Transformers

Diffusion Transformers with Mixture-of-Experts (DiT-MoE) improve model capacity under sparse activation, but diffusion inference is still bottlenecked by redundant computation across timesteps. Existing caching methods mainly operate at the token level, which becomes suboptimal in DiT-MoE because each token update is internally decomposed into multiple routed expert branches. Our analysis shows that cross-timestep redundancy in DiT-MoE is better characterized at the expert-branch level than at the whole-token level. Based on this observation, we propose MoECa, a fine-grained caching framework that performs branch-level feature reuse across timesteps. MoECa further introduces expert-aware adaptive control and synchronized cache updates across MoE and attention paths to maintain stable intermediate states. Experiments on multiple DiT-MoE models show that MoECa consistently achieves a better speed-quality trade-off than prior caching methods, with up to 2.83$\times$ inference speedup and minimal quality degradation.

Transformer architectures have dramatically advanced representation learning and inference in deep models through self-attention mechanisms. In parallel,associative memory (AM) frameworks map representations onto energy landscapes, offering interpretable retrieval mechanisms. However, their continuous-time inference dynamics lack the biological plausibility of classical Continuous Attractor Neural Networks (CANNs). To bridge this gap, we propose Controlled Dynamics Attractor Transformer (CDAT), which couples a mixture von Mises-Fisher (Mo-vMF) attention energy with a Hopfield refinement energy, while augmenting energy descent with a CANN-inspired excitation-inhibition modulation. CDAT instantiates a topology-constrained dynamical system whose couplings encode relational structure among tokens, thereby linking attractor-style dynamics to modern energy-based attention. We further provide a constructive dissipation analysis to formally establish their controlled inference dynamics. Benefiting from these robust and structured dynamics, CDAT achieves state-of-the-art performance across multiple benchmarks in graph anomaly detection and graph classification.

The considerable layer-wise redundancy in large language models (LLMs) has established non-uniform sparsity allocation across layers as the standard pruning approach for efficient compression. Existing layer-wise allocation methods that estimate allocation strategy from local signals such as activation outliers or weight spectra mainly derive from local layer importance, whereas the final post-pruning performance is also influenced by the network's subsequent compensatory capacity. In this paper, we directly characterize this property through controlled perturbation experiments. We make the following empirical findings. First, layers exhibit highly heterogeneous responses to pruning-scale perturbations. In most cases, early layers amplify perturbations, while middle and late layers actively absorb them, with relative L2 drift decreasing monotonically across depth and direction realigning toward the unperturbed hidden-state trajectory. Second, absorption is a large-perturbation phenomenon. Under small perturbations the network exhibits amplification across all layers, and the transition to absorption occurs smoothly as perturbation magnitude grows to pruning scale. This enriches the linearized accumulation theory underlying related works. Building on these findings, we define an absorption coefficient per layer and propose absorption-aware correction, an orthogonal augmentation that improves OWL and AlphaPruning by reducing perplexity by 7.13% and boosting zero-shot accuracy by 1.02% across multiple model families at 70% sparsity.

KV cache eviction methods typically use a single retention-rule family throughout a model, making eviction-method identity a model-level design choice. Yet Transformer layers differ substantially in their attention behavior, representations, and sensitivity to compression, suggesting that a uniform rule may overlook useful layer-wise structure. This raises a basic question: should eviction methods themselves vary across layers? We investigate this question by composing existing eviction methods across Transformer layers and systematically exploring the resulting layer-wise design space. Using simple offline profiles, we construct fixed routes and study how their quality varies with method placement and cache budget. On LongBench, heterogeneous routing improves performance on a majority of tasks over homogeneous policies at the same cache budget. Even when method counts are held fixed, the profile-guided placement ranks second among 100 evaluated assignments, demonstrating that routing quality depends strongly on where methods are placed. Moreover, the same fixed route outperforms the best of nine standalone baselines across all five tested cache budgets. Together, these results establish layer-wise method composition as an exploitable, placement-sensitive design dimension for KV cache compression.

We train a two-layer transformer encoder to classify rational elliptic curves $E/\mathbb{Q}$ of conductor $\leq 10000$ as either rank 0 or rank 1 from the first 128 normalized Frobenius traces. We achieve >99% accuracy on both classes, and accuracy is essentially unchanged on test curves with no isogeny or quadratic-twist relative in the training set. We then apply techniques from mechanistic interpretability such as attention analysis, linear probing, activation patching, logit attribution, and neuron-level circuit analysis to reverse-engineer the algorithm the (centroid in function space) model learned. We find that a sparse circuit of 20 out of 512 layer-1 MLP neurons is sufficient for rank prediction under a linear probe with an AUROC of 0.992 at plateau, implementing a push-pull detector architecture of rank-0 and rank-1 detectors with a one-sided readout. However, we notice that the model has sub-optimal readout problems indicating a mismatch in rank-order between the readout pathway and the discriminative circuit. Critically, the learned input weights of the top discriminating neuron match the Mestre-Nagao sum heuristic weights $\log(p)/(p\cdot \log{B})$ with a Spearman coefficient $r = 0.997$ and Pearson coefficient $r = 0.952$: the model has learnt a result from analytic number theory from the Frobenius trace data alone. We additionally find that all 50 independently trained models concentrate CLS attention on prime positions at 2-50$\times$ the rate of composite positions. The CLS embedding encodes $\log{L(E,1)}$ with $R^2 = 0.962\pm 0.011$ across the 50 models (after controlling for the conductor). Activation patching analysis reveals that attention weights are dissociated from causal information flow. Additionally, the 50 solutions from training are near-identical in function space (with pairwise agreement $>$98.8%) despite large weight space barriers.

Mechanistic interpretability methods summarize a transformer component by a single importance score, conflating two distinct roles: a component may matter because it transports task-relevant content, or because the forward computation degrades when its contribution is removed. We introduce \emph{Interchange-Group Sobol Decomposition} (IGSD), a paired-intervention framework that compares matched activation replacement with zero ablation on the same component, estimates two Sobol-style variance indices, and uses their signed difference to separate the two roles, with intervention validity monitored by a symmetric off-manifold diagnostic $\widehat{\mathrm{ST}}>1$. In factual recall, IGSD identifies an early-layer content channel in both GPT-2 small and Qwen2.5-1.5B that standard importance methods underestimate. A controlled subject and relation donor design shows that the early channel transports relation-frame content while late attention transports subject-retrieval content, refining at head granularity to the known $\mathrm{Attn}_{L9H8}$ head. Late-layer clamping confirms that the early signal is expressed through downstream transformations rather than residual pass-through. These results show that replacement and deletion are not interchangeable controls and their divergence provides a practical statistical diagnostic for content transport in transformer components.

When humans see a bird, they recognize far more than just "bird" -- they see a head, wings, and talons, a structured assembly of reusable parts that can be identified across every bird they have ever seen. We ask whether a self-supervised visual model can discover the same compositional structure on its own. To this end, we propose RATS (Register Attention Transformers), which decomposes the classification token into N learnable register tokens that route patch information through an L->N->N->L bottleneck via a three-step compress-communicate-broadcast attention. The N registers are partitioned across the H attention heads, so that registers assigned to different heads do not interact with each other. Without auxiliary losses or part annotations, each register spontaneously specializes into a proto-semantic region whose emerging structure resembles object parts. RATS surpasses all baselines by +12 mIoU on average across five segmentation benchmarks, with consistent gains on ADE20K (+1.11 mIoU) and COCO (+0.2 AP^m). Its register dictionary further exhibits part-level consistency and semantic proximity across related categories. Our results suggest that RATS may provide a useful architectural prior for structured and interpretable visual representation learning.

Linear attention replaces softmax attention's growing KV cache with a fixed recurrent state, but this compression limits exact state tracking and long-context memory. We introduce \emph{Semidirect Fourier Delta Attention} (SFDA), a phase-controlled generalization of Kimi Delta Attention that replaces real diagonal decay with block-rotational Fourier control: $$S_t=(I-β_t k_tk_t^*)Λ_tS_{t-1}+β_tk_tv_t^*, \qquad Λ_t=\diag(α_t\odot e^{iθ_t}).$$ Our main result is a constructive chunk-WY factorization for products $A_t=Λ_t-u_tr_t^*$, giving $$A_t\cdots A_1=Γ_t-Y_tM_tW_t^*$$ with rank growth bounded inside fixed chunks. This yields an exact affine chunk transfer, formal stability and complexity bounds, and a compact characterization of phase-plus-low-rank memory. We verify the algebra numerically and show in toy state-tracking experiments that SFDA learns cyclic memory where the phase-disabled KDA baseline remains near chance. Fused kernels and large-scale language-model comparisons are left to future work.

Transformer-based neural operators have shown remarkable performance for approximating solution operators of partial differential equations on complex geometries. However, existing approaches implicitly assume a fixed domain size, which limits their ability to generalize at inference. In this work, we investigate domain extension, namely zero-shot inference on spatial domains that are significantly larger than those encountered during training. We argue that this setting fundamentally requires spatial locality and translation equivariance. We propose to implement this locality via a decomposable bias in the attention logits computation, enabling finely controllable locality while remaining fully decomposable into query-key inner products and directly compatible with optimized attention kernels. Combined with rotary positional embeddings, it enables expressive embeddings with controllable spatial support without altering the transformer architecture. We empirically show that our approach substantially improves zero-shot generalization to larger domains across two PDE benchmarks and a 3D industrial atmospheric flow application. Our code and datasets are available at https://github.com/cerea-daml/domain-extension.

Mixture-of-experts (MoE) layers enable the scaling of transformer models while keeping the inference compute fixed. While task-expert specialization has been observed in empirical studies of frontier MoE transformer models, existing theoretical work analyzes this using continuous mixture models that cannot be used to model natural language effectively. An important open question is to \textit{theoretically explain task-expert specialization in transformer MoE models using discrete models of language}. To address this, we represent structured knowledge via syntactic templates and finite key-value dictionaries, and prove formally that a single-layer MoE transformer can encode knowledge by using experts that specialize in the corresponding tasks. Our construction shows how queries are routed to unique, task-specific experts whose size depends solely on the intrinsic complexity of the given task (i.e. the combined size of its syntactic templates and factual dictionary). Our construction provides a theoretical support for empirical results on localized knowledge circuits in MoE models. We support our theoretical findings with experiments evaluating model performance under varying MoE loss functions.

Transformer feed-forward networks (FFNs) are often treated as nonlinear stores of computation, yet how nonlinear a trained FFN block actually is has rarely been measured. We treat each FFN as a position-wise input-to-output map and split it into the exact least-squares linear approximation plus a residual. The held-out variance the closed-form linear map explains defines a block's linear recoverability (R^2_lin), an optimiser-free measure of its linearity. Across all twelve blocks of GPT-2, Pythia-160m, and llama-160m, R^2_lin is highly heterogeneous and non-monotone with depth, ranging from near-linear (>0.99) to strongly nonlinear (<0.3) between adjacent blocks, and is not set by the activation function: same-width GELU models GPT-2 and Pythia-160m have sharply different profiles, so recoverability is a learned property of individual trained blocks, not an architectural one. A low-rank bilinear probe of the residual recovers only a few points of R^2, with gain uncorrelated with residual nonlinearity: the unrecovered computation is not a single position-wise product but higher-order or distributed structure. The measurement also serves as a targeted compression signal: recoverable blocks admit large single-layer replacements (GPT-2's early FFN at 8x fewer parameters for +0.77 perplexity), while low-recoverability blocks flag where this is unsafe. It further exposes a methodological pitfall: trained linear baselines can badly under-converge on ill-conditioned transformer activations, so we report the exact closed-form least-squares ceiling throughout.

The development of generalizable electroencephalography (EEG) decoding models is essential for robust brain-computer interfaces (BCI) and objective neural biomarkers in mental health. Conventional approaches have been hindered by poor cross-subject and cross-task generalization, owing to high inter-subject variability and non-stationary neural signals. We address this challenge with a zero-shot cross-subject decoding framework on the large-scale Healthy Brain Network dataset, benchmarking a convolutional neural network baseline, a hybrid LSTM, and a Transformer-based foundation model. To adapt the Transformer for regression while averting catastrophic forgetting, we propose a novel progressive unfreezing strategy. The baseline yielded an nRMSE of 0.9991, whereas our fine-tuned Transformer achieved 0.9799 on unseen subjects. This work advances scalable, calibration-free EEG decoding for computational psychiatry and behavioral prediction.

The detection of gravitational waves has revolutionized our ability to explore fundamental aspects of the Universe. Traditionally, modeled gravitational-wave signals have been identified using template-based matched filtering, followed by coincidence analysis across multiple detectors in the signal-to-noise ratio time series. Recent advances in Machine Learning and Deep Learning have sparked growing interest in their application to both signal detection and parameter estimation. In this study, a hybrid Deep Learning strategy is proposed that leverages the effectiveness of Transformer encoders alongside well-established Convolutional Neural Network architectures in an attempt to estimate the intrinsic and extrinsic parameters of non-precessing binary black hole systems. The primary focus of this work is point estimation, producing single best-fit values for each parameter rather than full posterior distributions. This method is evaluated on both simulated signals embedded in Gaussian noise and real gravitational-wave events, and it demonstrates strong predictive performance and robustness across key astrophysical parameters.

Remote photoplethysmography (rPPG) transformers achieve low heart-rate error on benchmarks, yet their decisions remain opaque--a growing concern as rPPG moves toward clinical heart rate estimation. Existing rPPG XAI is dominated by qualitative heatmap inspection without quantitative faithfulness metrics or physiology-grounded validation, leaving a gap between visual plausibility and auditable evidence. We address this gap. First, we adapt four attribution methods (raw attention, rollout, flow, Beyond Intuition) to RhythmFormer's bi-level routing attention with top-$k$ selection. Second, we introduce a skin coverage metric quantifying how much attribution mass falls on skin regions. Third, we adapt the SaCo faithfulness coefficient from its original classification setting to rPPG regression by using the MAE between original and perturbed predicted rPPG waveforms as the perturbation impact. Applying these tools, we quantify a multi-hop leakage effect under sparse top-$k$ routing: attention rollout and flow almost completely restores the connections that individual refined-attention layers explicitly set to zero. Beyond Intuition mitigates this via its value-projection-weighted rollout and gradient-supported mask, attaining the highest median refined skin coverage ($0.83$ vs. $0.57$ for vanilla rollout) and faithfulness ($F=0.92$) among the evaluated methods on UBFC-rPPG. Validation across diverse datasets and model variants is needed. A case study on a low-SaCo outlier further shows all four methods recovering consistently once an artefactual region is replaced, suggesting consistent SaCo behavior across attribution families in this illustrative case. Together, these metrics move XAI for rPPG toward auditable numerical evidence about spatial alignment and perturbation faithfulness, i.e. trustworthy rPPG XAI.

Whole-body mobile manipulation requires coordinating mobile base and manipulator under shifting viewpoints, posing challenges in geometric perception and action generation. Current policies either rely on 2D features or sparse 3D representations that lack dense spatial structure, and typically encode arm and base within one action vector that ignores their distinct control demands. Moreover, existing dense fusion strategies risk corrupting pretrained representations under noisy depth while incurring heavy computational overhead. We present GeoHAT, an end-to-end diffusion-based framework built on a simple principle: geometry should be injected only where reliable and attended to only where needed. GeoHAT employs a lightweight Fourier spatial encoder that maps dense per-pixel 3D coordinates into geometric tokens without an additional 3D vision backbone. These tokens are then selectively injected into vision foundation model features through per-token gated fusion modulated by depth validity, preserving the semantic prior while enriching spatial understanding. For action generation, a Hybrid Whole-Body Action Decoder decomposes arm and base into distinct subspaces and lets each action modality attend to its task-relevant visual context through sparse cross-attention, while causal temporal modeling captures intra-timestep coordination and inter-timestep dependencies. Experiments on the ManiSkill-HAB simulation benchmark demonstrate that GeoHAT achieves a 79.3% mean success rate, surpassing the strongest baseline by 23.7%. Furthermore, real-world experiments on diverse tasks also confirm consistent improvements over all baselines.

Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale. We introduce MiniMax Sparse Attention (MSA), a blockwise sparse attention built upon Grouped Query Attention (GQA). A lightweight Index Branch scores key-value blocks and independently selects a Top-k subset for each GQA group, enabling group-specific sparse retrieval while maintaining efficient block-level execution; the Main Branch then performs exact block-sparse attention over only the selected blocks. Designed around a principle of simplicity and scalability, MSA is deliberately streamlined, making it straightforward to deploy efficiently across a broad range of GPUs. To translate sparsity into practical speedups, we co-design MSA with a GPU execution path that uses exp-free Top-k selection and KV-outer sparse attention to improve tensor-core utilization under block-granular access. On a 109B-parameter model with native multimodal training, MSA performs on par with GQA while reducing per-token attention compute by 28.4x at 1M context. Paired with our co-designed kernel, MSA achieves 14.2x prefill and 7.6x decoding wall-clock speedups on H800. Our inference kernel is available at: https://github.com/MiniMax-AI/MSA. A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3.

Accurate prediction of solubility remains a central challenge across materials science and sustainable chemistry. In particular due to emerging technologies like organic and hybrid photovoltaics, batteries, and catalysis, solvent usage is expected to increase significantly within the coming years. Therefore, substituting solvents with greener alternatives is vital. This is where machine learning can have substantial impact. However, the limited data on critical parameters of solubility significantly constraints machine learning efficacy. In this work, we transfer a pre-trained foundational model on QM9 targets to our application with minimal data requirements. Additionally, the pipeline integrates uncertainty quantification, allowing the user to gauge the confidence of the predictions. As baseline, we succeed in predicting the Hansen solubility parameters and Dielectric Constant for which extensive databases exist. Importantly, we achieve high model performance on additional targets, such as Gutmann Donor and Acceptor numbers, where the available data is extremely limited. Overall, we augment data on solubility descriptors by orders of magnitude with high quality predictions. For effective dissemination, we deploy easy-to-use, easily integrateable with high throughput labs, customizable tool for ranking and screening possible solvent substitutes. Finally, we rediscovered known green solvent alternatives and proposed new candidates proving its relevance for finding eco-friendly solvents.

Crack detection plays an important role in infrastructure inspection and Structural Health Monitoring (SHM). However, cracks typically appear as thin, low-contrast structures and are easily affected by background noise, posing challenges for existing object detection models. This study proposes an improved YOLO-based architecture with integrated attention mechanisms, termed YOLO-AMC (YOLO with Attention Mechanisms for Crack Detection), to enhance automated crack detection performance. Based on YOLOv11, the original C2PSA module is removed, and multiple attention mechanisms, including Global Attention Mechanism (GAM), Residual Convolutional Block Attention Module (Res-CBAM), and Shuffle Attention (SA), are introduced into the multi-scale feature fusion layers of the Neck to strengthen cross-scale feature integration. Experimental results demonstrate that YOLO-AMC consistently outperforms baseline models YOLOv11n and YOLOv8n across multiple evaluation metrics. Among the evaluated attention modules, GAM achieves the best detection performance, obtaining mAP@0.5 = 0.9917 and mAP@0.5:0.95 = 0.9506 on the test dataset, which are higher than those of YOLOv11 (0.9833 / 0.9112) and YOLOv8 (0.9707 / 0.8921). Furthermore, while maintaining a computational complexity of 7.6 GFLOPs, the proposed model achieves 110.95 FPS on an NVIDIA RTX 4090 platform and approximately 5 FPS on a Raspberry Pi 5 edge device, demonstrating a favorable trade-off between accuracy and deployment efficiency. The implementation code for this study is available on GitHub at https://github.com/CY-Tsai24/YOLO-AMC.

The standard basis of transformer hidden states is a training-free, architecture-general feature basis for detecting concepts and, in language models, steering them; with no learned dictionary. Individual dimensions act as binary registers read one at a time: their signs (+/-1) encode content, their magnitudes strength. A feature is just a subset of dimensions with a consistent sign pattern, read by counting sign agreements. We validate this Bag of Dims (BoD) framework across seven models spanning language, vision, and audio; reading dimensions one at a time loses nothing, as a full-capacity MLP adds zero AUC over per-dim reading. The same per-dimension signs appear in every modality, so they reflect transformer training itself, not the language objective. Sign alone carries predictive content: setting all magnitudes to unity preserves 60-93% top-5 next-token accuracy through the LM head. From a single-token cache (one forward pass per token, no labels) we detect 175 categories at AUC 0.97-0.99 by counting sign agreements, and from random seeds alone discovery scales to 1500 features per model. A trained probe adds only +0.018 AUC and converges to axis-aligned weights: the rotation dictionaries learn buys little. Signs are causally operative: they survive the attention projections, and flipping a concept's sign pattern in the live forward pass suppresses it. Reading and steering are separate roles in the same basis: a concept's reader dimensions are not its writer dimensions. The writer target is just as cheap, the sign of the summed unembedding rows over a few seeds, no training. Injected through the attention output pathway under closed-loop control, it steers concepts into fluent text on four language models (62-92% of twelve concepts). The signs were in the standard basis all along; the open problem is no longer finding the right rotation but cataloging what each dimension encodes.

Transformers dominate modern sequence modeling, but their quadratic attention incurs substantial computational cost. Subquadratic architectures offer a scalable alternative. However, it remains unclear which designs yield the most effective sequence models. We compare three leading approaches: xLSTM, Mamba-2, and Gated DeltaNet. We evaluate these models on tasks with complex dependencies: (1) code-model pre-training, (2) distillation of code models from large language models, and (3) pre-training of time-series foundation models. Across these settings, xLSTM delivers the strongest overall performance. To explain xLSTM's advantage, we present a unified formulation and analyze the underlying architectural mechanisms, focusing on state tracking and memory dynamics. Our results show that xLSTM enables more flexible and stable memory correction via its gating scheme. We corroborate these findings on controlled synthetic length-generalization tasks. Overall, our findings indicate that xLSTM's gains on complex tasks stem from robust state tracking and accumulation.

The Transformer architecture is widely regarded as the most powerful tool for natural language processing, but due to a high number of complex operations, it inherently faces the issue of high energy consumption. To address this issue, we consider Spiking Neural Networks (SNNs), which are an energy-efficient alternative to conventional Artificial Neural Networks (ANNs) due to their naturally event-driven approach to processing information. However, this inherently makes them difficult to train. Often, many SNN-based models circumvent this issue by converting pre-trained ANNs. More recently, attempts have been made to design directly trainable SNN-based adaptations of the Transformer model structure. Although the results showed great promise, the application field was computer vision. Moreover, the proposed model incorporates only encoder blocks. In this paper, we propose SpikeDecoder, a fully SNN-based implementation of the Transformer decoder block, for applications in natural language processing. In a series of experiments, we analyze the impact of exchanging different blocks of the ANN model with spike-based alternatives to identify trade-offs and significant sources of performance loss. We further investigate the role of residual connections and the selection of SNN-compatible normalization techniques. Besides the work on the model architecture, we formulate and compare different embedding methods to project text data into spikes. Finally, we demonstrate that our proposed SNN-based decoder block reduces the theoretical energy consumption by 87% to 93% compared to the ANN baseline.

Recent advances in Vision Transformers (ViTs) for face recognition (FR) have moved beyond the standard CLS-token paradigm. In this paradigm, a special classification token (CLS) is prepended to the patch embeddings and used as a representation of the input for downstream tasks. An alternative approach, Concatenated Patch Embeddings (CPE), instead leverages all patch tokens by concatenating them into a single vector, which is then projected into a compact face representation. CPE has been shown to improve recognition performance in comparison to CLS-based ones, but our qualitative analysis of attention maps showed the presence of artifacts that limit their interpretability. To address this issue, we incorporate register tokens, learnable tokens concatenated to the initial patch embeddings, and processed jointly through the ViT encoder blocks. This mechanism has been shown to produce more structured and interpretable attention maps compared to baseline ViT. We empirically demonstrate that these artifacts consistently appear across various ViT backbones, including small and large models, and that introducing register tokens effectively mitigates them. Adding four or eight registers significantly enhances interpretability, with eight registers providing the highest verification accuracies and smoothest attention structures. Our resulting model, ViT-8R, corresponds to a CPE-based ViT-B architecture augmented with eight register tokens achieves state-of-the-art performance among ViT-based FR models on large-scale IJB-B and IJB-C benchmarks. Also, ViT-8R produces substantially clearer attention maps compared with the baseline model, which offer deeper insight into the model's attention behavior (https://github.com/TaharChettaoui/ViT-FR-Registers)

The rapid progress of reasoning and agentic large language models (LLMs) has increased the demand for long-context inference, but self-attention (SA) scales quadratically with context length. To address this, we study SWARR (Sliding-Window Attention with Reinforced Adaptation for Math Reasoning), a practical recipe for adapting SWA models to mathematical reasoning. SWARR has two stages: (1) efficient conversion from a pretrained SA model to SWA with supervised fine-tuning (SFT), which avoids pretraining a new base model, and (2) policy adaptation with reinforcement learning (RL). We find that SWA still underperforms SA after SFT, and we hypothesize that this gap is caused in part by a data-architecture mismatch: most SFT data are prepared for SA models and may contain long-range dependencies that are difficult for SWA to model. Because on-policy RL optimizes self-generated trajectories under the SWA constraint, it can adapt trajectories to better match SWA. Experiments on mathematical reasoning benchmarks show that this recipe substantially narrows the gap between SWA and SA, recovering much of the accuracy lost during SWA conversion while preserving the efficiency benefits of linear-complexity attention. Our central contribution is the empirical finding that RL changes the conclusion one would draw from conversion and SFT alone about SWA's viability for math reasoning.

Vision sensors provide a lightweight solution for spacecraft proximity operations, but monocular spacecraft 6D pose estimation remains difficult under illumination variation, specular reflection, shadowing, weak texture, and background interference. These factors make local visual evidence spatially unreliable and can destabilize pose regression. This article proposes a Precision-Aware Illumination-Disentangled Vision Transformer (PAID-ViT) for robust spacecraft pose estimation.The proposed model separates pose-relevant structure tokens from illumination-sensitive appearance tokens, estimates patch reliability before pose aggregation, and uses foreground mask supervision to preserve silhouette cues. A parameter-free geometric recovery module converts normalized crop coordinates, log-depth, and a continuous 6D rotation representation into camera-frame rotation and translation. Experiments on SPEED+ V2, the SPEED+ validation/lightbox/sunlamp evaluation configuration used in this study, suggest that PAID-ViT reduces translation error and improves robustness in the challenging sunlamp domain, while ablation studies support the complementary roles of illumination disentanglement, reliability-aware token aggregation, mask supervision, and training-side regularization.

Robotic table tennis is a representative benchmark for high-speed, closed-loop robotic control in dynamic environments, where accurate and fast prediction of ball states is critical for reliable planning and control. Physics-based approaches rely heavily on accurate parameter identification and precise initial state, while learning-based methods often struggle to capture long-range temporal dependencies and are typically trained on limited or simulated data. We propose a transformer-based framework for table tennis ball state prediction that leverages attention mechanisms to model long-range temporal correlations directly from historical observations, without relying on explicit flight or bounce models. To support robust learning and generalization, we collected a large-scale real-world dataset from players of varying skill levels and diverse ball cannon configurations. The combination of a high-capacity transformer architecture and extensive real-world data enables accurate long-horizon forecasting. Building on this capability, we introduce a plug-and-play sim-to-real transfer strategy, Swap Predictor at Deployment (SPAD), which replaces the physics-based simulator used during training with the proposed real-world-trained predictor at deployment, improving the sim-to-real transferability of the policy without requiring retraining. We demonstrate that this simple substitution effectively narrows the sim-to-real gap while preserving the efficiency and scalability of simulation-based training.

We introduce SPADE (SPlit And Delay Embeddings), an autoregressive transformer for sequences whose tokens carry multiple features. Rather than embedding these features jointly, SPADE embeds them independently. Delaying each feature stream relative to the previous one allows intra-token correlations to be learned by the standard self-attention mechanism. Applied to point-cloud calorimeter shower generation in the highly granular ILD detector, SPADE is competitive with the state of the art AllShowers model on photon showers, and substantially outperforms its VQ-VAE-based predecessor OmniJet-$α_C$. The mechanism is applicable to any generative task with multi-feature tokens, enabling LLM-style pretraining workflows for higher-dimensional data.

Recent deep learning approaches for network intrusion detection increasingly incorporate temporal architectures such as recurrent networks and Transformers, often reporting near-perfect performance on CIC-IDS2017. However, many existing studies neither supply their temporal modules with genuine sequence inputs nor evaluate under realistic, leakage-free conditions, making it unclear whether reported gains arise from true sequence-modeling capability. In this work, we reformulate CIC-IDS2017 as a temporal intrusion-detection task by constructing ordered flow sequences from network conversations and benchmarking nine classical and deep learning architectures under a random split, two leakage-free splits, and a padding-scheme ablation. The central finding is that padding convention, not architecture, determines the Transformer's performance: on genuinely sequential (non-padded) windows the Transformer achieves the highest macro-F1 of any model in the experiment (0.89); under zero-pad+mask evaluation it drops markedly (-0.24 macro-F1), while LSTM, GRU, and 1D-CNN remain stable. Under leakage-free group evaluation the Random Forest is the most robust model (+0.009), while the Transformer's false-alarm rate grows from 0.04% to 2.7%, a 67-fold increase invisible under conventional protocols. These findings demonstrate that evaluation methodology -- specifically padding convention and split protocol -- has a larger effect on reported performance than architectural choice, and that widely used random splits with repeat-last padding can overestimate model robustness by up to 0.24 macro-F1. We advocate leakage-free splits, explicit padding disclosure, and sequence-aware benchmarking as standard practice in future IDS research. Code and implementation details are available at https://github.com/zachmocz/temporal-ids-bench.

Image steganalysis, the detection of hidden information embedded in digital images, is a core component of modern cybersecurity and digital forensics. Recent residual Transformer architectures, such as the Pixel-Difference-Convolution and Enhanced-Transformer-Network (PENet) [1], achieve strong detection accuracy, but their computational and memory demands hinder deployment in resource-constrained settings. We present PENet+, a lightweight steganalysis framework that preserves PENet's discriminative structure while substantially improving efficiency. Rather than redesigning or compressing the attention blocks, we retain PENet's self-attention topology for reproducibility and add a classifier-streamlining stage that progressively narrows the SPP-to-FC1 input channels (SPP: spatial pyramid pooling; FC1: first fully connected layer), yielding large reductions in parameters and FLOPs with negligible accuracy loss. We further refine the high-pass-filter (HPF) stem with an activation-aware mechanism that aggregates HPF responses early and selects a balanced SRM-Gabor top-K subset, and we replace PENet's backbone with a MobileNetV2-style inverted residual network. A balanced configuration with K=31 filters (16 Gabor + 15 SRM) matches or surpasses heavier settings at lower compute. Finally, we motivate PReLU from a steganalysis standpoint, arguing that preserving negative responses helps capture weak stego cues that ReLU suppresses. On a disjoint ALASKA2 JPEG QF90 protocol at 512x512 resolution (5,000 cover images for training, validation, and internal testing; a separate 19,000-cover evaluation set), PENet+ achieves up to 45.5% fewer parameters and about 97% fewer FLOPs than the re-evaluated PENet baseline, offering a computationally efficient direction for resource-constrained steganalysis. Device-level latency and power measurements remain future work.

Electroencephalography (EEG) is a widely adopted technique for monitoring brain activity, offering valuable insights into neurological states due to its high temporal resolution and cost-effectiveness. To enhance the analysis of complex EEG data, we propose EEG-TransNet, an architecture designed to capture temporal, regional, and synchronous features of EEG signals. EEG-TransNet introduces three key modules: 1) a preprocessing and feature extraction module leveraging ResNet and wavelet-based denoising, 2) a Local Self-Attention Block for regional feature learning, and 3) a Fuzzy-Attention Synchronous Transformer (FAST) to model spatiotemporal dependencies. Through extensive experiments on three EEG datasets (BETA, SEED, and DepEEG), the proposed model consistently outperforms other methods in terms of classification accuracy and robustness across varying signal lengths. Ablation studies confirm the contribution of the Local Self-Attention Block in improving performance, and the inclusion of depthwise separable convolutions in the decoder reduces computational complexity while maintaining high accuracy. EEG-TransNet's ability to generalize across subjects with minimal performance variation highlights its potential as a robust tool for EEG-based brain activity classification and emotion recognition tasks.

The scalability of Large Language Models (LLMs) to long contexts is fundamentally constrained by the quadratic complexity of standard attention, motivating the adoption of linear attention mechanisms with sub-quadratic cost. To improve representation capacity under long contexts, recent approaches organize memory in a multi-state manner. However, existing multi-state linear attention methods rely on fixed state merging policies that cannot adapt to dynamically varying token importance, irreversibly obscuring critical tokens and causing severe error accumulation over long sequences. To address this limitation, we propose DLA, a dynamic memory modeling framework for multi-state linear attention. DLA introduces (i) Information-Aware Dynamic State Merging, which adaptively determines state boundaries based on token-level information variation, preserving high-resolution representations around semantic transitions while aggressively summarizing stable regions, and (ii) Capacity-Bounded Memory Modeling, which maintains a fixed-size, chronologically ordered state cache by selectively merging adjacent low-information states to control memory growth with minimal information loss. We pre-train DLA on two different linear attention models and evaluate on 16 datasets across three categories. Experimental results demonstrate the superiority of DLA over state-of-the-art.

The recent receptance weighted key value (RWKV) model combines RNN-style recurrence, offering a linear-complexity alternative to Transformers' quadratic self-attention for modeling global dependencies. However, when directly applied to point clouds, RWKV, originally developed for sequential text, struggles to capture local geometric structures and model spatial dependencies effectively. To address this, we propose the \textbf{P-RWKV} block, which bridges the gap between sequence modeling and irregular 3D geometry while preserving the efficiency advantages of RWKV. It consists of a Local Perception Expansion (LPE) component to expand contextual perception along the spatio-temporal sequence and a Spatial Context Enhancement (SCE) component to strengthen spatial awareness. To validate the effectiveness of P-RWKV for point cloud understanding, we construct PointER, a single-modality self-supervised representation learning framework whose encoder is composed of stacked P-RWKV blocks. Furthermore, we extend P-RWKV to a cross-modality setting and integrate the proposed core sub-modules into multiple architectures, demonstrating strong plug-and-play flexibility and architectural generality. Extensive experiments show that the P-RWKV block and its key sub-modules achieve competitive performance across various tasks with lower computational cost and inference latency. Code will be released upon acceptance.

Snapshot Broadband Filter Array (BFA) imaging provides high light throughput for spectral reconstruction but introduces severe spectral aliasing due to complex modulation. Current deep learning approaches, limited to spatial denoising, often fail to address the global frequency-specific degradations caused by the mask structure. To address this, we propose a Physics-embedded Frequency-aware Transformer (PF-Trans) for high-fidelity remote sensing spectral reconstruction. Our method explicitly integrates the physical sensing model through mask injection and a gray-scale consistency loss to ensure physical fidelity. Furthermore, we introduce a Dual-domain Block with a parallel Fast Fourier Transform (FFT) branch, enabling the network to perceive and suppress aliasing artifacts in the frequency domain. Extensive experiments on multiple datasets demonstrate that PF-Trans achieves state-of-the-art performance, achieving a Peak Signal-to-Noise Ratio (PSNR) of up to 48.50 dB on the GF-5 Shanghai dataset, significantly outperforming comparison methods.

We present ViP-VL, an efficient Vietnamese Self-supervised speech Pretraining model leveraging Vector-quantization Learning. To bridge the gap between high-resolution audio and efficient processing, ViP-VL incorporates Acoustic Stacking and Receptive Field Alignment to enable a synchronized 8x subsampling rate within the ChunkFormer architecture, while further enhancing representation robustness through a specialized Mask Selection Strategy during pretraining on the BEST-RQ framework. Pretrained on 17,000 hours of unlabeled Vietnamese speech, our model establishes new state-of-the-art results across four major downstream tasks: Automatic Speech Recognition, Speech Emotion Recognition, Dialect Classification, and Speaker Verification. To facilitate future research and the development of high-performance Vietnamese speech technologies, we publicly release our pretrained weights and implementation at github.com/khanld/chunkformer.

Offsite conversion rate (OCVR) prediction is an important ranking problem in computational recommendation systems. This task presents a modeling challenge: click signals are abundant and exhibit short temporal horizons, whereas conversion signals are inherently sparse, long-delayed, and frequently unattributed. Despite these statistical disparities, both signal types must inform models that operate within strict serving-latency constraints. Prior pre-training approaches address this heterogeneity with a single, undifferentiated encoder applied uniformly across both data streams. We propose DUET (Dual User Embedding Transformers), a framework that explicitly partitions user behavioral data into two domain-coherent streams -- clicks and conversions -- and pre-trains dedicated transformer encoders with architectures tailored to each stream's statistical characteristics: multi-layer self-attention for the dense click stream and interleaved cross- and self-attention for the sparse conversion stream. The resulting complementary embeddings are jointly consumed by a downstream ranker without exceeding serving-latency budgets. Evaluation demonstrates up to 0.38% normalized entropy (NE) reduction relative to the strongest baseline, and A/B test shows consistent improvements in OCVR prediction accuracy.

Modern Diffusion Transformers for video generation provide limited control over the progression of time and the editing of temporal dynamics. We propose a temporal-control methodology that extends a pretrained DiT with explicit time editing, allowing control over motion speed and temporal structure without redesigning the backbone. Its core implementation augments the pretrained model with a lightweight temporal module, preserving the original generative prior while expanding its controllable dynamic range.

We tightly characterize the VC dimension of depth-$L$ Transformers with a total of $W$ parameters, mapping an input sequence of length $T$ to a single output, establishing an upper bound of $O(L W \log (T W))$ and a nearly matching lower bound of $Ω(L W \log (T W / L))$. We further tightly characterize the sample complexity of chain-of-thought learning using such a Transformer, showing teacher forcing (i.e. selecting a predictor consistent with the entire chain-of-thought on training data) learns with sample complexity $O\left(L W \log \left(\left(T+T^{\prime}\right) W\right)\right)$ and that any learning rule that uses chain-of-thought data requires at least $Ω\left(L W \log \left(\left(T+T^{\prime}\right) W / L\right)\right)$ examples, where $T$ is the input length and $T^{\prime}$ is the number of autoregressive steps.

Motivation: Transformer-based models are increasingly applied to large-scale single-cell transcriptomics, showing strong performance through self-supervised learning on millions of cells. However, most existing approaches treat genes as independent features, and largely ignore prior biological knowledge, which limits interpretability and robustness. In this paper, we explore whether explicitly incorporating gene regulatory information can improve both model performance and biological insight. Results: We present scTransformer, the first Transformer-based approach that builds a priori knowledge of biological mechanisms into the model's attention patterns. By constraining information flow according to known regulatory structures, the model learns representations that are more biologically meaningful. We evaluate scTransformer on a disease-relevant single-nucleus RNA-seq dataset using supervised cell-type classification. Compared to standard Transformers, our approach improves classification accuracy, enhances separation of cell types in embedding space, and produces attention patterns consistent with known regulatory programs. Overall, our results demonstrate that embedding biological structure into Transformer models can enhance interpretability without sacrificing performance, offering a principled step toward biologically grounded foundation models for single-cell omics.

Existing sparse attention and KV cache compression methods for long-context LLM inference typically apply fixed sparsity patterns or uniform budgets across all attention heads, overlooking the substantial variation in attention behavior among heads and contexts. We observe two distinct entropy patterns among attention heads: Rigid Heads, whose entropy stays near zero across input segments, and Dynamic Heads, whose entropy fluctuates significantly. Crucially, the distribution of these types is context-dependent and cannot be predetermined offline. We therefore propose EntropyInfer, a training-free framework that uses attention entropy to adaptively allocate compute at the granularity of individual heads and segments during prefilling. For decoding, we introduce a latent KV cache compression scheme that leverages generated output tokens, rather than prefill tokens alone, to identify and retain the most critical cache entries. Extensive experiments on Llama, Qwen and openPangu model series show that EntropyInfer consistently outperforms baselines including SnapKV, AdaKV, and CritiPrefill, achieving up to 2.39$\times$ end-to-end speedup beyond 100k tokens with minimal quality degradation compared to full attention. The code is released in https://github.com/SHA-4096/EntropyInfer.

Understanding how transformer representations evolve across layers, not merely what they encode, remains an open problem in mechanistic interpretability. We recast the transformer forward pass as a discrete population trajectory through a high-dimensional representation manifold, drawing on geometric tools from computational neuroscience. Rather than probing for pre-specified features, we characterize trajectory geometry using five metrics computed directly in the ambient space: trajectory length, curvature, a semantic convergence index, layerwise cosine similarity, and representational stability. Across three model families (GPT-2, TinyLlama, Qwen2.5) and five controlled prompt families, we report four findings. First, semantically related prompts converge significantly in middle-to-late layers (peak CI 0.41--0.58, p<0.001, Mann-Whitney U), consistent with attractor-like dynamics. Second, reasoning tasks produce trajectories of greater curvature than lexical variations (0.71--0.83 rad vs. 0.27--0.31 rad), suggesting curvature encodes computational complexity. Third, ambiguous tokens exhibit trajectory bifurcation with up to 5.6x representational separation by the final layer, absent in unambiguous controls. Fourth, layerwise cosine similarity reveals a universal three-phase structure: encoding, elaboration, and output preparation, consistent across all three architectures. All four effects vanish under shuffled-layer and random-embedding controls. We release a fully open-source, model-agnostic pipeline and argue that trajectory geometry constitutes a principled, probe-free lens for mechanistic interpretability.

During the training of large Transformer models, attention masks regulate the scope and direction of information flow across a sequence. Numerous mask variants exist, and operators such as FlexAttention already support arbitrary attention masks. Nevertheless, a systematic formal analysis of the information-flow structure induced by arbitrary masks has been missing. This paper develops a complete theoretical framework. We prove that, with sufficient depth, the information flow of a multi-layer Transformer converges to a Hasse diagram -- a directed acyclic graph representing a partial order. Building on this, we recast the design of parallel training tasks as the problem of finding a minimal common supergraph of Hasse diagrams, and we establish a criterion for the minimal common supergraph. This yields a constructive method to derive attention masks directly from a family of tasks. Applying the framework, we design two novel masks: a block-generation attention mask that ensures training-inference consistency (Block Two-Stream Attention), and a fully supervised bidirectional attention mask (Butterfly Attention). These results demonstrate the framework's capacity to discover new structures.

Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving. In this report, we propose Lookahead Sparse Attention (LSA), a novel inference paradigm powered by a Neural Memory Indexer built upon the DeepSeek-V4 architecture. Rather than passively attending to all historical tokens, LSA proactively predicts future context demands and preserves only the query-critical KV chunks in the GPU memory. Crucially, we instantiate this architecture via a backbone-free decoupled training strategy. By formulating the indexer as a standard dual-encoder architecture, we train it independently using standard retrieval training frameworks without ever loading the massive backbone model into GPU memory. We demonstrate that this "less is more" paradigm significantly maximizes serving efficiency while acting as an effective attention denoiser in tasks that rely on long-term global memory. Across primary long-context evaluation suites (e.g., LongBench-v2, LongMemEval, and RULER), FM-DS-V4 compresses the average physical KV cache footprint down to merely 13.5% of the full-context baseline, while consistently preserving or slightly elevating downstream accuracy (+0.6% absolute margin on average). Crucially, at extreme 500K scales, FlashMemory suppresses the physical KV cache overhead by over 90% without destabilizing the backbone's core reasoning capacities.

Fundamental investigations into how different molecular encoding methods affect molecular property prediction remain relatively limited. In this study, we extensively examined the optimal molecular encoding methods for molecular properties prediction using two prevalent structure designs: a classical neural network model (MLP) and a Transformer encoder-based model (MLP+TL). For molecular encoding methods, we investigated several types of fingerprints, including traditional topological fingerprints, substructure-based fingerprints, and string-based representations. These two models were trained on seven well-known molecular datasets to evaluate different input molecular encoding methods based on evaluation metrics. On several biologically relevant classification tasks, including toxicity, mutagenicity, and side-effect prediction, our models consistently achieved average AUC values above 0.9. Rather than relying on external post-hoc explanation methods such as the local interpretable model-agnostic explanation (LIME) or the Deep SHapley Additive exPlanations (SHAP), we leveraged the model's intrinsic attention weights as an internal interpretability signal for identifying potentially important feature. The MLP+TL model using MACCS and PubChem as input can capture chemically interpretable groups that determined the major blood-brain barrier (BBB) permeability and mutagenicity in Salmonella typhimurium. In particular, a comparison between Morphine and Heroin highlighted the role of hydroxyl-related substructures in BBB permeability prediction, which was consistently reflected in the attention weights. Overall, our findings provide practical guidance for selecting effective molecular encoding methods and contribute to the development of interpretable molecular informatics approaches for drug discovery.

Worldwide image geo-localization aims to determine where on Earth a single image was captured. However, visually similar scenes may lie thousands of kilometers apart, so methods that localize primarily by appearance often mistake a distant look-alike for the true location. We attribute this failure to a structural cause: in existing methods, GPS coordinates serve only as training supervision, and the distance relationships among locations never enter the learned representation. To address this, we propose GeoMetric, a retrieval-based framework that encodes GPS coordinates relationally rather than in isolation, injecting the distance structure among locations into both representation learning and inference. GeoMetric comprises three components: (1) a Transformer-based GPS encoder with distance-aware location attention that modulates inter-sample aggregation by great-circle proximity; (2) a trimodal contrastive objective that aligns images, geo-textual descriptions, and GPS embeddings in a unified space; and (3) a retrieval-augmented inference stage that supplies large multimodal models (LMMs) with contrastive candidate context for grounded coordinate reasoning. Extensive experiments on IM2GPS, IM2GPS3k, YFCC4k, and YFCC26k demonstrate that GeoMetric consistently outperforms state-of-the-art methods, improving street-level accuracy (within 1 km) by 1.5%, 0.9%, 6.9%, and 2.5%, respectively. Controlled ablations confirm that the gains originate from the proposed geographic encoding rather than any specific LMM.

Linear attention reformulates sequence modeling as recurrent state evolution, enabling efficient linear-time inference. Under the key-value associative paradigm, existing approaches restrict the role of the query to the readout operation, decoupling it from state evolution. We show that query-conditioned state readout induces a structured value prediction over accumulated memory that complements key-based retrieval. Based on this insight, we propose Q-Delta, a query-aware delta rule that integrates mixed key-query prediction errors into state evolution, enabling jointly corrective dynamics while preserving delta-rule efficiency. We establish stability guarantees for the resulting dynamics and derive a hardware-efficient chunkwise-parallel formulation with a custom Triton implementation. Empirical results demonstrate stable optimization, competitive throughput, and consistent improvements over strong baselines on language modeling and long-context retrieval tasks.

The analysis of internet memes in the Nepali language is complicated by frequent code-mixing and a lack of established baseline resources. While memes inherently combine visual and textual elements, this study focuses on a text-centric approach by extracting embedded text using an OCR layer and modeling it with Transformer-based architectures. We evaluate six distinct models and investigate the comparative effectiveness of Hard and Soft Voting ensemble strategies across two tasks: binary hate speech detection and three-class sentiment analysis. Experimental results show that a standalone decoder-only model achieved the highest performance for binary classification, whereas the Soft Voting ensemble performed best for the multi-class sentiment task, yielding a 15.8% relative improvement in Macro F1-score over the strongest standalone baseline. These findings suggest that ensemble strategies behave differently across binary and multi-class tasks, highlighting the importance of selecting aggregation methods suited to the classification objective.

Transformers consistently fail to learn certain simple functions that are provably expressible with specific parameter settings. This gap between learnability and expressivity is particularly prominent for sensitive functions -- functions whose output is likely to change if a single bit of the input is flipped -- for example, PARITY. While prior work has established that transformers exhibit a bias toward functions with low average sensitivity, the precise mechanism underlying this bias remains poorly understood. To shed light on this phenomenon, we study the geometry of transformers' parameter space. We show that sensitive functions -- even when representable -- occupy a vanishingly small region that random initialization is very likely to miss. Specifically, we shift the focus from average sensitivity to the full sensitivity profile -- the distribution of sensitivity values across all inputs -- and prove that randomly initialized transformers almost surely compute functions which have low-sensitivity strings. Consequently, any function that lacks such strings is provably unlearnable.

Temporomandibular joint osteoarthritis (TMJ OA) is a prevalent degenerative condition whose osseous changes are often subtle on cone-beam CT (CBCT), making automated detection challenging. We study how well the DINO family of self-supervised vision transformers -- DINOv1, DINOv2, DINOv2+reg, and RAD-DINO (a radiology-pretrained variant) -- transfers to CBCT, asking how much backbone adaptation is needed and of what kind. We propose a simple slice-based pipeline using Vision Transformer (ViT) backbones: axial CBCT slices are encoded per-slice by a frozen or partially adapted ViT and aggregated via attention-based multiple instance learning (MIL) for patient-level binary OA/Normal classification. Through systematic ablation across unfreezing strategies and aggregation designs on a multi-source CBCT dataset, we find that partial unfreezing of the final two transformer blocks is the decisive factor, improving AUC from 0.671 (fully frozen DINOv2) to 0.902. This outperforms DINOv1 (0.867), DINOv2+reg (0.774), and a supervised ImageNet ViT-B/16 baseline (0.843). Our results provide practical guidance for adapting DINO-family foundation models in low-data medical imaging settings, showing that adaptation strategy is a stronger driver of performance than backbone choice alone.

We introduce CHIAR-Former (CHIAroscuro Attention-based tRansFormer), an efficient transformer that routes each token to either DCT spectral mixing (O(d log d), sub-quadratic) or full self-attention (O(n^2 d), quadratic in sequence length n) based on per-token spectral entropy H(x) in [0,1], which measures the frequency-domain complexity of each token embedding x. We make three contributions: (1) we discover routing collapse -- a three-operator system collapses to DCT+Attention, revealing the optimal operator subset; (2) we propose a learned task-level MetaRouter g = sigma(Linear(x-bar)) in [0,1], where x-bar is the batch-mean embedding and g soft-blends spectral and identity paths end-to-end; and (3) we demonstrate 35-40% FLOP reduction at 400M parameters with a 3.93 PPL cost on WikiText-103 (Test PPL 27.51 vs. 23.58). Under mixed-dataset training, CHIAR-Former dramatically outperforms full attention on small corpora, confirming the regularisation value of spectral mixing. The MetaRouter stabilises at g ~ 0.22, indicating that at scale the model reaches a robust compute-quality equilibrium: attention layers absorb representational complexity while spectral preprocessing efficiently anchors low-frequency structure.

Passive long-wave infrared (LWIR) hyperspectral imaging under a standoff geometry depends on atmospheric absorption and emission, as well as reflected radiance, thus making atmospheric compensation essential to get knowledge of a target of interest. Despite its importance, this compensation has been largely overlooked due to its practical and modeling difficulty. In this paper, we present a lightweight set-based deep learning framework that takes multiple radiance measurements, collected at different standoff ranges, as input and jointly estimates transmittance, atmospheric path radiance, and a shared downwelling spectrum. We analyze the learned representation with a sparse autoencoder and observe that several latent features do activate on geographically coherent subsets of the test data despite the absence of location supervision. Experiments on a MODTRAN generated standoff LWIR dataset demonstrate low spectral distortion across all estimated products. The dataset and code is publicly available at: https://factral.co/SAE-LWIR/

In mechanistic interpretability, attention heads are commonly elevated to role claims (e.g., "this head represents addition") when they are necessary for a behavior, encode it linearly, and recover that behavior when restored after ablation. We show this evidence is insufficient: across three 7-8B instruction-tuned models and five computation families, heads passing all three checks routinely fail to transfer the computation when their activations are patched into a different prompt under matched controls. We introduce KID (Knowing / Intent / Doing), a role-assignment lens for attention heads, and pair it with a three-stage pipeline: capability-selective screening (CSS), singular value decomposition (SVD), and activation transduction under matched controls. Our results document a preliminary role taxonomy (including prompt-trajectory stabilizers, answer-side logit-bias heads, and soft computation-pattern carriers) and show that the same-answer control (a transduction target sharing the answer string but not the requested computation) is an underused check that exposes broad state transfer masquerading as semantic specificity.

Lossy text compression reduces data size while preserving core meaning, making it well-suited for summarization, automated analysis, and digital archives. Despite the dominance of transformer-based models in language modeling, integrating context vectors and entropy coding into Sequence-to-Sequence (Seq2Seq) generation remains underexplored. A key challenge lies in identifying the most informative context vectors from encoder output and incorporating entropy coding to enhance storage efficiency while maintaining high-quality outputs, even under noisy text. We introduce TextEconomizer, an encoder-decoder framework paired with a transformer neural network that reduces variable-sized inputs by 50% to 80% without prior knowledge of dataset dimensions. Our model achieves competitive compression ratios via entropy coding while delivering near-perfect text quality, assessed by BLEU, ROUGE, METEOR, and semantic similarity scores. TextEconomizer operates with approximately 153x fewer parameters than comparable models, achieving a 5.39x compression ratio without sacrificing semantic quality. We also evaluate an LSTM-based autoencoder achieving a state-of-the-art 67x compression ratio with 196x fewer parameters, and LLaMAFormer, a modified transformer with 263x fewer parameters than ICAE while maintaining competitive text quality. TextEconomizer significantly surpasses existing transformer-based models in balancing memory efficiency and high-fidelity outputs, marking a breakthrough in lossy compression with optimal space utilization.

As capacitance extraction accuracy of rule-based pattern matching becomes difficult to sustain at advanced nodes, a growing trend emerges to develop deep-learning-based 2D capacitance models. However, existing MLP- and CNN-based methods constrain their input to fixed metal-layer combinations in a specific process node, limiting their usability in practice. Recognizing the inherent similarity between capacitance matrix and the prevailing attention mechanism, we propose AttentionCap, a customized Transformer for capacitance matrix learning, with a Gram representation framework, a physics-aligned symmetric-attention output layer, and a novel normalized Laplacian loss. We also introduce a process-node embedding to enable multi-node learning. Trained on synthetic data, AttentionCap attains 0.67%/3.99% self/coupling-capacitance error on unseen real designs under a multi-layer and multi-node setting, surpassing the CNN-Cap baseline with 4.6$\times$/5.7$\times$ lower self/coupling error and 192$\times$ faster inference speed. A pretrained AttentionCap accurately transfers to an unseen node with only 5K samples and 4K finetuning steps. With sufficient accuracy on unseen real designs and strong transferability to new process nodes, AttentionCap offers highly practical value for modern EDA workflows. Code and data are available at https://github.com/THU-numbda/AttentionCap.

Vision Transformers (ViTs) achieve strong performance but suffer from high computational costs due to quadratic self-attention complexity. Although token reduction techniques such as pruning and merging mitigate this, they typically overlook how representations evolve across network depth. We propose RAPID, a depth-aware token reduction framework that adapts reduction strategies to the layer-wise characteristics of token representations. The primary methodological contribution is a bifurcated strategy: in shallow-to-middle layers, RAPID employs a redundancy-similarity aware pruning metric to eliminate over-represented local patterns. As features transition to global semantic concepts in deeper layers, the framework shifts to an importance-similarity aware merging mechanism. This stage leverages classification (CLS) token attention weights to protect semantically critical tokens while fusing less important but similar neighbors. Empirical validation on ImageNet-1K using ViT and DeiT architectures demonstrates that RAPID establishes a superior accuracy-compression Pareto frontier compared to plug-and-play baselines such as ToMe and ToFu. RAPID is particularly robust in aggressive compression regimes, achieving up to 4.29% higher accuracy than ToMe at extreme reduction rates. Our framework provides a training-free template for optimizing vision models by aligning reduction strategies with hierarchical feature evolution.

MOLOT (Malicious Operational Logic Observation Transformer) is a static malicious-code detection system designed for SAST setup where package metadata, maintainer history, and dynamic execution traces may be unavailable or unreliable. The system represents source code as behavior sequences derived from static call graphs, includes an explanation stage that ranks suspicious behavior activities and maps them back to source-code locations. The approach is evaluated on Python and JavaScript packages from PyPI and npm, compared with opensource detection tools, and validated under product constraints including runtime, memory use, and false-positive rates observed in a real moderation workflow. We also release Open Malicious-Code Bench, a public benchmark for reproducible evaluation of malicious-package detection methods. The results show that static behavior-sequence modeling can provide accurate, explainable, and deployable malicious-code detection for modern DevSecOps workflows.

Sparsity allows scaling model parameters without proportionally increasing computational cost. While mixture of experts (MoE) models are made increasingly sparse, individual experts typically remain large and dense. Here, we demonstrate that further increasing sparsity by shrinking each expert to consist of a single neuron and selecting a tiny fraction of many available neurons can improve compute efficiency and interpretability. Counterintuitively, the key to achieving both is removing the nonlinearity typically applied to the experts, resulting in a network of sparsely gated linear neurons (sgatlin). In an isoflop comparison, we find that replacing all transformer feedforward layers with sgatlin improves perplexity in language models across different compute budgets. At the same time, the sparsity and linearity of the resulting feedforward circuits present new opportunities for model interpretability. In a small-scale case study, we demonstrate that feedforward circuits in sgatlin can be interpreted without having to train additional replacement models. We find that they form semantically structured clusters and are causally implicated in factual recall. Our findings paint a possible path towards compute-efficient and interpretable transformer feedforward layers.

The attention mechanism is the dominant computational bottleneck in modern transformer-based AI. Its standard implementation incurs quadratic memory traffic in the sequence length~$n$, and DRAM accesses cost 100--1000$\times$ more energy than arithmetic operations on contemporary hardware, so any analysis focused solely on FLOP counts fundamentally mischaracterises the bottleneck. We present a Mathematics of Arrays (MoA) reformulation of scaled dot-product attention and its numerically stable softmax, deriving a Denotational Normal Form (DNF) that eliminates all intermediate arrays -- including the implicit transposed-key buffer and every softmax temporary -- by algebraic construction rather than empirical tuning. The DNF achieves $O(n_{dk} + n{_{dv}})$ data movement versus $O(n^2 + n_{dk} + n_{dv})$ for the standard implementation, where $n$ is the sequence length, $dk$ is the key dimensionality and $dv$ the value dimensionality, and is verified numerically against PyTorch at full double-precision floating-point on concrete inputs. Unlike hardware-specific accelerators or empirical tiling schemes such as FlashAttention, MoA simultaneously provides array fusion, shape-transformation correctness, and predictive cost models from a single algebraic framework. Memory minimality is a theorem established before any code is written. A predictive performance model projects $2$--$100\times$ speedup and $2$--$50\times$ energy reduction, with the advantage widening at exascale. The derivation establishes a formally verified pipeline from Python specification through (ONF) Operational Normal Form, and dimension-lifted hardware mapping, providing performance-portable AI kernels of direct relevance to DARPA edge-deployment and DOE exascale priorities.

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

Long-context prefill remains expensive because full/GQA layers still score the historical sequence, even in hybrid models with local, sparse, linear, or recurrent components. We study how much dense attention is needed to preserve task-level behavior under explicit support granularity and top-k budgets. We introduce an attention-mass top-k oracle for existing GQA checkpoints: for each layer and query position, it computes dense attention, selects head-averaged token support, and recomputes attention only on that support. The oracle is a diagnostic reference, not a deployable accelerator, and separates sparse-budget feasibility from indexer error and runtime realization effects. On Qwen-family retrieval-heavy evaluations, the longest per-query oracle rows stay within 1 point of dense, and a Qwen3.5-9B RULER-style sweep from 4K to 100K stays within 0.48 points. Guided by the oracle, we derive a head-collapsed auxiliary indexer trained by KL distillation from dense attention-mass distributions while keeping the backbone frozen. With separately distilled Qwen3.5-0.8B and Qwen3.5-9B indexers, the reported 16K/32K validation macro gaps are +2.04 and +1.13 points, treated as quality preservation rather than improvement; fused selection-block-shared support can introduce a larger realization gap. Preliminary single-card TTFT measurements show distilled-indexer sparse serving speedups of 1.71x for Qwen3.5-0.8B on NPU and 1.93x for Qwen3.5-9B on GPU against its dense FlashAttention-2 baseline. Additional random-init stress rows reach 3.44x, indicating sparse-runtime headroom but not validated output quality. This first release separates oracle feasibility, distilled-indexer quality, and runtime headroom, leaving a fully matched quality-latency frontier to future work.

Pose-guided text-to-image generation often suffers from limb distortions and feature crosstalk in complex multi-person scenarios. While existing UNet-based adapters struggle with long-range spatial dependencies, emerging Multimodal Diffusion Transformers (MM-DiTs) offer superior global modeling. However, naive signal concatenation in MM-DiTs severely disrupts pre-trained latent distributions. To address this, we propose TrioPose, a native pose-driven framework built upon the SD3.5M architecture. Specifically, we introduce a Triple-Stream Pose-Aware DiT (TSPA-DiT) that treats pose as an independent modality. It employs layer-wise activation and zero-initialized dual-residual injection to smoothly enforce geometric constraints while preserving pre-trained latent stability. To resolve severe multi-instance occlusions, we design a Learnable Relational Bias Mask that categorizes topological connectivity into fine-grained physical states, mapping them into continuous attention soft constraints to effectively decouple inter-instance interference. Furthermore, a Pose-Guided Spatial Loss Weighting strategy modulates the native diffusion objective using heatmap-derived error maps, focusing anatomical supervision strictly on distortion-prone regions. Extensive experiments demonstrate that TrioPose achieves state-of-the-art performance across challenging benchmarks, including Human-Art, CrowdPose, and OCHuman. Notably, it attains an AP of $64.33$ on Human-Art, representing a $30\%$ improvement over prior arts, while setting new standards for visual fidelity and text-image semantic alignment in complex multi-human generation.

Three-dimensional models are widely assumed preferable for volumetric medical imaging, yet their practical value depends on whether performance gains justify added computational cost and complexity. Rather than proposing a new architecture, we study how input dimensionality (2D, 2.5D, 3D) affects model behavior across convolutional neural networks (CNNs) and Vision Transformers (ViTs) under a fixed training protocol. Using a leakage-free NLST cohort (n = 1,977) with supporting LIDC-IDRI data, we find that the 2.5D CNN offers the most favorable discrimination-stability trade-off in our comparison (ROC-AUC 0.682, 95% CI [0.546, 0.799]) with a stable operating point. In contrast, 3D CNNs show threshold instability, and transformers exhibit degenerate predictions, such as all-positive predictions. Confidence intervals are wide and overlapping, so we present these results as a controlled resource-performance frontier and a failure-mode taxonomy rather than as definitive superiority claims. For class-imbalanced lung cancer screening classification, 2D and 2.5D inputs provide a more reliable trade-off between performance, stability, and computational efficiency than full 3D representations.