that channel-gain maps exhibit across environments, as dictated by the laws of physics and typical environmental characteristics (e.g. building materials and layouts). Adopting a metalearning perspective, a transformer-based estimator is proposed to implicitly learn this common structure from measurements collected in multiple environments. This enables CGME in new environments from significantly fewer measurements (five times less in our experiments). To maximize learning efficiency, the transformer is composed with a feature map that enforces the invariances of CGME, such as those following from reciprocity. Numerical experiments corroborate the merits of the proposed estimator relative to existing methods.

Video editing has evolved toward In-Context Learning (ICL) paradigms, yet the resulting quadratic attention costs create a critical computational bottleneck. In this work, we propose In-context Sparse Attention (ISA), the first near-lossless empirical sparse framework tailored for ICL video editing. Our design is grounded in two key insights: first, context tokens exhibit significantly lower saliency than source tokens; second, we theoretically prove and empirically validate that Query sharpness correlates with approximation error. Motivated by these findings, ISA implements an efficient pre-selection strategy to prune redundant context, followed by a dynamic query grouping mechanism that routes high-error queries to full attention and low-error ones to a computationally efficient 0-th order Taylor sparse attention. Furthermore, we build \textbf{\texttt{LIVEditor-14B}} , a novel lightning video editing model via ISA and a proposed video-editing data pipeline that curated a 1.7M high-quality dataset. Extensive experiments demonstrate that LIVEditor-14B achieves a $\sim$60% reduction in attention-module latency while surpassing state-of-the-art methods across EditVerseBench, IVE-Bench, and VIE-Bench, delivering near-lossless acceleration without compromising visual fidelity.

Continuous-time (CT) Transformers improve irregular and long-range modeling over CT-RNNs by exploiting inputs or outputs embeddings with continuous dynamics. However, the core scaled-dot-product-attention (SDPA) mechanism remains inherently discrete. We propose FLUID (Flexible Unified Information Dynamics), a CT Transformer that incorporates continuous dynamics directly into the attention computation by replacing it with Liquid Attention Network (LAN). LAN reinterprets attention logits as continuous dynamical system and reformulates them as the solution to a linear ODE modulated by input-dependent nonlinear recurrent gates. Theoretically, we establish stability guarantees for LAN dynamics and show that it serves as an interpolating middle ground between SDPA and CT-RNNs, recovering each as special case under well-defined parameterization of its gating functions. LAN also introduces an explicit attention-sink gate to eliminate disproportionate attention mass on uninformative nodes. FLUID replaces standard residual connections with input-dependent Liquid Hyper-Connections to adaptively regulate interlayer information flow. Empirically, we evaluate FLUID on a broad set of learning tasks, including (i) irregular time-series, (ii) long-range modeling, (iii) lane-keeping control of autonomous vehicles, and (iv) learning physical dynamics under a scarce data regime. Across all the tasks, FLUID consistently matches or outperforms CT baselines, achieving improvements of up to 47% in certain scenarios and enhancing generalization under distributional shifts. Additionally, FLUID demonstrates superior noise robustness and a self-correcting inductive bias in autonomous vehicle control. We also provide a detailed analysis of key hyperparameters to guide tuning and show that FLUID occupies an intermediate position among competing approaches in terms of runtime and memory efficiency.

Recent work has shown that Transformers' compositional generalization is governed by \emph{complexity control}, initialization scale and weight decay, which steers training toward low-complexity reasoning solutions rather than high-complexity memorization. Existing analyses, however, treat complexity control as a single static hyperparameter choice, leaving open \emph{when} during training this control is actually decisive. We show that the memorization-versus-reasoning fate of a Transformer is determined within a sharp, identifiable window of training. On a controlled compositional task we find that (i)~weight decay applied for a single 25%-of-training window matches full-training weight decay in out-of-distribution (OOD) accuracy ($0.93$ vs $0.91$); (ii)~holding total regularization budget constant, placing it in the middle of training yields $5{-}9\times$ higher OOD accuracy than placing it early; (iii)~the boundary of the critical window is remarkably sharp, window onset shifted by as little as $100$ optimization steps causes mean OOD to jump from chance ($0.15$) to reasoning-regime ($0.61$); (iv)~the window's position depends systematically on initialization scale, but the basin of attraction for reasoning solutions \emph{shrinks} at small initialization, contradicting the prevailing recommendation that smaller initialization is uniformly better. We further show that the critical-window phenomenon is task-specific: it does not appear on grokking with modular arithmetic, where properly tuned constant weight decay matches scheduled weight decay.

We investigate the scaling properties of implicit deductive reasoning over Horn clauses in depth-bounded Transformers. By systematically decorrelating provability from spurious features and enforcing algorithmic alignment, we find that in sufficiently deep models with a bidirectional prefix mask, implicit reasoning approaches explicit CoT performance across graph topologies and problem widths, though CoT remains necessary for depth extrapolation.

Transformer self-attention can be interpreted as a gradient flow on the unit sphere, in which tokens evolve under softmax interaction potentials and tend to form clusters. While prior work has established clustering behavior for single-head attention, the multi-head setting remains less understood due to geometric interference between heads, which invalidates standard monotonicity arguments. In this work, we develop a theoretical framework for multi-head self-attention dynamics and resolve several open questions. We show that, under suitable conditions on the score matrices, a natural multi-head energy functional is non-decreasing along both flat and spherical dynamics. We identify the key obstruction to per-head monotonicity as radial shadow terms, which are projections of each head's output onto token directions, persisting even under orthogonality assumptions. We introduce a sufficient condition ensuring monotonicity and establish robustness to approximate orthogonality. In a simplified scalar-head regime with equiangular token configurations, we derive a closed-form expression for the critical inverse temperature governing clustering behavior, and show that heterogeneous heads exhibit super-additive clustering rates. In this regime, we also prove a separation in clustering time between ReLU and softmax attention in the linearized dynamics. Finally, we establish an entropy production identity and show that attention entropy increases monotonically toward equilibrium as clustering progresses. Our results provide a unified perspective on the dynamics of multi-head attention and clarify the mechanisms underlying clustering and stability in transformer models.

Visual correspondence across image-to-image (2D-2D), image-to-point cloud (2D-3D), and point cloud-to-point cloud (3D-3D) geometric matching forms the foundation for numerous 3D vision tasks. Despite sharing a similar problem structure, current methods use task-specific designs with separate models for each modality combination. We present UniCorrn, the first correspondence model with shared weights that unifies geometric matching across all three tasks. Our key insight is that Transformer attention naturally captures cross-modal feature similarity. We propose a dual-stream decoder that maintains separate appearance and positional feature streams. This design enables end-to-end learning through stack-able layers while supporting flexible query-based correspondence estimation across heterogeneous modalities. Our architecture employs modality-specific backbones followed by shared encoder and decoder components, trained jointly on diverse data combining pseudo point clouds from depth maps with real 3D correspondence annotations. UniCorrn achieves competitive performance on 2D-2D matching and surpasses prior state-of-the-art by 8% on 7Scenes (2D-3D) and 10% on 3DLoMatch (3D-3D) in registration recall. Project website: https://neu-vi.github.io/UniCorrn

Normalization Equivariance (NE) is a structural prior that improves robustness to distribution shift in image-to-image tasks. A function $f$ is normalization equivariant iff $f(a y + b\mathbf{1}) = a f(y) + b\mathbf{1}$ for all $a>0$ and $b\in\mathbb{R}$. Existing NE methods constrain every internal layer to NE-compatible operations. These constraints add runtime cost and exclude standard transformer components such as softmax attention and LayerNorm. We introduce Wrapped Normalization Equivariance (WNE), a parameter-free wrapper that normalizes the input, applies any backbone, and denormalizes the output. We prove every NE function admits this factorization, so the wrapper exactly parameterizes the class of NE functions. On blind denoising, wrapping CNN and transformer architectures improves robustness under noise-level mismatch with no measurable GPU overhead, while architectural NE baselines are up to $1.6\times$ slower.

SwiGLU is the dominant MLP activation in modern large language models, yet its intermediate tensor materialization costs 9-37% of MLP execution time. We present two complementary CUTLASS-based SM90 kernels that fuse SwiGLU into GeMM at the tile level. Kernel-1 overlaps Swish computation on the Gate accumulator with Up-tile loading using the Pingpong warp-specialized schedule; Kernel-2 interleaves SwiGLU with tile stores via a custom Epilogue Visitor Tree. Evaluated on Qwen-2.5 models (0.5B-72B) on NVIDIA H100, our kernels achieve up to 2.47x speedup over PyTorch, shifting workloads from memory-bound to compute-bound and reaching 79.5% peak BF16 utilization. We demonstrate that torch.compile cannot replicate this fusion (3-7x slower than our kernels), validating the need for hand-crafted tile-level design. Our fused kernels are also numerically superior, achieving zero mismatches compared to 4.5-11% for cuBLAS.

AI-generated text is nowadays produced at scale across domains and heterogeneous generation pipelines, making robustness to distribution shift a central requirement for supervised binary detectors. We train transformer-based detectors on HC3 PLUS and calibrate a single decision threshold by maximising balanced accuracy on held-out validation; this threshold is then kept fixed for all downstream test distributions, revealing domain- and generator-dependent error asymmetries under shift. We evaluate in-domain on HC3 PLUS, under cross-dataset transfer to the multi-domain, multi-generator M4 benchmark, and on the external AI-Text-Detection-Pile. Although base models achieve near-ceiling in-domain performance (up to 99.5% balanced accuracy), performance under shift is brittle and strongly model-dependent. Feature augmentation via attention-based linguistic feature fusion improves transfer, with our best model (DeBERTa-v3-base+FeatAttn) achieving 85.9% balanced accuracy on M4. Multi-seed experiments confirm high stability. Under the same fixed-threshold protocol, our model outperforms strong zero-shot baselines by up to +7.22 points. Category-level ablations further show that readability and vocabulary features contribute most to robustness under shift. Overall, these results demonstrate that feature augmentation and a modern DeBERTa backbone significantly outperform earlier BERT/RoBERTa models, while the fixed-threshold protocol provides a more realistic and informative assessment of practical detector robustness.

Several recent Transformer architectures expose later layers to representations computed in the earliest layers, motivated by the observation that low-level features can become harder to recover as the residual stream is repeatedly transformed through depth. The cheapest among these methods add static value residuals: learned mixing coefficients that expose the first-layer value projection V_1 uniformly across tokens and heads. More expressive dense or dynamic alternatives recover finer-grained access, but at higher memory cost and lower throughput. The usefulness of V_1 is unlikely to be constant across tokens, heads, and contexts; different positions plausibly require different amounts of access to early lexical or semantic information. We therefore treat early-representation reuse as a retrieval problem rather than a connectivity problem, and introduce Selective Access Transformer (SATFormer), which preserves the first-layer value pathway while controlling access with a context-dependent gate. Across models from 130M to 1.3B parameters, SATFormer consistently improves validation loss and zero-shot accuracy over the static value-residual and Transformer baselines. Its strongest gains appear on retrieval-intensive benchmarks, where it improves over static value residuals by approximately 1.5 average points, while maintaining throughput and memory usage close to the baseline Transformer. Gate analyses suggest sparse, depth-dependent, head-specific, and category-sensitive access patterns, supporting the interpretation that SATFormer learns selective reuse of early representations rather than uniform residual copying. Our code is available at https://github.com/SkyeGunasekaran/SATFormer.

Human attention is the gateway to conscious perception, memory and decision-making. However, its role in modern transformer models remains largely unexplored. As these systems increasingly influence what people see, prefer and buy, the question arises as to whether they encode principles of human interest or merely exploit large-scale correlations. Addressing this issue is crucial for understanding cognition and ensuring the responsible use of AI in communication and marketing. In order to address this issue, the concept of visual interest was examined within the multimodal vision-language-model Qwen3-VL-8B, using a pre-defined Common Interestingness (CI) score derived from large-scale human engagement data on the photo-sharing platform Flickr. Here, we analyzed internal representations across vision and language components using methods from the neurosciences. Our analyses revealed that CI information is linearly decodable from final-layer embeddings, indicating that it is aligned with human-derived measures of visual interestingness. Dimensionality reduction and Generalized Discrimination Value (GDV) analyses demonstrate that CI-related hidden representations emerge in intermediate vision transformer layers and becomes progressively more distinguishable across language model layers. Concept vectors derived using geometric, probe, and Sparse Auto-Encoder based methods converge in higher layers, as confirmed by representational similarity analysis. This indicates a robust and structured encoding of visual interestingness without explicit supervision. Future work will seek to identify shared computational principles linking human brain dynamics and transformer architectures, with the ultimate goal of uncovering the organizing mechanisms that give rise to attention and interest in both biological and artificial systems.

Generative models trained on synthetic plan data are a promising approach to generalized planning. Recent work has focused on finding any valid plan, rather than a high-quality solution. We address the challenge of producing high-quality plans, a computationally hard problem, in sub-exponential time. First, we demonstrate that, given optimal data, a decoder-only transformer can generate high-quality plans for unseen problem instances. Second, we show how to self-improve an initial model trained on sub-optimal data. Each round of self-improvement combines multiple model calls with graph search to generate improved plans, used for model fine-tuning. An experimental study on four domains: Blocksworld, Logistics, Labyrinth, and Sokoban, shows on average a 30% reduction in plan length over the source symbolic planner, with over 80% of plans being optimal, where the optimum is known. Plan quality is further improved by inference-time search. The model's latency scales sub-exponentially in contrast to the satisficing and optimal symbolic planners to which we compare. Together, these results suggest that self-improvement with generative models offers a scalable approach for high-quality plan generation.

Real-world clinical data is inherently multimodal, providing complementary evidence that mirrors the practical necessity of jointly assessing multiple related outcomes. Although multi-task learning can improve efficiency by sharing information across outcomes, existing approaches often fail to balance shared representation learning with outcome-specific modeling. Hard parameter sharing can trigger negative transfer when task gradients conflict, while flexible sharing may still entangle shared and task-specific signals. To address this, we propose a multi-task framework built on a unified Transformer for multimodal fusion, augmented with Orthogonal Task Decomposition (OrthTD) to split patient representations into shared and task-specific subspaces and impose a geometric orthogonality constraint to reduce redundancy and isolate task-specific signals. We evaluated OrthTD on a real-world cohort of 12,430 surgical patients for predicting four outcomes. OrthTD achieved average AUC (area under the receiver operating characteristic curve) of 87.5% and average AUPRC (area under the precision-recall curve) of 37.2%, consistently outperformed advanced tabular and multi-task methods. Notably, OrthTD achieves substantial gains in AUPRC, indicating superior performance in identifying rare events within imbalanced clinical data. These results suggest that enforcing non-redundant shared and task-specific representations can improve multi-outcome prediction from multimodal clinical data.

The exponential growth of e-commerce platforms in Indonesia has generated a massive volume of user-generated product reviews. Analyzing the sentiment of these reviews is critical for measuring customer satisfaction and identifying product issues at scale. This paper benchmarks traditional Machine Learning (ML) approaches against a Transformer-based Deep Learning model for a three-class sentiment analysis task (positive, neutral, negative) on the Tokopedia Product Reviews 2025 dataset. We implemented Term Frequency-Inverse Document Frequency (TF-IDF) feature extraction coupled with three algorithms: Logistic Regression, Linear Support Vector Machine (SVM), and Multinomial Naive Bayes as robust baselines. Subsequently, we fine-tuned the IndoBERT model (indobenchmark/indobert-base-p1) for contextual sequence classification. To computationally address the severe class imbalance inherent in e-commerce feedback, we applied balanced class weights for the baseline models and engineered a custom weighted cross-entropy loss function within the IndoBERT training loop, following the broader motivation of imbalanced-learning research. Our comprehensive evaluation using Accuracy, Macro F1-score, and Weighted F1-score revealed that the traditional Linear SVC model significantly outperformed the IndoBERT model in our experimental setup, achieving an Accuracy of 97.60% and a Macro F1-score of 0.5510, compared to IndoBERT's 88.70% and 0.5088. Detailed analysis indicates that this performance gap was primarily driven by discrepancies in the data sampling regimes, where baselines utilized the full corpus while the Transformer was constrained to a sampled subset. Finally, we demonstrate the practical viability of our pipeline by deploying the final sentiment classification model as an interactive Gradio web application.

This paper describes a submission to the Environment-Aware Speech and Sound Deepfake Detection Challenge (ESDD2) 2026, which addresses component-level deepfake detection using the CompSpoofV2 dataset, where speech and environmental sounds may be independently manipulated. To address this challenge, a dual-branch deepfake detection framework is proposed to jointly model speech and environmental contextual representations from input audio. Two pretrained models, XLS-R for speech and BEATs for environmental sound, are used to extract complementary contextual representations. A Matching Head is introduced to model representation differences through statistical normalization and representation interaction, enabling estimation of the original class. In parallel, multi-head cross-attention enables effective information exchange between speech and environmental components. The refined representations are processed with residual connections and layer normalization, and passed to an AASIST classifier to predict speech-based and environment-based spoofing probabilities. The model outputs original, speech, and environment predictions. On the test set, the proposed system achieves an F1-score of 70.20% and an environmental EER of 16.54%, outperforming the baseline system.

Large language models often fail at simple counting tasks, even when items to count are in the prompt. We investigate whether this failure occurs because transformers do not represent counts internally, or because they cannot convert representations to the correct output tokens. Across three model families: Pythia, Qwen3, and Mistral, ranging from 0.4B to 14B parameters, we find evidence for the second explanation. Linear probes recover the correct count from intermediate layers with $R^2>0.99$, showing that the information is present. However, the internal directions that encode counts are nearly orthogonal to digit-token output-head rows ($|\cos| \leq 0.032$). In other words, the model stores the count in a form that the digit logits do not naturally read out. We localize this failure with two interventions. Updating only the digit rows of the output head (36,864 parameters) substantially improves constrained digit prediction (60.7--100.0% on four tasks), but it does not fix unconstrained generation (0%); we do not claim that digit-row repair fixes open-ended text. By contrast, small LoRA on attention Q/V (7.67M parameters) improves upstream routing and achieves 83.1%$\pm$7.2% in true greedy autoregressive generation (deployable fix). Logit-lens at layer 35 (entity counting; correct-digit rank): (i) median over 3 seeds drops from order-$10^4$ to 1; (ii) seed 42 shows $54{,}332 \to 838$ (median top-1 while one seed stays far below). Norm, logit-lens, and cross-task analyses generalize the bottleneck to counting, addition, and list length; nulls on MMLU and GSM8K and limited DROP transfer. These results identify counting failure as a geometric readout bottleneck, not an internal-representation failure: the model knows the count but the output pathway is misaligned with tokens needed to express it.

A method is presented for reducing the cost of representative token selection in transformer attention layers by exploiting the coherence of the representative set across depth. Activation Decorrelation Attention (ADA) selects $r \ll T$ representative tokens at each layer via a Gram threshold and computes attention on the compressed $r \times r$ problem, but the selection requires a $T \times T$ Gram matrix at every layer. The cascade mechanism introduced here inherits the representative set from layer $l$ to layer $l+1$, validates it via a $(T - r) \times r$ cross-Gram computation, and updates it with a small number of additions and removals. The cost of the selection step drops from $O(T^2 d)$ to $O(T r d)$ per layer. Validation on three model families (GPT-2 124M, GPT-J 6B, OPT 6.7B) on AMD MI300X demonstrates Gram operation savings of $22\%$ to $63\%$ with mean Jaccard overlap of $0.83$ to $0.94$ between consecutive layers. The cascade reveals that the set of informative tokens is a structural property of the input that propagates coherently through the depth of the network: the same tokens carry the non-redundant information at layer $l$ and at layer $l+1$.

A method is presented for accelerating inference in transformer language models by exploiting the low effective rank of the token activation manifold at each layer. The method decomposes each activation vector into a subspace component and a residual, computes the linear-layer output on the subspace component via a cached low-rank weight image at reduced memory bandwidth, and applies a per-token gate that determines whether the residual correction is computed or skipped. The gate ensures that the output distribution is preserved to within a controllable tolerance. Validation on three model families (GPT-2 124M, GPT-J 6B, OPT 6.7B) on AMD MI300X demonstrates effective speedups of 3.0x to 10.5x on linear-layer weight reads with perplexity ratios below 1.00 and top-1 token agreement above 98%. The method requires no retraining, no architectural modification, and no approximation of the attention mechanism. At the operating point (k = 256, ε = 0.05) on GPT-J 14 6B, the accelerated model produces character-for-character identical output to the baseline.

Understanding whether deep neural networks are effectively optimized remains challenging, as training occurs in highly nonconvex landscapes and standard metrics provide limited visibility into layer-wise learning quality. This challenge is particularly acute for transformer-based language models, where training is expensive, models are often reused in frozen form, and poorly optimized layers can silently degrade performance. We propose a layer-wise peeling framework for monitoring training dynamics, in which each transformer layer is locally optimized against intermediate representations of the trained model. By constructing lightweight, layer-specific reference solutions and projecting layers onto multiple intermediate outputs via different permutations, we obtain achievable baselines that enable fine-grained diagnosis of under-optimized layers. Experiments on decoder-only transformer models show that these layer-wise reference bounds can match or even surpass the trained model at various stages of training, exposing inefficiencies that remain hidden in aggregate loss curves. We further demonstrate that this analysis remains effective under binarization and quantized settings, where training dynamics are particularly fragile. Across all numerical results, the proposed bounds consistently separate apparent convergence from effective optimality, highlighting optimization opportunities that are invisible when relying on training loss alone.

While linear-complexity attention mechanisms offer a promising alternative to Softmax attention for overcoming the quadratic bottleneck, training such models from scratch remains prohibitively expensive. Inheriting weights from pretrained Transformers provides an appealing shortcut, yet the fundamental representational gap between Softmax and linear attention prevents effective weight transfer. In this work, we address this conversion challenge from two perspectives: architectural alignment and representational alignment. We identify Test-Time Training (TTT) as a linear-complexity architecture whose two-layer dynamic formulation is structurally aligned with Softmax attention, enabling direct inheritance of pretrained attention weights. To further align representational properties, including key shift-invariance and locality, we introduce key instance normalization and a lightweight locality enhancement module. We validate our approach by linearizing Stable Diffusion 3.5 and introduce SD3.5-T$^5$ (Transformer To Test Time Training). With only 1 hour of fine-tuning on 4$\times$H20 GPUs, SD3.5-T$^5$ achieves comparable text-to-image quality to the fine-tuned Softmax model, while accelerating inference by 1.32$\times$ and 1.47$\times$ at 1K and 2K resolutions. Code is available at https://github.com/LeapLabTHU/Transformer-to-TTT.

Evidence derived from large-scale real-world data (RWD) is increasingly informing regulatory evaluation and healthcare decision-making. Administrative claims provide population-scale, longitudinal records of healthcare utilization, expenditure, and detailed coding of diagnoses, procedures, and medications, yet their potential as a substrate for healthcare foundation models remains largely unexplored. Here we present ReClaim, a generative transformer trained from scratch on 43.8 billion medical events from more than 200 million enrollees in the MarketScan claims data spanning 2008-2022. ReClaim models longitudinal trajectories across diagnoses, procedures, medications, and expenditure, and was scaled to 140 million, 700 million, and 1.7 billion parameters. Across over 1,000 disease-onset prediction tasks, ReClaim achieved a mean AUC of 75.6%, substantially outperforming disease-specific LightGBM (66.3%) and the transformer-based Delphi model (69.4%), with the largest gains for rare diseases. These advantages held across retrospective and prospective evaluations and in external validation on two independent datasets. Performance improved monotonically with scale, and post-training added 13.8 percentage points over pre-training alone. Beyond disease prediction, ReClaim captured financial outcomes and improved real-world evidence (RWE) analyses: for healthcare expenditure forecasting it increased explained variance from 0.28 to 0.37 relative to LightGBM, and in a target trial emulation it reduced systematic bias by 72% on average relative to Delphi. Together, these results establish administrative claims as a scalable substrate for healthcare foundation models and show that learned representations generalize across time periods and data sources, supporting disease surveillance, expenditure forecasting, and RWE generation.

Long-term time series forecasting requires models that simultaneously capture rapid oscillations, medium-range periodicities, and slowly evolving macro-trends from a fixed look-back window. Existing lightweight MLP-based models typically operate on a single temporal resolution, limiting their ability to explicitly model patterns at multiple scales. We propose MSMixer, a channel-independent multi-scale MLP architecture that addresses this limitation through three complementary innovations: (i) three parallel scale branches at down-sample factors {1x, 4x, 16x} with independent MLP blocks, (ii) a learnable softmax gate that dynamically weighs branch outputs, and (iii) a DLinear complementary shortcut that provides full-window trend and seasonality context. MSMixer contains only 112K parameters at H=96 and runs at O(T) complexity. Evaluated on four ETT benchmarks with standard chronological splits and three random seeds, MSMixer achieves the lowest average MSE (0.357) among lightweight models, outperforming DLinear (0.386, -7.4%) and NLinear (0.365, -2.1%), winning 12 of 16 configurations. Against five Transformer-based baselines from the literature, MSMixer achieves best or second-best MSE in 9 of 16 configurations while using 5x fewer parameters than PatchTST. Ablation and sensitivity analyses confirm the complementary contributions of the multi-scale branches and the DLinear shortcut.

DeepSeek-V3.2 and V4 introduce Compressed Sparse Attention (CSA): a lightning indexer (a learned scoring projection over compressed keys) scores them, the top-k are selected per query, and a sparse attention kernel reads only those. Public CSA implementations materialize a [B, S, H_I, T] FP32 score tensor before the top-k reduction. With H_I=64 indexer heads and the V4-Flash compression ratio m=4, that intermediate is 256 GB at sequence length S=65,536, exceeding any single-GPU high-bandwidth-memory (HBM) budget. We present StreamIndex, a Triton implementation of the CSA pipeline whose central component is a chunked partition-merge top-k driver that never materializes the full intermediate. On synthetic-but-realistic V4-shaped inputs at the indexer-step (layer) level on a single NVIDIA H200, the materialize path runs out of memory (OOMs) at S=65,536 with V4-Flash dimensions; StreamIndex runs the same indexer to S=1,048,576 with 6.21 GB peak HBM, a 32x regime extension. Set-overlap recall against the materialize ground truth is bit-exact at small S where both fit; across three 5-point design-space sweeps (chunk size, key-tile size, top-k), mean recall rounds to 1.0000 with min recall at least 0.9980 in every cell. The chunked driver composes with TileLang's pipelined attention kernel: at S=262,144 with V4-Flash dimensions, the materialize indexer paired with TileLang attention OOMs while the chunked indexer paired with the same attention runs in 1.97 s at 18.56 GB peak. Our contribution targets the indexer step; we make no claim of a faster attention kernel or of real-checkpoint end-to-end behavior. Code: https://github.com/RightNow-AI/StreamIndex.

Spiking Neural Networks (SNNs), particularly Spiking Transformers, offer energy-efficient processing of event-based sensor data for healthcare applications. Yet current architectures are rigid: they are trained and deployed as static networks with fixed parameter counts and computational graphs. This limits deployment on neuromorphic hardware such as Loihi and SpiNNaker, where on-chip constraints often require smaller models that trade accuracy for feasibility. We introduce the Elastic Spiking Transformer, a runtime-adaptive architecture that brings elasticity into the spiking paradigm. Inspired by Matryoshka-style representation learning, it embeds nested elasticity in the Feature Extractor, Spiking Self-Attention, and Feed-Forward blocks. Through granularity-aware weight sharing, a single universal model can dynamically slice network width and attention heads at inference time without retraining. This design provides two key advantages for SNNs. First, it allows the model to adjust its parameter footprint to different hardware memory budgets. Second, reducing active neurons also lowers spike firing rates, yielding proportional reductions in synaptic operations, an energy benefit not directly available in standard artificial neural networks. We evaluate the approach on CIFAR10/100, CIFAR10-DVS, and the EHWGesture clinical gesture understanding dataset. Results show that one Elastic Spiking Transformer spans a broad range of complexity-accuracy trade-offs, matching or surpassing independently trained baselines while supporting adaptive, real-time gesture recognition on resource-constrained edge devices.

This paper presents the first comprehensive comparative analysis of modern machine learning architectures for transliteration between Tajik (Cyrillic script) and Persian (Arabic script). A key contribution is the creation and validation of a unique parallel corpus aggregated from multiple heterogeneous sources, including crowdsourced projects, lexicographic pairs, parallel texts of "Shahnameh", diplomatic articles, texts of "Masnavi-i Ma'navi", official terminology lists, and transliterated correspondences. The initial dataset comprised 328,253 sentence pairs; a representative subset of 40,000 pairs was formed using stratified random sampling. The experiment compared six classes of models: rule-based baseline, LSTM with attention, character-level Transformer, G2P Transformer (trained from scratch), pre-trained multilingual models (mBART, mT5 with LoRA), and byte-level ByT5. Results demonstrate the overwhelming superiority of ByT5 (chrF++ 87.4 for Tajik to Farsi, 80.1 for reverse). The G2P Transformer significantly outperformed mBART (72.3 vs. 62.2 chrF++) despite limited data. Models using subword tokenization (mT5) failed completely (chrF++ less than 18.5). The findings demonstrate that for accurate transliteration of the Tajik-Farsi pair, architectures operating at the byte or character level are unequivocally more effective than traditional multilingual Seq2Seq models relying on subword tokenization.

Self-attention in Transformers is typically implemented as $\mathrm{softmax}(QK^\top/\sqrt{d})V$, where $Q=XW_Q$, $K=XW_K$, and $V=XW_V$ are learned linear projections of the input $X$. We ask whether these learned projections are necessary, or whether they can be replaced by a simpler similarity-based diffusion operator. We introduce \textbf{Gaussian Kernel Attention} (GKA), a drop-in replacement for dot-product attention that computes token affinities directly using a Gaussian radial basis function (RBF) kernel applied to per-head token features. Each head learns only a bandwidth parameter $σ_h$, while a single output projection $W_O$ preserves compatibility with the standard Transformer interface. GKA can be interpreted as normalized kernel regression over tokens, linking modern Transformer architectures to classical non-local filtering and kernel smoothing methods. We evaluate GKA in both vision and language modeling settings. For autoregressive language modeling within the \texttt{nanochat} framework, we implement causal masking and sliding-window constraints by masking and renormalizing the Gaussian kernel. At depth 20, a GKA model with $0.42\times$ the parameters and $0.49\times$ the total training FLOPs of a standard attention baseline trains stably, exhibits a near-zero train-validation gap, and demonstrates competitive behavior on standard benchmarks, albeit with higher bits-per-byte (BPB) at this compute scale. Overall, GKA provides a minimal, interpretable attention mechanism with an explicit locality scale, offering a dimension in the accuracy-efficiency trade-off for Transformer design.

Capturing user interests across extensive watch histories is critical for short-form video recommendation, yet scaling sequence length is limited by two bottlenecks: the semantic sparsity of atomic Video IDs and the quadratic computational complexity of Transformers. Traditional orthogonal Video IDs fail to capture content relationships and demand large embedding tables, while the quadratic complexity of self-attention restricts the maximum sequence length under strict industrial latency and resource constraints. In this work, we present a production-deployed framework for modeling ultra-long user behavior sequences at a billion-user scale. We first address the representation bottleneck by adopting content-native Semantic IDs. By utilizing depth-truncated, coarse-grained Semantic IDs, we shrink the embedding table size from corpus cardinality. This compact representation naturally generalizes to cold-start content through shared semantic prefixes. Second, to overcome the sequence scaling barrier, we introduce a Global-Aware Compression Transformer that leverages non-parametric temporal folding and unified global query integration to effectively condense the sequence, alleviating both the memory and computational bottlenecks of standard self-attention. Offline profiling on our computing infrastructure demonstrates an order-of-magnitude reduction in peak memory footprint and a drastic decrease in computational overhead. This efficiency gain enables supporting longer sequence lengths at an affordable cost in production, yielding substantial online gains in satisfied user engagement and satisfied content consumption in large-scale online A/B tests.

Autoregressive decoding becomes bandwidth-limited at long contexts, as generating each token requires reading all $n_k$ key and value vectors from KV cache. We present Stochastic Additive No-mulT Attention (SANTA), a method that sparsifies value-cache access by sampling $S \ll n_k$ indices from the post-softmax distribution and aggregates only those value rows. This yields an unbiased estimator of the post-softmax value aggregation while replacing value-stage multiply-accumulates with gather-and-add. We introduce stratified and systematic sampling to design variance-reduced, GPU-friendly variants. Evaluated on Llama-3.1-8B-Instruct at 32k-token contexts, S$^2$ANTA matches baseline accuracy while achieving up to $1.5\times$ decode-step attention-kernel speedup over FlashInfer and FlashDecoding on an NVIDIA RTX 6000 Ada. In batched long-context generation, these kernel gains translate to up to $1.25\times$ end-to-end decode-latency speedup. Finally, we propose Bernoulli $qK^\mathsf{T}$ sampling as a complementary technique to sparsify the score stage, reducing key-feature access through stochastic ternary queries. Both methods are complementary to upstream quantization, low-rank projection, KV-cache compression, and KV-cache selection methods. Together, they point toward sparse, multiplier-free, and energy-efficient inference. We open-source our kernels at: https://github.com/OPUSLab/SANTA.git

Existing large-scale sign language resources typically provide supervision only at the level of raw video-text alignment and are often produced in laboratory settings. While such resources are important for semantic understanding, they do not directly provide a unified interface for open-world recognition and translation, or for modern pose-driven sign language video generation frameworks: 1. RGB-based pretrained recognition models depend heavily on fixed backgrounds or clothing conditions during recording, and are less robust in open-world settings than style-agnostic pose-processing models. 2. Recent pose-guided image/video generation models mostly use a unified keypoint representation such as DWPose as their control interface. At present, the sign language field still lacks a data resource that can directly interface with this modern pose-native paradigm while also targeting real-world open scenarios. We present SignVerse-2M, a large-scale multilingual pose-native dataset for sign language pose modeling and evaluation. Built from publicly available multilingual sign language video resources, it applies DWPose in a unified preprocessing pipeline to convert raw videos into 2D pose sequences that can be used directly for modeling, resulting in a consolidated corpus of about two million clips covering more than 55 sign languages. Unlike many laboratory datasets, this resource preserves the recording conditions and speaker diversity of real-world videos while reducing appearance variation through a unified pose representation. Toward this goal, we further provide the data construction pipeline, task definitions, and a simple SignDW Transformer baseline, demonstrating the feasibility of this resource for multilingual pose-space modeling and its compatibility with modern pose-driven pipelines, while discussing the evaluation claims it can support as well as its current limitations.

Machine Learning surrogates for Computational Fluid Dynamics (CFD), particularly Graph Neural Networks (GNNs) and Transformers, have become a new important approach for accelerating physics simulations. However, we identify a critical bottleneck in the field: while architectures have advanced significantly, the common underlying training paradigms remain bound to naive assumptions, such as node-wise supervision and explicit Euler time-stepping. These legacy choices ignore the stiff dynamics and local flux continuity inherent to numerous partial differential equations resolution methods, such as Finite Element, Difference, or Volume (FEM). In this work, we propose a unified framework to bridge the gap between geometric deep learning and rigorous numerical analysis. We introduce three key innovations: (1) Multi Node Prediction, a stencil-level objective that predicts field values for a node's full local topology, enforcing spatial derivative consistency; (2) Temporal Correction, replacing unstable explicit schemes with a predictor-corrector via temporal Cross-Attention; and (3) Geometric Inductive Biases, leveraging 3D Rotary Positional Embeddings (RoPE) to robustly capture rotational symmetries in unstructured meshes. We evaluate this framework across three architectures (MeshGraphNet, Transolver, and a Transformer) on diverse physics datasets. Our approach yields consistent improvements in accuracy and stability, particularly in long-horizon rollouts, while producing latent representations that generalize to unseen subtasks such as Wall Shear Stress or Pressure prediction. Code is available at https://github.com/DonsetPG/graph-physics.

Pansharpening aims to generate high-resolution multispectral (HRMS) images by fusing low-resolution multispectral (LRMS) images with high-resolution panchromatic (PAN) images. However, the current mainstream frequency-based pansharpening methods employ fixed frequency filters, which cannot precisely adapt to complex and spatially diversified frequency distributions in PAN and MS images. Furthermore, existing denoising strategies insufficiently exploit frequency components for denoising and struggle to suppress various noise types accurately. To address these challenges, we propose CGFformer, a cluster-guidance frequency Transformer that focuses on varying frequency distribution and interactions between frequency and spatial components. Specifically, we design an adaptive separation module that integrates local features and non-local information through K-means clustering, enabling more precise separation of high- and low-frequency components. Subsequently, we introduce a dual-stream refinement module combined with Transformer-based cross-attention to remove various noise, allowing the network to jointly suppress frequency-relevant and irrelevant disturbances. In addition, we develop a frequency-spatial fusion module designed to enhance detail and facilitate spatial-frequency interaction, ensuring more effective reconstruction of spatial structures in the fused results. Extensive experiments on multiple benchmark datasets demonstrate that the proposed CGFformer achieves notable improvements over existing pansharpening approaches.

Recently, Convolutional Neural Network (CNN) or Transformer architecture based models have been proposed to overcome the limitations of Recurrent Neural Network (RNN) based models in spatiotemporal prediction. These models prevent the inefficiency of parallelization limitation due to the sequential properties and stacked error due to the recursive method, and show high performance. Novertheless, there are still some challengies. First, CNN based models have difficulty considering global information due to the local properties of the kernel, and their performance is limited. In addition, information is mixed because the time axis is combined with the channel axis of the image for processing. Models based on Transformer architecture have high complexity due to the self-attention calcuation and take a long training time. In this paper, we propose a new structure model called CNN-based Multi-In-Multi-Out model for Efficient Spatiotemporal Prediction (MIMO-ESP) to overcome these limitations. MIMO-ESP considers global information and significantly improves complexity by configuring a Transformer architecture based on CNN. In addition, it treats the time axis as an independent axis without combining it, and effectively considers spatiotemporal information together by applying dilation. This structure makes MIMO-ESP efficient and high performance. Extensive experiment results on three promising benchmark datasets which including video, traffic, and precipitation prediction tasks demonstrate that the usefulness of MIMO-ESP due to the achieved competitive efficiency while outperforming existing models. Furthermore, the ablation study results demonstrate the usefulness of the components of MIMO-ESP, emphasizing the potential of the proposed approaches.

Transformer-based models have achieved remarkable success across a wide range of domains, yet our understanding of their training dynamics remains limited. In this work, we identify a recurrent focus-dilution cycle in attention learning and provide a rigorous explanation in a one-layer Transformer setting for Markovian data via gradient-flow analysis. Using stage-wise linearization around critical points, we show that a single focus-dilution cycle can be decomposed into a sequence of distinct stages. First, embedding and projection rapidly condense to a rank-one structure, while attention parameters remain effectively frozen. Then, the attention parameters begin to increase, inducing a frequency-driven focus toward high-frequency tokens. As attention continues to evolve, it generates next-order perturbations in embeddings, leading to a mass-redistribution mechanism that progressively dilutes this focus. Finally, small asymmetries among low-frequency tokens lift a degenerate critical point, opening new embedding directions and initiating the next cycle. Experiments on synthetic Markovian data as well as WikiText and TinyStories corroborate the predicted stages and cyclical dynamics.

Generating expressive conducting gestures from music is a challenging cross-modal motion synthesis problem: the output must follow long-range musical structure, preserve beat-level synchronization, and remain plausible as a fine-grained 3D human performance. Existing conducting-motion studies are often limited by sparse pose representations, small-scale data, or evaluation protocols that do not directly measure whether music and gesture are mutually aligned. This paper presents TransConductor, a Transformer-based framework for music-driven conducting gesture generation. We introduce ConductorMotion, a SMPL-parameter data construction pipeline that recovers detailed body motion from conducting videos and forms a dataset targeted at professional conducting gestures. Given acoustic descriptors extracted from audio and an initial pose, TransConductor uses a Trans-Temporal Music Encoder and a Trans-Temporal Conducting Gesture Decoder to autoregressively predict SMPL pose parameters. To better assess artistic correspondence, we further build a retrieval-based evaluation model that embeds music and gestures into a shared space and yields FID, modality distance, multi-modality distance, and diversity metrics. Experiments show that TransConductor outperforms dance-generation and conducting-generation baselines, while ablations verify the benefits of the Transformer backbone and the proposed alignment loss.

We study the use of transformers to reconstruct the compositions of tensor products of two-dimensional rational conformal field theories (RCFTs) based on their low-energy spectra. The task is challenging due to its combinatorial nature. The constituent theories are characterized by their central charges and affine Lie algebra labels. We achieve 98% accuracy in recovering the constituents of tensor products theories constructed from Wess-Zumino-Witten models. We further demonstrate that our method generalizes to CFTs with larger central charge and unseen classes of RCFTs by adding a small number of out-of-domain examples. Our results show that transformers are effective at this task and point towards a new tool for bulk reconstruction in AdS/CFT.

The transformer is the most popular neural architecture for language modeling. The cornerstone of the transformer is its global attention mechanism, which lets the model aggregate information from all preceding tokens before generating the next token. One common variant of attention is called local attention, which restricts each token to aggregating information from a bounded window of predecessors, reducing the quadratic cost of global attention to linear. Although this restriction is usually motivated by efficiency, it has also been found to improve model quality, a phenomenon that has so far lacked a satisfactory explanation. We provide a formal account of this phenomenon in terms of recognizer expressivity. It has been shown that fixed-precision transformers with global attention correspond to a fragment of linear temporal logic containing a single past operator. We additionally prove that adding local attention introduces a second temporal operator, strictly enlarging the class of recognizable regular languages. Moreover, global and local attention are expressively complementary: neither subsumes the other, and combining them yields the richest fragment. Experiments on formal language recognition and natural language modeling corroborate the theory, showing that hybrid global--local transformers outperform their global-only counterparts.

Reasoning language models often generate long chain-of-thought (CoT), which accumulates a massive KV cache during the decoding phase and incurs high decoding latency and limited throughput. To address these issues, KV cache compression has emerged as a promising technique for reducing memory overhead by selectively removing unimportant KV pairs while preserving useful ones for subsequent decoding. Nevertheless, we identify two key limitations in existing KV cache compression methods: 1) their threshold-triggered compression policy may provide limited throughput improvement or even reduce throughput, and may fully eliminate KV pairs from certain blocks of the sequence, potentially worsening information loss. 2) they typically retain either isolated KV pairs or fixed-size chunks with rigid boundaries, failing to preserve important flexible-sized chunks at arbitrary token positions. To overcome these limitations, we propose Kara, a sliding-window KV cache compression method that performs decoding-time compression by operating only on the recently generated context. Kara leverages bidirectional attention to score and select informative KV pairs in the window. To enable flexible preservation of important semantic information, we design a Token2Chunk module to expand a subset of selected KV pairs into chunks. Furthermore, we adapt Kara to PagedAttention and develop KvLLM, an inference framework built upon vLLM, which reduces KV cache memory usage and effectively improves output throughput. Extensive experiments demonstrate consistent performance improvements of proposed Kara and KvLLM.