Transformers

Latent diffusion models (LDMs) enable efficient high-resolution image synthesis by denoising in a VAE-compressed latent space. However, fixed visual tokenizers can discard fine textures and structural details, while separate representation and diffusion training creates a mismatch between reconstruction and generation objectives. These limitations have renewed interest in pixel-space diffusion, which models raw pixels directly, removes the VAE bottleneck, and supports end-to-end optimization. This formulation better matches the demands of high-fidelity generation but introduces challenges in high-dimensional modeling, including noise scheduling, loss weighting, token efficiency, and scalable architecture design. Pixel-space modeling also offers a promising basis for unified multimodal systems: raw pixels, text, and task conditions can be represented in a shared token space and jointly processed by a single Transformer, narrowing the gap between visual understanding and generation. This paper reviews Pixel-Space Diffusion Transformers (pDiTs) from the perspectives of model architecture, continuous generative mechanisms, and unified multimodal modeling. We summarize representative methods, identify key technical challenges, and discuss future directions toward high-fidelity, end-to-end vision foundation models that integrate generation and understanding.

Graph foundation models (GFMs) with global attention are increasingly used to represent mixed-integer linear programs (MILPs), aiming to capture structure beyond the locality of standard graph neural networks. We study their expressive power through graph isomorphism testing, asking which MILP instances they map to identical representations. We prove that a broad class of hierarchical graph transformers combining global linear attention, edge-weighted cross-attention, and bipartite message passing is bounded by the one-dimensional Weisfeiler-Leman (1-WL) test: under any parameter setting, 1-WL-equivalent MILP graphs receive identical graph embeddings. Our compositional proof shows that each architectural component is a symmetric multiset function and thus preserves 1-WL equivalence. We validate this characterization across ten diverse graph encoders, including Graphormer-, GraphGPS-, Set-Transformer-, and Gasse-style models. Across model capacities, graph scales, and pooling operators, every tested encoder maps 1-WL-equivalent non-isomorphic graph pairs to numerically identical embeddings. Consequently, graph invariants that vary within a 1-WL equivalence class cannot be recovered from these representations. We further show that expressiveness beyond 1-WL arises from input encoding rather than attention: random-walk positional encodings separate the constructed pairs, while additional constructions expose the limits of this remedy. These results characterize the expressive power of global-attention GFMs and provide an encoder-agnostic diagnostic for detecting 1-WL-induced representation equivalence.

Deepfake generation has raised growing concerns regarding digital media authenticity, misinformation, identity fraud, and public trust. Recent studies show that combining spatial and frequency features leads to stronger detection results than using independently. This paper presents MSCA-FFT, a Fast Fourier Transform (FFT)-based multi-scale cross-attention framework for image-level deepfake detection. The model combines a partially fine-tuned Xception spatial branch with an FFT-based frequency branch. The frequency branch processes the log-scaled FFT magnitude spectrum through shallow convolutional layers, avoiding inverse frequency-to-image reconstruction used in DCT-based pipelines. The spatial and frequency representations are refined by transformer encoders, fused through cross-attention, and passed to an MLP classifier for real/fake prediction. Experimental results show that MSCA-FFT achieves consistently higher performance than the DCT-based state-of-the-art spatial-frequency fusion method and the compared baseline models. The ablation study further indicates that the FFT-based frequency branch provides complementary spectral cues when fused with spatial features. In addition, FFT-based frequency analysis and Grad-CAM/LIME explanations show consistent evidence around manipulation-sensitive facial regions, including the eyes, mouth, nose, and facial boundaries.

Linear attention promises constant-time recurrent inference but degrades sharply on associative recall. We formulate attention recall as a spherical-packing problem and introduce Kernelized Linear Attention Activations (KATA), a framework whose feature maps are derived from first principles by certifying nonnegative attention weights through a self-dual homogeneous cone. Building on this observation, we show that rank-one positive semi-definite (PSD) features offer a favorable capacity--interference tradeoff. KATA recovers a parameter-free convex output gate and characterizes associative capacity through the Welch interference floor. For tolerances above this floor, KATA enlarges the state without adding parameters and admits spherical codes with exponentially many keys in the projection dimension. We implement KATA as fused Triton kernels at two operating points: a flash-attention-style forward up to ${\sim}1.6\times$ FlashAttention-2 throughput, and an exact $O(T)$ chunked-state form that reaches ${\sim}11\times$ FlashAttention-2 forward throughput at $131$k tokens. An associative scan of the first-order feature lowers the inter-chunk recurrence depth to $O(\log(T/C))$ for chunk size $C$ and averages ${\sim}2.4\times$ the throughput of a matched sequential linear-attention baseline. On long-range MQAR and repeated-key overwrite, several KATA variants outperform Gated DeltaNet, with parameter counts and state sizes reported alongside accuracy. Induction preserves near-perfect recall, while kernel benchmarks show that the maps can be implemented efficiently. KATA retains $0.985$ MQAR at a $16\times$ out-of-distribution length, approaching the softmax with roughly one quarter of the KV-cache entries. Experiments on 340M-parameter LLMs reveal a feature-dependent fluency trade-off and clarify how positional embeddings, delta rules, and decay gates interact with feature geometry.

Nuclear fusion has made significant progress in recent years and is expected to become one of the most important pathways to addressing global energy challenges. This paper focuses on observing plasma using visible-light cameras, analyzing its spatio-temporal motion cues, and predicting the two-dimensional spatial distribution of light intensity, aiming to provide a foundational basis for future scientific experiments using deep neural networks. Specifically, we propose Delta-InvFormer, a novel backbone network centered on a differential Transformer. The key insight is that by taking consecutive video frames as input, we can better capture the dynamics of the plasma. Moreover, spatial and temporal differential self-attention effectively mitigates interference from noisy signals, ensuring high-quality feature extraction. These features are then fused into a compact and informative representation, which is fed into a decoder network to predict the distribution. Based on real experimental data collected from the Experimental Advanced Superconducting Tokamak (EAST) large-scale scientific facility, our results demonstrate that the proposed model not only significantly accelerates traditional methods for distribution prediction but also achieves competitive reconstruction accuracy. The source code of this paper will be released on https://github.com/Event-AHU/OpenFusion

We present a practical recipe for building a compact Hindi text-to-speech (TTS) model by distilling a large flow-matching teacher (IndicF5, 337M-parameter DiT) under a severe data budget (~17.6 hours). Training a small model from scratch on this much data fails outright. Instead we warm-start the student from the teacher by pruning depth only: keeping the teacher's width, text dimension, attention heads, and mel/text I/O fixed so all non-block tensors copy one-to-one, and retaining an evenly-spaced subset of transformer blocks. We first measure how much depth the teacher tolerates (it remains near-functional at -27% blocks but collapses past -50%), then descend gradually (22 -> 16 -> 12 -> 8 -> 6 blocks), re-fine-tuning after each prune, with each step gated by an objective ASR word-error-rate (WER) check. The resulting students reach WER 0.00 on unseen sentences at 249M and 190M parameters, and remain robust down to 131M; at 102M we observe a clear capacity cliff that we attribute to the data budget rather than the recipe. We also document two train/inference feature- and library-parity failures (mel filterbank and rotary-embedding library versions) that silently degrade audio, and a version-independent fix. The method yields a high-quality Hindi voice that runs in real time on a 6 GB laptop GPU. An independent 50-sentence FLEURS benchmark compares the released 190M student against its teacher and MMS-TTS-hin.

Prostate cancer diagnosis with multiparametric MRI (mpMRI) is commonly based on PI-RADS assessment or binary classification, which suffer from subjectivity and fail to capture clinically relevant pathological heterogeneity. To address this limitation, we construct a Prostate Cancer Histopathology Spectrum Dataset (PCa-HSD) and formulate a clinically meaningful four-class classification task, addressing the underrepresentation of benign lesions that are easily confounded with prostate cancer in existing datasets. We propose Language-guided Segmentation-assisted Diagnostic Transformer model (LSDT), which leverages zero-shot segmentation to provide anatomical priors and performs effective multi-modal slice fusion for classification. Our proposed method consistently improves accuracy across backbones, achieving the best average accuracy of 0.633 and JointRecall of 0.768 in five-fold cross-validation on a cohort of 344 patients. These results demonstrate that integrating pathology supervision and anatomical priors significantly enhances fine-grained prostate MRI classification and provides a more clinically relevant paradigm for risk stratification. Code will be made publicly available in a future revision.

We present a continuous geometric framework that models the discrete algebraic operations of the Transformer architecture as an integro-differential equation (IDE) on a semantic fiber bundle $\calE = \calM \times \R^d$. Beginning from a single geometric axiom -- that the token sequence forms a discrete $1$-manifold equipped with a canonical measure lattice -- we translate every core component of the modern Transformer (RMSNorm, RoPE, Softmax Attention, FFN, Residual Stream, SGD, Weight Decay) into a cohesive vocabulary of differential geometry, measure theory, and stochastic calculus. The resulting framework yields quantitative predictions spanning entropic optimal transport (Attention as a Schrödinger bridge) and non-equilibrium thermodynamics (SGD as Itô diffusion violating detailed balance). We conduct a six-part experimental campaign across five architectures (Qwen3, LLaMA\nobreakdash-3.1, Gemma\nobreakdash-3, GPT-2, Mistral) spanning $124$M to $8$B parameters. The empirical observables are quantitatively consistent with the geometric predictions: the $ε^{-1/2}$ Lipschitz scaling calibration at machine precision ($R^2 = 1.000$), the Lie--Trotter operator-splitting torsion, the symmetric ablation instability confirming the Dual-Law of Topological Stability, the $\calO(1/\sqrt{k})$ thermodynamic suppression of Poincaré recurrence on the RoPE torus, the thermodynamic context-limit phase transition, and the Non-Equilibrium Steady State parameter vortex -- verified across two optimizers (AdamW and Pure SGD) to exclude momentum artifacts. The results demonstrate that analyzing Transformers through the lens of continuous stochastic differential geometry provides a predictive descriptive vocabulary for the stability limits, context bounds, and optimization dynamics of Large Language Models.

The deployment of reliable lithium-ion battery management systems is crucial for accelerating electrification, yet the joint prognosis of State of Health (SOH) and Remaining Useful Life (RUL) remains severely hindered by task heteroscedasticity. Conventional multi-task learning frameworks fail to balance the bounded, low-variance noise of SOH estimation with the unbounded, nonlinearly expanding uncertainty of long-term RUL predictions. Here, we present the Rotary SOH-Injected Prior Battery Transformer (RoSIP-Batt), a unified co-estimation framework that resolves these optimization conflicts. By formulating joint prediction as a Bayesian multi-task objective, RoSIP-Batt introduces a homoscedastic uncertainty weighting mechanism to dynamically scale task-specific gradients based on learned residual noise levels. The architecture leverages decoupled dual classification tokens and a per-dimension gated fusion mechanism, secured by a gradient-detachment operator to prevent high-variance RUL updates from corrupting the stable SOH representation space. To capture electrochemical degradation patterns without relying on absolute cycle steps, Rotary Position Embedding (RoPE) is incorporated into a shared Transformer backbone to model translation-invariant relative temporal profiles. Crucially, the intermediate SOH estimate is directly injected into the RUL regression head as a physical degradation prior. Evaluations across the NASA, MIT-Stanford, and HUST datasets show that RoSIP-Batt significantly outperforms state-of-the-art baselines, reducing SOH estimation error to 1.994% MAE on NASA and restricting RUL prediction error to 62.85 cycles on Stanford. These findings establish RoSIP-Batt as a highly generalizable, computationally efficient solution suitable for real-time embedded BMS deployment.

The deployment of reliable lithium-ion battery management systems is crucial for accelerating electrification, yet the joint prognosis of State of Health (SOH) and Remaining Useful Life (RUL) remains severely hindered by task heteroscedasticity. Conventional multi-task learning frameworks fail to balance the bounded, low-variance noise of SOH estimation with the unbounded, nonlinearly expanding uncertainty of long-term RUL predictions. Here, we present the Rotary SOH-Injected Prior Battery Transformer (RoSIP-Batt), a unified co-estimation framework that resolves these optimization conflicts. By formulating joint prediction as a Bayesian multi-task objective, RoSIP-Batt introduces a homoscedastic uncertainty weighting mechanism to dynamically scale task-specific gradients based on learned residual noise levels. The architecture leverages decoupled dual classification tokens and a per-dimension gated fusion mechanism, secured by a gradient-detachment operator to prevent high-variance RUL updates from corrupting the stable SOH representation space. To capture electrochemical degradation patterns without relying on absolute cycle steps, Rotary Position Embedding (RoPE) is incorporated into a shared Transformer backbone to model translation-invariant relative temporal profiles. Crucially, the intermediate SOH estimate is directly injected into the RUL regression head as a physical degradation prior. Evaluations across the NASA, MIT-Stanford, and HUST datasets show that RoSIP-Batt significantly outperforms state-of-the-art baselines, reducing SOH estimation error to 1.994% MAE on NASA and restricting RUL prediction error to 62.85 cycles on Stanford. These findings establish RoSIP-Batt as a highly generalizable, computationally efficient solution suitable for real-time embedded BMS deployment.

A Bayes-filtered transformer (BFT) is a transformer trained on sequences that are generated in two steps: first a latent task is drawn from a prior, then observations are drawn conditional on that task. Trained under autoregressive log loss, the BFT's next-token prediction, in the idealized limit, is the Bayesian posterior predictive distribution (PPD) induced by that prior and that conditional law. In practice the trained BFT is only an approximation of this ideal PPD, raising an interpretive question: what prior and posterior over the latent task has the trained BFT actually internalized? Existing work answers this question by comparing the trained BFT's predictions against the predictions of various "reference" posteriors, each standing in for a different candidate algorithm or computation the BFT might be implementing. This prediction-space comparison is fragile: different posteriors can share the same posterior-mean predictions. We use predictive Monte Carlo (PMC) as a general interpretability tool for any BFT: using only next-token generation, PMC returns an approximation to the implicit prior and posterior over the latent task, answering the interpretive question directly in latent space. We apply PMC to three stylized task families spanning 0-Markov and 1-Markov exchangeability. The phenomena previously reported in these settings remain visible in latent space. Code is available at https://github.com/afiq-aswadi/bft-pmc

Time series forecasting (TSF) is vital to many applications, yet existing models often struggle to capture the heterogeneous long-range global patterns and short-range local variations in multivariate time series. While some approaches partially model these dependencies, they often do not jointly exploit temporal and feature-wise information. To address this challenge, we propose HyBDM, a multi-scale hybrid model that decomposes temporal dynamics into global patterns and local variations, which are modeled by two specialized experts. The Global Patterns Expert employs an enhanced BiConv-Mamba module that integrates bidirectional convolutions, an M-SSM layer, a forgetting mechanism, and a GDD-MLP module for cross-channel modeling. The Local Variations Expert uses a Local Window Transformer (LWT) to perform efficient locality-aware attention with reduced computational complexity. In addition, a Multi-Scale Patcher and a Long-Short Router enable multi-resolution representations and adaptive fusion of the two experts. Experiments on six benchmark datasets show that HyBDM outperforms state-of-the-art methods in both forecasting accuracy and computational efficiency, demonstrating its effectiveness in bridging global-local dependencies for multivariate TSF.

Task arithmetic, sequential fine-tuning, activation steering, and first-order random search all operate through relatively small perturbations around an already trained checkpoint, and they rely on different local approximations: individual perturbations should be first-order predictable, task updates should compose with controlled interference, useful tangent structure should be stable and possible to estimate, and weight edits should have counterparts in representation space. We measure 8 such properties with the same harness around a multitask LoRA operating point, on 9 transformers (82M-7B), with a prospectively registered property list, thresholds, and test split. We find a shared one-direction validity window up to the tested scale $10^{-2}$, but no universal radius for pairwise composition or update ordering. Along individual directions, changes of the probe loss remain first-order predictable throughout the grid: a perturbation's effect on the loss is essentially its projection onto the gradient, which is also what makes local random search work. Pairwise structure, however, proves to be far more fragile: on over a third of the measured (model, task pair) combinations, two-update order sensitivity sets in strictly inside that window; task-gradient subspaces rotate within tens of steps; additivity under our fixed activation probe fails at full task-vector scale on several models, including both held-out 7B models; and no model median passes the registered global mean-vector weight-to-steering correspondence bar. For two sequential task-gradient steps, the leading order-dependent term is the Lie bracket $H_B\textbf{g}_A-H_A\textbf{g}_B$; its normalized prediction $c(η)=ηκ+O(η^2)$ tracks the measured defect at median ratio 1.002, while the onset scale $η^\dagger\approx0.10/κ$ spans three orders of magnitude across models and task pairs.

Linear-attention models replace the growing KV cache with recurrent states, but autoregressive decoding still reads, updates, and writes these states one token at a time. Speculative decoding can reduce this cost by verifying several draft tokens in one target pass, yet existing speculative systems are designed for Transformer KV caches. For stateful linear-attention targets, verification must follow recurrent dependencies across chains and branches, acceptance must update only the accepted state trajectory, and the drafter must avoid submitting candidates that waste stateful verification work. This paper presents SpecLA, a speculative decoding runtime for stateful linear-attention models. SpecLA verifies chains and trees with topology-aware kernels, stores compact factors produced during verification to recover accepted states, and uses confidence pruning plus a target-aligned EAGLE-style drafter to feed useful candidates to the verifier. On an NVIDIA H100 with a public GDN-1.3B target, SpecLA achieves up to 1.70x end-to-end speedup over autoregressive decoding.

Video Diffusion Transformers process long spatio-temporal sequences, making self-attention the main bottleneck in high-resolution video generation. Training-free sparse attention reduces this cost, but adaptive Top-$p$ routing creates uneven per-head workloads under multi-GPU sequence parallelism. The resulting workload heterogeneity turns sparse attention into a rank-level straggler problem. We present \method{}, a training-free sparse-attention system that improves the distributed execution efficiency of adaptive sparse attention under multi-GPU sequence parallelism. \method{} uses Top-$p$ routing, a Top-$k$ safety floor, and video-aware block organization as the sparse-routing frontend, then repairs the materialized mask at runtime. Runtime Load Balancing migrates a small number of heavy heads via P2P communication to shorten the current critical path. Slack-Aware Sparse Augmentation fills residual non-critical-rank slack with additional high-value blocks, while overlap hides scheduling and migration overhead behind existing computation. On step-distilled Wan2.2 I2V, \method{} reduces average load imbalance from 1.34 to 1.08 and delivers a $4.41\times$ attention speedup over FlashAttention, while achieving a $2.02$--$2.11\times$ DiT inference speedup with competitive video quality.

While large language models (LLMs) can solve advanced reasoning problems in seconds, we show that even frontier models fail to perform a much simpler operation: exactly copying an input string that lies well within their context windows. We attribute this failure to positional encodings in Transformer architectures, whose inductive bias favors copying through a shortcut based on matching local contexts rather than carefully locating the corresponding input positions. To address this issue, we introduce 2D-RoPE, which organizes text into a 2D grid rather than a 1D sequence and assigns each token a row ID and a column ID. Under this view, copying becomes simply retrieving input tokens at a fixed column offset, which makes the task easy to learn. In synthetic copy experiments, shallow Transformers with 2D-RoPE achieve perfect copying at input lengths hundreds of times longer than those seen during training, whereas standard positional encodings fall far behind. We further show that the advantage of 2D-RoPE language models on copy tasks consistently holds in large-scale pretraining on DCLM with model sizes up to 1.4B parameters. Overall, our results suggest that viewing text in 2D can benefit language modeling, and we hope this encourages future work to further explore the potential of 2D positional encodings.

In-context learning is a remarkable property of transformers and has recently received a lot of interest. In many studies of in-context learning, it has been shown that transformers are capable of implementing solver for linear and non-linear regression problems, in which the most of them implement gradient descent algorithm. However, it is still unclear whether those implementations have actually been acquired through training. In this paper, we construct a transformer with linear self-attention, which in-context learns the least squares estimate in a simple regression task. The point here is that the closed form (analytical) solution is approximately obtained by using layer normalization rather than an approximate solution based on gradient descent algorithm. Then, we show an experimental example, in which our implementation is mainly used in the transformer trained with l1 regularization when the target output is the least squares estimate.

Accurate load forecasting at multiple grid levels is essential for future smart grids, ranging from aggregated control area forecasts for balancing supply and demand to forecasts of individual end-consumer loads for demand-side management and energy management systems. We present a comprehensive benchmark for load forecasting across grid levels, comprising three datasets that represent a transmission system operator control area, low-voltage grid feeders, and individual end consumers. We evaluate ten methods for short-term load forecasting and find that Transformer-based approaches consistently outperform established methods, reducing forecast error by 6.6-10.7 %. To analyze the impact of architectural design, we introduce YAformer, a flexible Transformer architecture that integrates modifications from prior work and is optimized via hyperparameter optimization. However, the standard Transformer achieves superior performance, suggesting that these architectural modifications are not required for accurate load forecasting. We further evaluate the Transformer-based time-series foundation model Chronos-2, which demonstrates competitive zero-shot performance on two datasets but fails to accurately capture special events in the TSO data. Detailed analyses reveal model-specific strengths and weaknesses, and ablation studies highlight the importance of long input contexts, covariates and continuous retraining - aspects that are often overlooked in the time-series forecasting literature.

Recent advances in AI-generated content have driven widespread adoption of Diffusion Transformers (DiTs) for high-resolution, long-duration content generation. While parallelization techniques accelerate diffusion inference, they face significant scalability challenges due to excessive communication overhead in multi-node environments. We observe that sequence partitions in Context Parallelism (CP) exhibit distinct heterogeneity: spatially proximate partitions contribute more significantly to attention computation results. By mapping this heterogeneous pattern to hierarchical communication topology, we can access high-contribution partitions with reduced communication cost. This insight motivates our novel selective attention state mechanism that strategically balances partial attention computation and historical result reuse across denoising steps. We present DiTango, an efficient parallel framework for DiT generation. DiTango features an anchor-guided state selection planner that optimizes computation-reuse decisions for each partition, complemented by a runtime that orchestrates efficient state-centric operations. This design achieves superior system efficiency while preserving generation quality. Experimental evaluation on popular diffusion models demonstrates that DiTango achieves up to 1.9x end-to-end and 3.2x attention speedup with near-linear scaling in multi-node settings, while maintaining generation quality comparable to state-of-the-art approaches.

Machine learning-based intrusion detection systems (IDSs) often suffer from class imbalance and vulnerability to adversarial attacks, leading to degraded detection performance and reduced robustness. This study proposes a TabTransformer framework augmented by the Boundary-Seeking Generative Adversarial Network (BGAN) for flow-based intrusion detection using the CICIDS2017 dataset. BGAN serves a dual purpose by generating synthetic minority-class samples to mitigate data imbalance and producing adversarial samples to evaluate model robustness. Experimental results demonstrate that BGAN augmentation improves TabTransformer's Macro-F1 score from 82.96% to 86.50%, with the largest class-wise improvement observed for Web_Attack (F1 score: 0.29 to 0.61). Robustness evaluation shows that all non-augmented models experienced a 100% Performance Drop Rate (PDR) under adversarial testing, whereas all BGAN-augmented models achieved negative PDR values, indicating improved resilience. Furthermore, the augmented TabTransformer maintained stable and low False Triggered Rate (FTR) values (1.51%-2.92%) across all noise levels, compared with the BGAN-augmented Decision Tree, which reached 49.09% under benign perturbations. These findings demonstrate that BGAN consistently enhances both class balance and adversarial robustness, while the proposed BGAN-TabTransformer framework provides an effective and adaptive intrusion detection solution for adversarial network environments.

Transformer reasoning is limited by autoregressive decoding, which repeat edly compresses rich hidden computation through token space and makes it difficult for intermediate reasoning states to persist across time. We in troduce Transformers with Temporal Middle-Layer Recurrence (T2MLR), a transformers-based latent reasoning architecture that fuses a cached middle layer representation from the previous token directly into an earlier layer of the current token position, enabling abstract intermediate computation to persist across decoding steps with little inference overhead. Across natural-language pretraining and multi-hop reasoning finetuning, T2MLR consistently outperforms data- and parameter-matched Transformer base lines. Moreover, applying recurrence to only a localized middle-layer block (as little as 20% of the network) often outperforms full-layer recurrence. Im portantly, T2MLR does not require pretraining from scratch: retrofitting the recurrent pathway into an existing pretrained 1.7B Transformer and briefly finetuning substantially improves math reasoning, lowering the barrier to practical adoption. These results suggest that effective latent reasoning in Transformers does not require looping over all layers as in previous works, but can instead emerge more strongly from targeted middle-layer recurrence.

Understanding newspaper images remains a challenging task due to their complex, nested hierarchical structures and dense, heterogeneous layouts. In this paper, we explore two complementary approaches for newspaper structure understanding. First, we present a modular bottom-up pipeline that combines state-of-the-art open-source models: YOLO for layout detection, LayoutReader for reading order prediction, and a custom algorithm for article segmentation. This approach leverages existing robust components while maintaining flexibility and interpretability. Second, we introduce Tiramisu (Tiered Transformers for Hierarchical Structure Understanding), a novel end-to-end transformer-based architecture that explicitly models document hierarchy through an iterative tiered process. Tiramisu performs section and article separation, block localization, semantic categorization, and reading order prediction using highly parallelized attention mechanisms. Finally, we release Finlam La Liberté, a new dataset designed specifically for evaluating hierarchical information retrieval in historical newspapers. Experimental results demonstrate the effectiveness of both approaches in reconstructing complex newspaper hierarchies, with comparative analysis highlighting their respective strengths for scalable document digitization. The Tiramisu training code, including the synthetic newspaper generator, is available at https://git.litislab.fr/tiramisu/tiramisu-newspaper-articles-extractor.

Real-time video generation demands fast decoding as much as fast denoising, yet current latent video diffusion models rely on 3D convolutional decoders that are slow and memory-intensive at high resolutions or for long video. We introduce FlashDecoder, a fast, memory-efficient pure-Transformer video decoder that decodes latents to pixels frame by frame. At each step, the current frame attends only to a fixed-size window of past frames through a rolling KV cache. The fixed temporal window keeps decoding fast and memory bounded regardless of video length, enabling constant-latency streaming. Because frames are processed sequentially, temporal causality is enforced without explicit attention masks, enabling training at resolutions up to 1080p and matching the reconstruction quality of convolutional decoders. On the Wan2.1 and Wan2.2 latent spaces, FlashDecoder matches each convolutional decoder in reconstruction quality (e.g., 41.55dB vs. 41.49dB PSNR at 1080p) while decoding 3.6x-4.7x faster with up to 11x less memory on a single H100 GPU. With architecture-aware inference optimizations, the speedup widens to 12x.

This work introduces FilmGPT, an autoregressive transformer designed to address the challenge of video montage--turning a collection of raw, "unwatchable" footage into coherent cinematic sequences. Inspired by language learning in modern LLMs, we train a long-context autoregressive transformer on a large corpus of movies. The aim is to implicitly capture the "grammar" of film directly from data rather than from hand-coded rules. Unlike other generative models, FilmGPT does not generate any new video frames. Instead, at inference time, we introduce a footage-constrained decoding algorithm to select the best next shot from the input raw footage according to the statistical patterns learned from films. We first evaluate these learned statistics directly by using the FilmGPT autoregressive model for next shot prediction on a standard benchmark of shot sequence ordering, outperforming the previous state of the art. We then evaluate our footage-constrained decoding algorithm on the full film editing task via a user study, and find that our FilmGPT-based editing significantly outperforms previous approaches. Finally, we demonstrate the applicability of FilmGPT to a wide range of applications in video montage, from automatic video segment trimming to human-in-the-loop film editing.

Contact-rich manipulation requires policies to infer interaction state from signals that are often weakly observable through vision and kinematics alone. Action Chunking with Transformers (ACT) has shown strong performance in fine-grained manipulation, but many deployments collect demonstrations through leader-follower teleoperation, where tracking error between commanded leader motion and executed follower motion implicitly encodes contact, resistance, and constraint violation. This paper examines whether ACT's apparent force-awareness depends on this hidden interaction cue. We introduce an observation-centric ACT variant that predicts future follower joint states instead of leader commands, thereby removing the teleoperation-induced discrepancy signal while preserving the rest of the learning pipeline. We then evaluate whether simple joint-torque proxies, derived from onboard motor current or joint effort, can recover contact-aware behavior without external force/torque sensors. Across four real-world tasks spanning surface following, insertion, stiffness discrimination, and force-based stopping, removing the implicit cue leads to severe failures in force-critical phases. In contrast, torque-augmented policies recover robust contact behavior and improve the base ACT policy. These results demonstrate that, on real hardware, the implicit teleoperation cue is a recoverable source of force-awareness, where torque signals are available, a simple proxy matches, surpasses, or further enhances it.

Vision Transformer (ViT) has been widely used as a powerful framework for modeling global dependencies among image patches. However, its core component, self-attention assigns softmax-normalized relative weights to all patches, making it difficult to evaluate the relevance between patches independently. In visual recognition, images often contain many background or redundant patches, yet self-attention cannot explicitly reject such irrelevant patches, which may introduce unnecessary information into feature aggregation. To address this limitation, Screening has been proposed in the field of language modeling, where the relevance of each token is independently evaluated based on query-key similarity and low-relevance tokens are explicitly excluded through thresholding. In this work, we propose VisionScreen, a new vision model that extends Screening mechanism to visual recognition. VisionScreen treats image patches as tokens arranged on a two-dimensional grid and extends absolute relevance estimation based on query-key similarity to the two-dimensional spatial domain. This allows each patch to selectively aggregate only content-wise and spatially relevant patches without relying on competition among patches. Experiments on image classification benchmarks demonstrate that the proposed method outperforms conventional ViT. These results suggest that Screening can be effective for visual recognition, offering an alternative to relative feature aggregation based on softmax attention.

Single-task fine-tuning of graph neural networks (GNNs) for power grid problems exhibits a systematic failure mode: models that achieve the lowest in-distribution error degrade the most under topology shift. We term this topology overfitting: the tendency of task-specific gradient signals to encode relational structure particular to the training topologies rather than the underlying physics, causing models to fail on unseen grids despite strong in-distribution performance. To expose and address this failure mode, we introduce MxGPS (Multiplex GPS), a multiplex graph transformer that runs K task-specialised GPS branches over a shared node encoder, jointly trained on Static State Estimation (SSE) and AC Power Flow (PF) via a self-supervised pre-training and multi-task fine-tuning protocol, with a cross-branch attention module evaluated in ablation. The joint SSE+PF objective forces the shared encoder to simultaneously satisfy complementary gradient signals, preventing it from overfitting to topology-specific relational structure. Under a 3-fold sliding-window cross-validation spanning four unseen topologies (14-, 24-, 162-, and 300-bus), MxGPS attains 0% boundary violation rate (BVR) on all four zero-shot Power Flow topologies. Critically, models with substantially lower in-distribution PF error degrade by 190% to 1400% under topology shift, whereas MxGPS degrades by only 39%, an inversion that directly implicates topology overfitting as the failure mechanism rather than insufficient model capacity. With only 1.6M parameters (12x fewer than the GridFM reference baseline), MxGPS demonstrates that multi-task joint training is a principled and parameter-efficient mechanism for topology-agnostic generalisation in power grid foundation models.

Looped Transformers scale sequential computation by applying a compact stack of physical blocks for multiple rounds, increasing unrolled depth without increasing stored parameters. This reuse changes the residual-scaling problem: in an untied Transformer, each residual branch receives and applies its own parameter update, whereas in a looped Transformer one shared update aggregates gradients from repeated visits and is read back by those same visits in the next linearized forward pass. We formalize this tied-depth effect through a first-order perturbation bound controlled by a visit-alignment coefficient $κ_R$. The bound recovers the DeepNorm exponent when visits decorrelate, but in the conservative aligned regime it requires the exponent to increase from $1/4$ to $1/2$ as loop count grows at fixed physical depth. The resulting method, \textbf{DeepLoop}, keeps the Post-LN DeepNorm architecture and sets $α=(2N)^{1/2}$ and $β=(8N)^{-1/2}$ for unrolled depth $N$. On GPT-style looped language models at GPT-2 small and GPT-2 medium scale, DeepLoop is neutral when no physical block is revisited and improves validation loss and downstream accuracy once recurrent depth is activated. These results show that stable recurrent depth requires residual scaling rules that account for parameter visits, not only nominal layer count.

Scene Text Recognition (STR) remains challenging due to the diversity of text appearances, including curvature, rotation, and perspective distortion. Recent Transformer-based approaches perform well but usually rely on one-dimensional positional encodings that ignore the 2D spatial structure of text images. Axial 2D extensions of Rotary Position Embedding (RoPE) exist for vision Transformers, but they assume roughly square, isotropic image content and apply the rotation only within encoder self-attention. Scene text violates both assumptions: crops are markedly anisotropic, and STR models are encoder-decoder, so the decoder must relate its queries to the encoder's 2D layout through cross-attention. We introduce 2D-RoPE-STR, which adapts axial 2D-RoPE to this setting through (1) an anisotropic row/column dimension allocation matched to the aspect ratio of text, and (2) an extension of the rotary coupling into encoder-decoder cross-attention, letting autoregressive decoding steps attend to encoder tokens by their 2D layout, a setting not addressed by prior encoder-only formulations. Both changes are essentially parameter-free and require no architectural redesign beyond the positional-encoding module. We further introduce a diagnostic protocol (a controlled ablation pair isolating only the positional encoding, an image-level net-win disagreement analysis, and encoder attention visualization) that identifies where and why relative 2D position helps: curved, rotated, and perspective-distorted layouts where reading order departs from a straight horizontal line. On six standard benchmarks (IIIT5K, SVT, ICDAR 2013, ICDAR 2015, CUTE80, SVTP), gains concentrate on exactly these irregular layouts, with ablations isolating each design choice against 1D RoPE and 2D sinusoidal and learnable alternatives.

Attention-based KV cache eviction (H2O and its descendants) compresses the memory-constrained state of a long-context model by ranking tokens on accumulated attention mass, treated here as signal energy, and keeping the heaviest. On schema-dense input streams such as nested JSON, this score acts as a non-stationary filter that disproportionately retains noise: a non-content sink role (delimiters or whitespace) carries an order of magnitude more energy than any content role, and structural KEY tokens are over-retained at roughly 1.8x the rate of the answer-carrying VALUE tokens, collapsing exact-match accuracy from 88% to 0% at a 5% budget as the signal-to-noise ratio of the retained state degrades. A counterfactual experiment establishes that suppressing KEY tokens is the best deployable filter. Our retraining-free, role-conditional allocation over SnapKV's windowed score, governed by a single tuned hyperparameter, closes 63-98% of the H2O gap at sub-20% budgets and, at higher budgets, modestly matches or exceeds full-cache accuracy -- a small, seed-sensitive denoising effect (borderline significant at B=0.50; not distinguishable from zero at B=0.30 over four seeds). A 15 MB linear role probe supplies these labels at negligible inference cost, though matching parser-level downstream accuracy remains open.

Transformer language models are increasingly used as software components, yet biased outputs remain difficult to localize and repair inside the model. Existing fairness testing and repair methods largely operate at the input-output or retraining level, while recent work suggests that bias-related behavior can concentrate in a small set of attention heads. This paper studies whether attention heads can be localized and repaired through a targeted inference-time intervention. We introduce ROBIN, a white-box head-level fairness debugging method that ranks attention heads using sensitivity to fairness probes and removes a small bias subspace from selected head outputs. In a four-model pilot study, ROBIN reduces the measured WinoBias gap across all models while preserving language-modeling quality better than whole-head zeroing. These preliminary results suggest that head-level bias repair should consider not only which heads are selected, but also how selected heads are modified.

The $\mathcal{O}(N^2)$ complexity of attention over $N$ tokens remains a computational bottleneck in transformer models. Vector-Quantized (VQ) attention reduces this to $\mathcal{O}(MN)$ by representing keys with $M$ codewords, but applies uniform codebook capacity regardless of where attention mass concentrates: high-attention regions of key space may be coarsely approximated while low-attention regions waste representational capacity. We propose Adaptive Vector-Quantized (AVQ) Attention, which adaptively allocates codebook capacity based on attention importance. Starting from a small set of codewords, our method identifies the most important codes during the forward pass and refines them with pre-learned child codewords, achieving fine-grained quantization where it matters most while maintaining coarse quantization elsewhere. We develop an implementation using custom Triton kernels that enables the full adaptive refinement process, including importance scoring, child codeword insertion, and parent contribution replacement, to be carried out within the tiled computation paradigm of Flash Attention with minimal overhead. Our approach maintains $\mathcal{O}(MN)$ complexity while achieving improved accuracy-efficiency trade-offs compared to fixed-codebook VQ-attention.

The key-value (KV) cache has become the dominant memory cost of transformer inference. It grows with batch size, context length, and depth, and at long context it, rather than the model weights, sets the ceiling on throughput. Two families of methods reduce it. Low-rank methods factor two-dimensional slices of the cache, either per-head matrices or cross-layer feature blocks, and quantization methods lower the bit-width of every entry. Neither family exploits the fact that the cache at a layer is naturally a third-order tensor whose three axes, the heads, the tokens, and the features, carry very different amounts of redundancy. We take this tensor view directly. Our method, JoLT, applies a partial Tucker decomposition that compresses only the token and feature axes while leaving the head and layer axes intact, and then restores the energy that truncation discards with a Johnson-Lindenstrauss (JL) rotated low-bit residual. A single Lagrangian dual allocates the Tucker ranks and the residual bit-widths together, per layer group and separately for keys and values, under one byte budget. The result is a near-lossless 2-3x compression: perplexity, GSM8K accuracy, and RULER needle-in-a-haystack retrieval all stay at or within statistical noise of the uncompressed baseline on both a grouped-query-attention model (Mistral-7B-v0.3) and a multi-head-attention model (LLaMA-2-13B). At 2x, JoLT reconstructs the cache to relative Frobenius error 0.009 (K) and 0.006 (V) on both architectures, roughly an order of magnitude below cross-layer SVD and 4-bit quantization. A randomized-SVD variant, FlashJoLT, delivers a 5-13x compression-time speedup at matched quality.

Existing traffic forecasting models commonly focus on extracting spatial dependencies, particularly global spatial information, which characterizes the representations obtained through interactions between each individual node and all nodes across the traffic network. However, the underlying mechanism by which such global information is modeled and extracted remains insufficiently investigated. Whether global information must be extracted by high-degree-of-freedom adaptive attention or can be captured by a simple global aggregation operator remains unclear. For this purpose, we design a controlled ablation framework that replaces only the spatial mixing module to test attention-based global interaction. Across six traffic benchmarks, uniform full-range mixing and standard spatial attention each achieve lower MAE on three datasets, with only a 0.14% difference in mean MAE, while the former reduces node-scale spatial mixing complexity from O(N2) to O(N). Mechanism analysis further decomposes spatial attention into a row-uniform global background and a non-uniform residual. The residual shows dataset-dependent marginal value, suggesting that spatial attention should be justified by stable gains beyond a row-uniform global background. The corresponding source code is publicly available at: https://github.com/uuesti/U-Trans

Attention can be viewed as an online learner over context, yet existing test-time memories cannot certify that dropping a token leaves outputs unchanged or delete its influence outright. We introduce Support Vector Attention (SV-Attention), a max-margin memory whose weights are support coefficients of a one-class SVM with fixed box parameter C. Its active-set partition gives reserve tokens exactly zero weight, certifying output-preserving eviction; a reversible incremental solver deletes a token to recover the state produced by retraining without it under the same C. In fp64 experiments, decrement and refit recover identical partitions whenever the optimum is unique, and their decision functions match to a median deviation of about 10^-9 (10^-13 on learned keys); the 10^-2 worst case is confined to ill-conditioned duplicates and remains below coefficient decay in every regime. The exact path reuses the maintained KKT inverse in a custom backward. Training uses a separate stabilized batched approximation and does not carry the exact-deletion certificate; it reaches 9,125 tokens/s on a 3.22M-parameter model, while remaining 35.8 times slower than an MPS softmax reference. At matched budgets, certified selection reaches 0.86 vs. 0.32 rare-item recall and retains 0.80 vs. 0.05 deterioration hours on real MIMIC-IV streams. We also demonstrate surgical forgetting, exact editing, patient-record deletion, and a forgettable retrieval memory over real sentence embeddings. On enwik8, the hybrid obtains 2.178 BPC vs. 2.383 for a matched-state sliding-window Transformer across seven seeds (8.6% paired improvement, p=0.001); a three-seed TinyStories result is directionally positive but not significant (p=0.057).

We present a theoretical framework to explain the emergence of inductive reasoning abilities in Transformer language models. While previous works on Transformer learning dynamics have so far been mostly tied to specific tasks, we study a generalized class of inductive tasks that unifies several synthetic tasks known in the literature, including in-context n-grams and multi-hop reasoning. In this class, we theoretically prove that the training dynamics of attention models can be confined to a highly interpretable, low-dimensional invariant manifold. On this manifold, the learning dynamics are captured by a handful of interpretable coordinates rather than millions of parameters, making both theoretical and empirical analysis more tractable. Using this framework, we characterize how data statistics govern the competition between in-context and in-weights learning, we study how random initializations determine the `winning' circuit when multiple solutions are possible, and we demonstrate that the coordinate frame associated with the manifold can be used to automatically detect which circuits have been learned in trained models. By casting circuit formation as a low-dimensional dynamical phenomenon, we take a step toward a predictive theory of how Transformers learn.

A theoretical understanding of Transformers is crucial to better understand the capacities and limitations of large language models (LLMs). There is much work analyzing the expressivity of attention-based models. By proposing handcrafted weights or using computational complexity arguments, a large amount of past theoretical works have sought to characterize which tasks are and which are not in the hypothesis class of Transformer models. However, little work investigates the learnability of such solutions. In this work, we make progress towards this goal. Inspired by recent loss landscape analysis work, we propose preliminary sample complexity bounds for learning C-RASP constructions with Transformers.

Reconstructing lineages from live-imaging microscopy requires linking cell detections across time, including through cell divisions. A common approach is to construct a candidate graph and associate cell segmentations (nodes) across frames. However, these and other existing methods overlook two structural obstacles in candidate tracking graphs: (i) cell divisions entangle distinct lineage paths in the node embedding space, and (ii) edges sharing a node have near-random label agreement, so the candidate-graph topology carries no useful information for graph neural networks to aggregate. We propose the \textbf{Higher-Order Cell Tracking Transformer} (HOCT), an edge-centric architecture in which candidate cell links attend to one another under a 3D geometric prior, resolving both issues. Evaluated on the Cell Tracking Challenge and a bacteria division benchmark, HOCT achieves state-of-the-art results without deep pre-trained image encoders. Moreover, the proposed approach is easier to fine-tune, quickly reducing tracking errors by 59% with 400 annotations in a human-in-the-loop setting, outperforming LoRA fine-tuning of competing transformer baselines (6.75% improvement).

Feedforward network (FFN) blocks account for a large fraction of the parameters and computation in Transformer architectures, yet their internal structure remains difficult to interpret due to the additive superposition induced by the residual stream. We examine whether the activation of an FFN neuron can be explained by a sparse set of preceding neuron activations and attention outputs. We introduce a training-free attribution method that estimates the relative influence of upstream neurons and attention outputs on a target neuron's activation. Empirically, across models and layers, we find that small subsets of preceding activations and attention outputs suffice to preserve neuron activations with high fidelity when all remaining inputs are masked with their average values. Effective sparsity is even greater when accounting for the inherent activation sparsity of upstream layers. Moreover, applying the neuron-specific masks in all layers simultaneously, such that the induced deviations propagate through the network, leaves model perplexity largely unchanged at moderate sparsity levels. These results demonstrate that, despite dense parameterization, FFNs exhibit sparse and structured inter-layer dependencies at the neuron level. Our method provides a practical, scalable tool for circuit-level interpretability and identifies candidate sparse pathways with potential implications for efficient inference.

To ensure the overall quality of AI-enabled software, not only traditional software components but also AI components need to be tested and repaired. Among AI components, Transformer models are increasingly integrated into software systems, which makes their misbehaviors critical. Although prior work in the software engineering community has proposed deep neural network (DNN) repair methods, most overlook Transformer-specific structures. We propose RepTran, a search-based repair method for Transformer models. It targets their feed-forward networks (FFNs), which play a central role in the architecture. RepTran identifies suspicious weights by combining two types of scores: a variance-based neuron score and an existing bidirectional score. It then iteratively optimizes these weights using differential evolution. Our evaluation includes 18 fault benchmarks constructed from CIFAR-100 and Tiny-ImageNet. We compare RepTran against three baselines: random weight selection, Arachne (a state-of-the-art DNN repair method), and ArachneW, which enables Arachne to control the number of selected weights. RepTran achieved an average repair rate of 74.7%, statistically outperforming random selection and Arachne across all benchmarks. Effect size analysis revealed that RepTran achieved higher repair rates than ArachneW regardless of the number of selected weights. These results suggest that RepTran is effective for enhancing the reliability of AI-enabled software.

Reliable visual data association is fundamental to visual SLAM (V-SLAM), as it directly determines the quality of the camera pose estimation and map consistency. However, the handcrafted descriptors used by most mature real-time systems degrade under illumination and viewpoint changes, while learning-based front-ends that address this weakness typically require replacing the extraction-and-matching pipeline and introduce substantial computational overhead. Descriptor enhancement offers a compromise by refining existing descriptors within their original format, yet current methods rely on simplified attention mechanisms whose limited contextual modeling constrains the achievable matching quality. To resolve this trade-off between contextual expressiveness and efficiency, we propose Desc++, a lightweight enhancement module that jointly encodes descriptor representations and keypoint geometry and aggregates spatial context through a hybrid architecture that combines order-agnostic global attention with geometry-aware sequential modeling in linear time. The enhanced descriptors retain their original dimensionality and matching interface, enabling integration into deployed V-SLAM systems without modifying the pipeline. Experiments across descriptor matching, correspondence analysis, and system-level benchmarks with four different V-SLAM systems demonstrate that Desc++ improves matching accuracy over the state-of-the-art enhancement method, translates these gains into more accurate and stable trajectory estimation, and achieves a favorable balance between accuracy and efficiency for practical integration into existing real-time V-SLAM pipelines.

Diffusion Transformers (DiTs) have advanced video generation with high-quality, temporally coherent results. However, extending them to motion transfer, which requires following reference motion while aligning with a target prompt, remains challenging due to limited understanding of motion and structure representations within DiTs. We analyze video DiTs at the attention-head level and identify distinct heads specialized for motion and spatial structure. Based on this insight, we propose a head-aware controllable motion transfer framework that requires no parameter updates. Our method refines motion cues from motion-specialized heads via semantic correspondence guidance and preserves structure through selective feature injection. This head-level control not only enables accurate motion transfer but also provides an interpretable foundation for controllable video generation with DiTs.

Indexer-TopK, the operation to compute the scores and select the top-k candidates, is widely used by sparse attention kernels in large language models and vector retrieval in recommendation systems and vector databases. However, existing GPU-based Indexer-TopK kernels like DeepSeek Sparse Attention (DSA) remain inefficient due to excessive global memory traffic, costly synchronization, and prohibitive memory overhead. In this work, we exploit the curse of dimensionality in high-dimensional spaces, where distances between high-dimensional vectors tend to concentrate within a narrow range, to design LITETOPK, a novel and efficient fused Indexer-TopK kernel. LITETOPK first samples a small subset of data to estimate query-data score ranges, then uses these estimates to partition candidate results into bins online. This organization allows the LITETOPK kernel to maintain a tight approximate threshold, write back only promising candidates, reduce unnecessary I/O, substantially lower memory overhead, and still preserve exact Top-k correctness. Experimental results show that LITETOPK accelerates the prefill stage of GLM 5.2 by 1.2x in real-world deployment scenarios while incurring lower memory overhead.

Perineural invasion (PNI) is a critical prognostic factor in cholangiocarcinoma. Non-invasive prediction from 3D MRI is challenging, demanding models that efficiently capture both fine-grained details and global context. We propose the Multi-window Mixture-of-Head Attention Transformer (MMA-Former), a novel end-to-end 3D architecture featuring a Coarse-Fine Transformer (CFT) structure for parallel multi-scale feature extraction. We advance this structure by integrating a novel Window-Specific Mixture-of-Head attention (WS-MoH) mechanism. Unlike standard Multi-Head Self Attention (MSA), WS-MoH generates a representation for each 3D window and dynamically routes the entire window to specialized or common attention heads. This enables spatially adaptive feature extraction tailored to the local context of each window, enhancing specialization and reducing redundancy without increasing parameters. Evaluated on a retrospective dataset of 168 T1-weighted MRI scans, MMA-Former achieved an AUC of 0.752, outperforming other 3D architectures, including the best CNN (AUC of 0.708) and Transformer baselines (AUC of 0.681).

Small-data inverse design is challenging in engineering informatics when observations are heterogeneous, mixed-type, and constrained by physical relations among design variables. This work proposes a topology-aware surrogate framework guided by an Incremental Transformer (INCRT) for physics-constrained inverse design, applied to geopolymer mixture design. The method integrates intrinsic-dimensionality analysis, mixed-variable design-space representation, tabular surrogate prediction, INCRT-based manifold rationalisation, and constrained inverse optimisation. Using a public benchmark of fly-ash and slag-based geopolymer concrete mixtures with compressive-strength and carbon-emission targets, the high-dimensional design space proves strongly redundant, organising around fewer effective mixture regimes. Compressive strength requires nonlinear tabular surrogates, while carbon emission is largely determined by composition and well recovered by regularised linear models. INCRT thus acts not as a replacement for tabular predictors but as a rationalisation layer providing prototype regimes and a manifold-support score for inverse design. Three strategies are compared: unconstrained surrogate optimisation, physics-constrained optimisation, and topology-aware physics-constrained optimisation. Unconstrained optimisation can match target strength but may yield physically invalid or off-manifold candidates; physics-only constraints do not always ensure data support. The topology-aware strategy yields candidates balancing target compliance, carbon reduction, physical admissibility, and proximity to the learned feasible manifold. The framework aims not to replace experimental validation but to support screening of credible candidate mixtures from small, mixed, physically constrained engineering datasets.

Driving world models built on large video-diffusion backbones generate realistic scenes but are hard to control: enforcing a traffic norm typically means retraining the backbone or conditioning it on hand-built layouts. We ask whether controllability requires training at all. Our experiment shows that a rectified-flow driving world model, which jointly generates future video and a planned ego trajectory, can have its planned trajectory steered entirely at sampling time by differentiable energy functions that encode driving norms, without knowledge-specific retraining of the diffusion backbone. Concretely, we demonstrate that a world model built on Open-Sora 2.0 MM-DiT backbone can be steered to brake at a counterfactual target by injecting energy guidance at sampling time. However, we find that the generated video does not yet follow the steered trajectory through the backbone's joint self-attention and identify the cross-stream coupling as a crucial requirement for end-to-end-controllable rollouts.

In pre-LayerNorm looped transformers, LayerNorm inside the recurrent block acts as an implicit gain controller: by coupling the block's local Lipschitz constant inversely to the activation scale, it renders the recurrence Jacobian non-normal -- asymptotically contractive at every verified fixed point even where its operator norm exceeds 1 -- so the true stability budget is the spectral margin, not an operator-norm bound. That margin depletes as the carry $ρ\to 1$, and a minority of initializations never converge to a fixed point at all, so the diagonal carry constraint $ρ(\bar{A}) < 1$ is necessary but not sufficient for convergence of the full recurrence. Training experiments across six tasks, including a controlled ablation, reveal that the linear carry is not the depth-memory mechanism: gradient descent routes memory through the block's more expressive nonlinear recurrence and leaves the stability-constrained carry at rest -- the carry's role is stabilization, not memory. We characterize the boundary of this claim: on tasks with axis-aligned per-channel structure, gradient descent does recruit the carry. All results are derived analytically and verified in a from-scratch, CPU-scale implementation; verification at larger scale is needed.

Self-attention is a ubiquitous primitive in modern sequence models, yet its operator-level geometry is only partially understood. We view a token sequence as a vector field over the token-position graph and identify attention as a connection walk: messages are aggregated by a nonnegative walk matrix while being transported along each edge by a learned linear map. Within this framework, we prove that single-head attention (SHA) is exactly a connection propagation step with constant transport, and that multi-head attention (MHA) is exactly a single edge-dependent connection walk whose effective transport is an attention-gated mixture of headwise transports. We further clarify the conditions under which the corresponding generator reduces to a random-walk connection Laplacian, highlighting the roles of stochasticity, reversibility, and metric-compatible transports. Empirically, we find that trained Transformers across scales (from 124M to 8B) and structures (encoder/decoder) exhibit geometric structure consistent with our theory: effective attention graphs converge to stable geometric operators in deeper layers, learned transports self-organize into approximate scaled isometries, and both phenomena strengthen consistently with scale. Overall, the paper provides a precise connection-walk formalism that links self-attention to classical geometric operators, along with a set of operator-level tools for analyzing transformer models from a geometric perspective.

Normalization is a critical component for stabilizing Transformer training, yet the choice between static strategies such as Layer Normalization (LN) and adaptive alternatives remains largely task-dependent. In this paper, we investigate a key optimization challenge in differentiable normalization gating. Our experiments show that, on relatively stationary vision tasks, the high gradient variance introduced by Gumbel-Softmax gating can hinder convergence of the routing mechanism, causing learned gates to underperform simple random selection. In contrast, on non-stationary language modeling and classification tasks, sustained gating diversity enables the model to learn more effective layer-wise normalization policies. Motivated by these observations, we propose AutoNorm-S (Stabilized), a training strategy that mitigates optimization instability through a gate-freezing schedule. AutoNorm-S achieves competitive or improved performance across multiple benchmarks, outperforming adaptive normalization baselines on NLP datasets, including PTB and SST-2, while remaining competitive on standard vision benchmarks. These results suggest that decoupling normalization selection from optimization noise provides a practical and principled approach for adaptive normalization in Transformer architectures.

Brain field potentials are scale-free: their power spectra follow a $1/f^β$ law whose aperiodic exponent $β$ tracks cortical state, and sleep depth in particular is a shift in $β$. We ask whether a transformer endowed with an explicit renormalization-group (RG) inductive bias the RG-Flow Transformer, which couples ordinary self-attention to a scale-aware stream with a learnable anomalous dimension $γ$, block-spin coarse-graining, and an entropy-gated synchronization bridge has an advantage over a parameter-matched vanilla transformer on \emph{real, scarce} EEG. Using the PhysioNet Sleep-EDF corpus with a strict leakage-free by-subject hold-out, we (i) benchmark RG-Flow against a param-matched vanilla transformer and a hierarchy-only ablation on 5-class AASM sleep staging, (ii) sweep the per-subject data budget to look for the inductive-bias crossover predicted when data are scarce, and (iii) test whether RG-Flow's learned $γ$ tracks the measured spectral exponent $β$ out-of-sample a quantity the vanilla model does not possess. Across $5$ subjects and $5$ seeds under leave-one-subject-out cross-validation, RG-Flow and the vanilla transformer are statistically indistinguishable on 5-class staging (77.3% vs 77.0% accuracy; paired $p=0.294$), and the predicted scarce-data crossover does not appear: vanilla is numerically ahead at every data-limited budget. What does separate the models is interpretability RG-Flow recovers the continuous spectral exponent out-of-sample ($β$-recovery $R^2 = 0.416$), a capability the vanilla architecture has no analogue for.

We present a clinically deployed end-to-end auto-contouring system for cervical cancer radiotherapy planning, anchored by the Boundary-Aware Transformer with Region-Aware Mamba (BAT-RM), a hybrid architecture that integrates Sobel-gated boundary attention, a linear-time, multi-directional Mamba module for long-range context, and a boundary-skeleton-guided fusion gate. This design achieves linear-time complexity for long-range context modeling, avoiding the quadratic cost of full spatial self-attention. The full pipeline spans multi-institutional data collection, rigorous inter-rater quality assurance, external validation in an independent cohort, and a web-based clinical interface natively compatible with Varian, RayStation, and Monaco. Against four baselines, BAT-RM achieves superior performance across seven anatomical classes, with statistically significant improvements in target volumes, including GTV and CTV, and in organs at risk such as the rectum and bladder. A prospective multi-center reader study involving 13 radiation oncologists demonstrated that AI assistance elevates junior oncologists' IoU from 0.899 to 0.965, approaching senior-level accuracy, while reducing contouring time by more than 80%. The system also reduced expert consultation rates and improved inter-reader consistency, reflecting gains in both efficiency and quality assurance. Following clinical deployment at a partner hospital, the system reduced patient wait times from days to hours without additional staffing, enabling same-day or next-day initiation of treatment for routine cases. BAT-RM demonstrates that a rigorous research pipeline, from data curation to clinical deployment, can translate directly into measurable patient benefit in resource-constrained settings where the demand for radiotherapy far exceeds specialist capacity.

Vision Transformers (ViTs) are difficult to interpret because current methods of relevance propagation and attention flow do not fully consider some key architectural features, such as the uneven importance of attention heads and residual connections. Prior approaches typically assume uniform importance across attention heads; furthermore, they model skip connections as identity paths, leading to inaccurate relevance attribution. To address these issues, we introduce GradSkip, a novel relevance propagation method for ViTs based on adaptive head weighting and skip-aware propagation. GradSkip models the different importance of the attention heads and dynamically distributes relevance between the attention and residual paths. Experiments on ImageNet1K and BloodMNIST demonstrate a state-of-the-art faithfulness of GradSkip while requiring over 14 times fewer GFLOPs than the best-performing existing approaches. Additional evaluations using transformer-based segmentation confirm improved localization and alignment with ground-truth regions.

Large language models (LLMs) store factual knowledge in their parameters. While recent work has shown that this knowledge resides in MLP layers, existing constructive and mechanistic interpretability models of fact-storage in LLMs fail to explain the surprising empirical phenomenon that they store facts at an information-theoretically optimal rate. In this work, we develop a theoretical account of this phenomenon. We develop the first Transformer-compatible fact-storing MLP closed-form construction that satisfies the following three properties empirically observed in LLMs: it (i) attains optimal fact storage scaling, (ii) handles arbitrary input/output geometries, and (iii) works inside Transformers. Key to our work is to analyze the decoding margin of MLPs, whereas prior work only studies MLP fact storage. Under isotropic embeddings, our construction achieves information-theoretically optimal storage capacity scaling and requires $10$-$104\times$ fewer parameters at matched fact count than prior constructions. For arbitrary key and value embeddings, we show that our construction attains the same storage capacity scaling, up to penalization factors depending on the embedding geometries. Moreover, we demonstrate that our constructed MLPs can be used within Transformer blocks for factual recall tasks at optimal capacity scaling, requiring $15$-$63\times$ fewer parameters at matched fact count than prior constructions. Finally, as a proof-of-concept, we show that fact-storing MLPs enable modular fact editing by swapping a Transformer's MLP with a new one.

CorrNet is a strong baseline for continuous sign language recognition (CSLR) because it models inter-frame correlations inside the visual encoding stage. In this paper, we study two natural extensions of a reproduced CorrNet system: replacing the BiLSTM temporal head with a Transformer encoder, and injecting motion cues after temporal pooling. We find that the Transformer head does not outperform the BiLSTM baseline, even with a training strategy adjusted for the Transformer, and the two heads have almost the same computational and runtime cost. For the second extension, we design a lightweight module called MotionGate. In our experiments, MotionGate consistently collapses to an identity-like mapping: the gate loses motion selectivity, and the injected residual becomes a weak, non-selective perturbation of the pooled features. These results suggest that explicit motion injection after CorrNet's correlation-based encoding is largely redundant, and that natural-looking architectural extensions in CSLR should be tested carefully instead of being assumed to help.

Precision industrial contact manipulation requires reliable robot policies under pose perturbations and contact-force constraints. Vision-language-action models offer broad generalization but often introduce high inference latency and GPU-memory cost, while vision-action chunking policies are more suitable for real-time industrial control. However, these policies are usually trained by behavior cloning and suffer from distribution shift in contact-rich tasks. This paper proposes PAC-ACT, a reinforcement-learning post-training framework for pretrained Action Chunking Transformer policies. PAC-ACT reformulates policy optimization at the chunk level, constructs an ACT-transferred actor-critic architecture, and introduces a hybrid behavior-prior constraint to preserve the pretrained action distribution during online fine-tuning. Experiments on industrial precision-contact benchmarks show that PAC-ACT improves task success, contact stability, and force safety while retaining low latency and low GPU-memory usage. On the Contour task, PAC-ACT significantly reduces peak contact force and decreases the proportion of force readings above 60 N by 46 times. Sparse-reward ablations further show that the proposed behavior-prior constraint enables effective exploration under randomized initial poses.

Self-supervised pretrained foundation models (FM) have shown early promise for non-invasive electroencephalogram (EEG) decoding applications. Many recent large-scale models converged on the approach of tokenizing raw EEG followed by masked reconstruction pretraining. However, this recipe has been shown to be suboptimal for data, like EEG, with high noise amplitude and information confined to limited dimensions such as narrow frequency bands. Building on this insight, we develop a novel contrastive-pretrained EEG model with multiscale temporal convolution input layers and Transformer encoder blocks (CoCoT). CoCoT matches or beats state-of-the-art reconstruction-pretrained EEG models on extensive benchmark decoding tasks with heterogeneous electrode configurations. Furthermore, CoCoT trained from scratch outperforms previous single-task decoding models and even rivals pretrained models, showcasing the architecture's flexibility and data efficiency. Through systematic ablations, including model architecture and pretraining objective, we demonstrate the viability of contrastive learning for building EEG FMs while suggesting key architectural design considerations, prompting further investigations in alternative large-scale pretraining strategies.

The growing adoption of large language model-based agents within operating system workflows has increased the importance of energy-efficient inference on laptop-class systems-on-chip (SoCs). While cloud offloading remains common, it introduces reliability and privacy concerns that are particularly problematic for agentic workloads. Recent laptop SoCs, therefore, incorporate neural processing engines (NPUs) optimized for energy efficiency; however, effectively mapping attention mechanisms onto NPUs remains challenging due to architectural diversity and explicit data-movement programming models. In this work, we present STEEL, the first open-source implementation of FlashAttention targeting XDNA-like NPUs. STEEL introduces a dataflow formulation of prefill attention, enabling efficient exploitation of spatial parallelism and on-chip memory. Furthermore, STEEL addresses the load imbalance induced by the causal mask by leveraging a sparsity-aware pipeline placement onto the NPU array, reducing synchronization overhead and improving utilization. We evaluate STEEL on the AMD Ryzen AI 9 HX 370 SoC and compare its performance against optimized CPU and GPU implementations. Experimental results show that STEEL reduces energy consumption by an average of 9.17x and 1.75x relative to CPU and GPU baselines, respectively. On XDNA 1, STEEL achieves an average 9.6x latency reduction over the prior state of the art, and delivers a 22.8x speedup on average compared to a layer-by-layer attention implementation on XDNA 2.

Egocentric 3D human pose estimation from head-mounted stereo cameras is challenging due to fisheye distortion, severe self-occlusion, and frequent truncation of body joints outside the camera field of view. Recent stereo egocentric methods have improved performance through heatmap lifting, stereo correspondence, and transformer-based refinement, but they often rely heavily on frame-local evidence or use temporal information only as auxiliary pose-level context. This limits robustness when current-frame stereo cues are weak, occluded, or ambiguous. We propose TSR-Ego, a temporally guided stereo framework that couples short-term motion evidence with projection-guided feature sampling. The model first enriches dense stereo feature maps using a causal depthwise-separable temporal convolution, allowing past visual evidence to influence the feature space before deformable cross-attention. A single-stage causal stereo decoder then refines learned 3D joint queries through temporal self-attention, joint self-attention, and fisheye deformable stereo cross-attention, using the evolving pose estimate to generate 2D sampling references. Unlike methods that apply temporal reasoning mainly after pose prediction, TSR-Ego uses motion context to shape both the sampled stereo features and the joint representations while preserving online inference without future frames. Experiments on UnrealEgo2 and UnrealEgo-RW show state-of-the-art performance, with especially strong gains on real-world sequences.

Block sparse attention is a hardware friendly way to alleviate the key-value (KV) cache read bottleneck in large language models (LLMs). However, it is not prevalent among leading open-weight LLMs, which rely instead on dense attention or fine-grained selection, thereby motivating our analysis. We study DeepSeek's Native Sparse Attention (NSA) as a representative method, whose three-branch design lets us isolate block selection, the most challenging and consequential stage. We formalize selection and reduce it to ranking blocks by a single quantity, the attention mass: the sum of a block's attention scores. We show that if selection retrieves the blocks with the largest attention mass, block sparse attention can match the quality of dense attention. However, computing the exact attention mass requires reading every key, so the problem of block selection ultimately reduces to approximating this mass from a compact summary instead of the full keys. Via a cumulant expansion, we show why existing methods falter: their selection strategies attempt to estimate the attention mass, but are confined to a first-order approximation. Therefore, we propose COBS (Cumulant Order Block Sparse Attention), an attention method that builds on NSA, incorporating a novel selector that stores a compressed second-order statistic per block. On the 32k RULER long-context retrieval benchmark, COBS raises the NSA baseline's mean score from 0.2999 to 0.8195, approaching dense attention at 0.9040 and closing about 86% of the gap, while using only 1.21x the KV cache read traffic of the NSA baseline and 15.15x less read traffic than dense. The same model preserves short-context behavior and attains lower position-wise negative log-likelihood (NLL) than dense attention in our comparison.

Vision-Language Models (VLMs) have achieved success using homogeneous Transformers to process multimedia data. Recent studies show that heterogeneous structures interleaving efficient mechanisms, like linear attention, improve both performance and inference latency over homogeneous designs. However, these efforts rely on handcrafted static mixing patterns, which are sub-optimal and difficult to adapt to specific hardware. To bridge this gap, we propose Multi-Objective Search for Adaptive Inter-layer Composition (MOSAIC), a hardware-aware search method that automatically transforms homogeneous models into optimized heterogeneous architectures. MOSAIC integrates diverse efficiency mechanisms--including linear, sparse, and low-rank operators--into a unified search space. By formulating the selection as a multi-objective Mixed Integer Programming (MIP) problem, our method identifies optimal configurations that maximize downstream performance under strict hardware latency constraints. To mitigate performance degradation from structural transitions, we introduce a two-stage parameter recovery process: global off-policy distillation to stabilize internal representations, followed by a dual-teacher on-policy distillation leveraging a 235B oracle for knowledge expansion and the original 4B teacher for distributional stability. We validate MOSAIC through MOSAIC-4B, derived from Qwen3-VL-4B-Instruct. Results demonstrate that MOSAIC-4B matches the baseline's performance across multiple benchmarks while requiring less than 2% of the original training cost. Furthermore, it substantially improves inference efficiency, achieving 1.76x prefilling and 2.54x decoding speedups.