Transformers

We present AttentionBender, a tool that manipulates cross-attention in Video Diffusion Transformers to help artists probe the internal mechanics of black-box video generation. While generative outputs are increasingly realistic, prompt-only control limits artists' ability to build intuition for the model's material process or to work beyond its default tendencies. Using an autobiographical research-through-design approach, we built on Network Bending to design AttentionBender, which applies 2D transforms (rotation, scaling, translation, etc.) to cross-attention maps to modulate generation. We assess AttentionBender by visualizing 4,500+ video generations across prompts, operations, and layer targets. Our results suggest that cross-attention is highly entangled: targeted manipulations often resist clean, localized control, producing distributed distortions and glitch aesthetics over linear edits. AttentionBender contributes a tool that functions both as an Explainable AI style probe of transformer attention mechanisms, and as a creative technique for producing novel aesthetics beyond the model's learned representational space.

Dropout is a widely used regularization technique in deep learning, but its effects are typically realized through stochastic masking rather than explicit optimization objectives. We propose a deterministic formulation that expresses dropout as an additive regularizer directly incorporated into the training loss. The framework derives explicit regularization terms for Transformer architectures, covering attention query, key, value, and feed-forward components with independently controllable strengths. This formulation removes reliance on stochastic perturbations while providing clearer and fine-grained control over regularization strength. Experiments across image classification, temporal action detection, and audio classification show that explicit dropout matches or outperforms conventional implicit methods, with consistent gains when applied to attention and feed-forward network layers. Ablation studies demonstrate stable performance and controllable regularization through regularization coefficients and dropout rates. Overall, explicit dropout offers a practical and interpretable alternative to stochastic regularization while maintaining architectural flexibility across diverse tasks.

Leveraging the natural spatiotemporal energy decay in video diffusion offers a path to efficiency, yet relying solely on rigid static masks risks losing critical long-range information in complex dynamics. To address this issue, we propose \textbf{DynamicRad}, a unified sparse-attention paradigm that grounds adaptive selection within a radial locality prior. DynamicRad introduces a \textbf{dual-mode} strategy: \textit{static-ratio} for speed-optimized execution and \textit{dynamic-threshold} for quality-first filtering. To ensure robustness without online search overhead, we integrate an offline Bayesian Optimization (BO) pipeline coupled with a \textbf{semantic motion router}. This lightweight projection module maps prompt embeddings to optimal sparsity regimes with \textbf{minimal runtime overhead}. Unlike online profiling methods, our offline BO optimizes attention reconstruction error (MSE) on a physics-based proxy task, ensuring rapid convergence. Experiments on HunyuanVideo and Wan2.1-14B demonstrate that DynamicRad pushes the efficiency--quality Pareto frontier, achieving \textbf{1.7$\times$--2.5$\times$ inference speedups} with \textbf{over 80% effective sparsity}. In some long-sequence settings, the dynamic mode even matches or exceeds the dense baseline, while mask-aware LoRA further improves long-horizon coherence. Code is available at https://github.com/Adamlong3/DynamicRad.

The quadratic complexity of softmax attention presents a major obstacle for scaling Transformers to high-resolution vision tasks. Existing linear attention variants often replace the softmax with Gaussian kernels to reduce complexity, but such approximations lack theoretical grounding and tend to oversuppress mid-range token interactions. We propose LaplacianFormer, a Transformer variant that employs a Laplacian kernel as a principled alternative to softmax, motivated by empirical observations and theoretical analysis. To address expressiveness degradation under low-rank approximations, we introduce a provably injective feature map that retains fine-grained token information. For efficient computation, we adopt a Nyström approximation of the kernel matrix and solve the resulting system using Newton--Schulz iteration, avoiding costly matrix inversion and SVD. We further develop custom CUDA implementations for both the kernel and solver, enabling high-throughput forward and backward passes suitable for edge deployment. Experiments on ImageNet show that LaplacianFormer achieves strong performance-efficiency trade-offs while improving attention expressiveness.

Sparse attention can reduce the cost of long-context inference, but most variants introduce new architectural components. We introduce Simplified Sparse Attention (SSA), a simpler approach to sparse attention that requires no architectural changes. Concretely, we first perform continued pretraining on sequences interleaved with gist tokens. We optimize the standard next-token loss as usual, but the gist tokens use an attention mask to restrict what parts of the context the language model can attend to; this teaches the model to pack each chunk's important information into the gist tokens. At inference time, SSA scores chunks via attention between the current query and the small set of gist tokens, selectively unfolding the top-k chunks by reintroducing their corresponding raw tokens. Since the query is scored only against the gist tokens, we avoid the memory-bandwidth cost associated with naive scoring against the full KV cache, without requiring the auxiliary KV cache approach used by sparse attention methods. On LongBench, SSA consistently outperforms compression and inference-time sparse-attention baselines under the same compression ratio. More strikingly, in retrieval-augmented generation, SSA can even outperform full attention after continued pretraining by over 5.7 points. We attribute this to the ability of SSA's selective unfolding, which concentrates attention on the query-relevant chunks and effectively filters out noise. SSA further extends to a hierarchical gist-of-gist variant (H-SSA) that achieves log-linear decoding complexity while maintaining or improving accuracy at high compression ratios up to 32x. The code is available at https://github.com/yuzhenmao/simplified-sparse-attention/.

Vision Transformers (ViTs) have achieved remarkable success across vision tasks, yet recent studies show they remain vulnerable to backdoor attacks. Existing patch-wise attacks typically assume a single fixed trigger location during inference to maximize trigger attention. However, they overlook the self-attention mechanism in ViTs, which captures long-range dependencies across patches. In this work, we observe that a patch-wise trigger can achieve high attack effectiveness when activating backdoors across neighboring patches, a phenomenon we term the Trigger Radiating Effect (TRE). We further find that inter-patch trigger insertion during training can synergistically enhance TRE compared to single-patch insertion. Prior ViT-specific attacks that maximize trigger attention often sacrifice visual and attention stealthiness, making them detectable. Based on these insights, we propose PASTA, a twofold stealthy patch-wise backdoor attack in both pixel and attention domains. PASTA enables backdoor activation when the trigger is placed at arbitrary patches during inference. To achieve this, we introduce a multi-location trigger insertion strategy to enhance TRE. However, preserving stealthiness while maintaining strong TRE is challenging, as TRE is weakened under stealthy constraints. We therefore formulate a bi-level optimization problem and propose an adaptive backdoor learning framework, where the model and trigger iteratively adapt to each other to avoid local optima. Extensive experiments show that PASTA achieves 99.13% attack success rate across arbitrary patches on average, while significantly improving visual and attention stealthiness (144.43x and 18.68x) and robustness (2.79x) against state-of-the-art ViT defenses across four datasets, outperforming CNN- and ViT-based baselines.

For state-of-the-art image understanding, Vision Transformers (ViTs) have become the standard architecture but their processing diverges substantially from human attentional characteristics. We investigate whether this cognitive gap can be shrunk by fine-tuning the self-attention weights of Google's ViT-B/16 on human saliency fixation maps. To isolate the effects of semantically relevant signals from generic human supervision, the tuned model is compared against a shuffled control. Fine-tuning significantly improved alignment across five saliency metrics and induced three hallmark human-like biases: tuning reversed the baseline's anti-human large-object bias toward small-objects, amplified the animacy preference and diminished extreme attention entropy. Bayesian parity analysis provides decisive to very-strong evidence that this cognitive alignment comes at no cost to the model's original classification performance on in- (ImageNet), corrupted (ImageNet-C) and out-of-distribution (ObjectNet) benchmarks. An equivalent procedure applied to a ResNet-50 Convolutional Neural Network (CNN) instead degraded both alignment and accuracy, suggesting that the ViT's modular self-attention mechanism is uniquely suited for dissociating spatial priority from representational logic. These findings demonstrate that biologically grounded priors can be instilled as a free emergent property of human-aligned attention, to improve transformer interpretability.

Despite the remarkable success of Vision Transformers (ViTs) across a wide range of vision tasks, recent studies have revealed that they remain vulnerable to adversarial examples, much like Convolutional Neural Networks (CNNs). A common empirical defense strategy is adversarial training, yet the theoretical underpinnings of its robustness in ViTs remain largely unexplored. In this work, we present the first theoretical analysis of adversarial training under simplified ViT architectures. We show that, when trained under a signal-to-noise ratio that satisfies a certain condition and within a moderate perturbation budget, adversarial training enables ViTs to achieve nearly zero robust training loss and robust generalization error under certain regimes. Remarkably, this leads to strong generalization even in the presence of overfitting, a phenomenon known as \emph{benign overfitting}, previously only observed in CNNs (with adversarial training). Experiments on both synthetic and real-world datasets further validate our theoretical findings.

Reconstructing physically plausible 3D human-scene interactions (HSI) from a single image currently presents a trade-off: optimization based methods offer accurate contact but are slow (~20s), while feed-forward approaches are fast yet lack explicit interaction reasoning, producing floating and interpenetration artifacts. Our key insight is that geometry-based human--scene fitting can be amortized into fast feed-forward inference. We present GRAFT (Geometric Refinement And Fitting Transformer), a learned HSI prior that predicts Interaction Gradients: corrective parameter updates that iteratively refine human meshes by reasoning about their 3D relationship to the surrounding scene. GRAFT encodes the interaction state into compact body-anchored tokens, each grounded in the scene geometry via Geometric Probes that capture spatial relationships with nearby surfaces. A lightweight transformer recurrently updates human meshes and re-probes the scene, ensuring the final pose aligns with both learned priors and observed geometry. GRAFT operates either as an end-to-end reconstructor using image features, or with geometry alone as a transferable plug-and-play HSI prior that improves feed-forward methods without retraining. Experiments show GRAFT improves interaction quality by up to 122% over state-of-the-art feed-forward methods and matches optimization-based interaction quality at ${\sim}100{\times}$ lower runtime, while generalizing seamlessly to in-the-wild multi-person scenes and being preferred in 64.8% of three-way user study. Project page: https://pradyumnaym.github.io/graft .

Transformers have become a common foundation across deep learning, yet 3D scene understanding still relies on specialized backbones with strong domain priors. This keeps the field isolated from the broader Transformer ecosystem, limiting the transfer of new advances as well as the benefits of increasingly optimized software and hardware stacks. To bridge this gap, we adapt the vanilla Transformer encoder to 3D scenes with minimal modifications. Given an input 3D scene, we partition it into volumetric patch tokens, process them with full global self-attention, and inject positional information via a 3D extension of rotary positional embeddings. We call the resulting model the Volume Transformer (Volt) and apply it to 3D semantic segmentation. Naively training Volt on standard 3D benchmarks leads to shortcut learning, highlighting the limited scale of current 3D supervision. To overcome this, we introduce a data-efficient training recipe based on strong 3D augmentations, regularization, and distillation from a convolutional teacher, making Volt competitive with state-of-the-art methods. We then scale supervision through joint training on multiple datasets and show that Volt benefits more from increased scale than domain-specific 3D backbones, achieving state-of-the-art results across indoor and outdoor datasets. Finally, when used as a drop-in backbone in a standard 3D instance segmentation pipeline, Volt again sets a new state of the art, highlighting its potential as a simple, scalable, general-purpose backbone for 3D scene understanding.

Metal-organic frameworks (MOFs) are a major target of machine-learning-based property prediction, yet most models assume that a single framework representation maps to a single property value. This assumption becomes problematic for experimental MOFs, where samples reported as the same framework can exhibit different properties because of differences in crystallinity, phase purity, defects, and other sample-dependent factors. Here we introduce Experimental X-ray Diffraction Integrated Transformer (EXIT), a multimodal transformer for sample-aware prediction of MOF properties that combines MOFid with X-ray diffraction (XRD). In EXIT, MOFid encodes MOF identity, whereas XRD provides complementary information about the experimentally realized sample state. EXIT is pre-trained on one million hypothetical MOFs with simulated XRD to learn transferable representations, leading to improved downstream performance relative to existing approaches. EXIT is fine-tuned on literature-derived experimental datasets for surface area and pore volume prediction. Incorporating experimental XRD improves predictive performance relative to models without experimental XRD, and attention analysis and sample-level case studies further show that EXIT assigns different predictions to samples sharing the same MOF identity when their XRD patterns differ. These results establish a practical step from framework-aware to sample-aware MOF property prediction and highlight the value of incorporating experimental characterization into porous materials informatics.

Multi-view crowd tracking estimates each person's tracking trajectories on the ground of the scene. Recent research works mainly rely on CNNs-based multi-view crowd tracking architectures, and most of them are evaluated and compared on relatively small datasets, such as Wildtrack and MultiviewX. Since these two datasets are collected in small scenes and only contain tens of frames in the evaluation stage, it is difficult for the current methods to be applied to real-world applications where scene size and occlusion are more complicated. In this paper, we propose a Transformer-based multi-view crowd tracking model, \textit{MVTrackTrans}, which adopts interactions between camera views and the ground plane for enhanced multi-view tracking performance. Besides, for better evaluation, we collect and label two large real-world multi-view tracking datasets, MVCrowdTrack and CityTrack, which contain a much larger scene size over a longer time period. Compared with existing methods on the two large and new datasets, the proposed MVTrackTrans model achieves better performance, demonstrating the advantages of the model design in dealing with large scenes. We believe the proposed datasets and model will push the frontiers of the task to more practical scenarios, and the datasets and code are available at: https://github.com/zqyq/MVTrackTrans.

Goal-conditioned navigation models for ground robots trained using supervised learning show promising zero-shot transfer, but their collision-avoidance capability nevertheless degrades under distribution shift, i.e. environmental, robot or sensor configuration changes. We propose ViLiNT a multimodal, attention-based policy for goal navigation, trained on heterogeneous data from multiple platforms and environments, which improves robustness with two key features. First, we fuse RGB images, 3D LiDAR point clouds, a goal embedding and a robot's embodiment descriptor with a transformer architecture to capture complementary geometry and appearance cues. The transformer's output is used to condition a diffusion model that generates navigable trajectories. Second, using automatically generated offline labels, we train a path clearance prediction head for scoring and ranking trajectories produced by the diffusion model. The diffusion conditioning as well as the trajectory ranking head depend on a robot's embodiment token that allows our model to generate and select trajectories with respect to the robot's dimensions. Across three simulated environments, ViLiNT improves Success Rate on average by 166% over equivalent state-of-the-art vision-only baseline (NoMaD). This increase in performance is confirmed through real-world deployments of a rover navigating in obstacle fields. These results highlight that combining multimodal fusion with our collision prediction mechanism leads to improved off-road navigation robustness.

Scaling Transformers typically necessitates training larger models from scratch, as standard architectures struggle to expand without discarding learned representations. We identify the primary bottleneck in the attention mechanism's linear projections, which strictly confine feature extraction to fixed-dimensional subspaces, limiting both expressivity and incremental capacity. To address this, we introduce Nexusformer, which replaces linear $Q/K/V$ projections with a Nexus-Rank layer, a three-stage nonlinear mapping driven by dual activations in progressively higher dimensional spaces. This design overcomes the linearity constraint and enables lossless structured growth: new capacity can be injected along two axes via zero-initialized blocks that preserve pretrained knowledge. Experiments on language modeling and reasoning benchmarks demonstrate that Nexusformer matches Tokenformer's perplexity using up to 41.5% less training compute during progressive scaling (240M to 440M). Furthermore, our analysis of growth dynamics reveals that zero initialization induces a stable convergence trajectory, allowing us to derive a geometric scaling law that accurately predicts performance across expansion scales.

Linear attention mechanisms have emerged as promising alternatives to softmax attention, offering linear-time complexity during inference. Recent advances such as Gated DeltaNet (GDN) and Kimi Delta Attention (KDA) have demonstrated that the delta rule, an online gradient descent update, enables superior associative recall compared to simple additive updates. While KDA refined the coarse head-wise decay gate into channel-wise decay, the learning rate $β_t$ in the delta update remains a scalar, limiting the model's capacity for dimension-specific adaptation. We introduce FG$^2$-GDN, which replaces the scalar $β_t$ with a channel-wise vector analogous to the transition from SGD to per-coordinate adaptive optimizers such as AdaGrad and Adam. We further propose FG$^2$-GDN+, which decouples the scaling for keys and values, enabling independent control of erasure strength and write strength. Experiments on synthetic and real-world benchmarks show that FG$^2$-GDN and its variant improve associative recall and long-context understanding over GDN and KDA, with comparable computational efficiency.

Hyperspectral image (HSI) classification remains challenging due to high spectral dimensionality, redundancy, and limited labeled data. Although convolutional neural networks (CNNs) and Vision Transformers (ViTs) achieve strong performance by exploiting spectral-spatial information and long-range dependencies, they often incur high computational cost and large model size, limiting practical use. To address these limitations, a unified hybrid framework, termed ConvVitMamba, is proposed for efficient HSI classification. The architecture integrates three components: a multiscale convolutional feature extractor to capture local spectral, spatial, and joint patterns; a Vision Transformer based tokenization and encoding stage to model global contextual relationships; and a lightweight Mamba inspired gated sequence mixing module for efficient content-aware refinement without quadratic self-attention. Principal Component Analysis (PCA) is used as preprocessing to reduce redundancy and improve efficiency. Experiments on four benchmark datasets, including Houston and three UAV borne QUH datasets (Pingan, Qingyun, and Tangdaowan), demonstrate that ConvVitMamba consistently outperforms CNN, Transformer, and Mamba based methods while maintaining a favorable balance between accuracy, model size, and inference efficiency. Ablation studies confirm the complementary contributions of all components. The results indicate that the proposed framework provides an effective and efficient solution for HSI classification in diverse scenarios. The source code is publicly available at https://github.com/mqalkhatib/ConvVitMamba

Looped transformers scale computational depth without increasing parameter count by repeatedly applying a shared transformer block and can be used for iterative refinement, where each loop rewrites a full fixed-size prediction in parallel. On difficult problems, such as those that require search-like computation, reaching a highly structured solution starting from noise can require long refinement trajectories. Learning such trajectories is challenging when training specifies only the target solution and provides no supervision over the intermediate refinement path. Diffusion models tackle this issue by corrupting data with varying magnitudes of noise and training the model to reverse it in a \textit{single step}. However, this process misaligns training and testing behaviour. We introduce Denoising Recursion Models, a method that similarly corrupts data with noise but trains the model to reverse the corruption over \textit{multiple} recursive steps. This strategy provides a tractable curriculum of intermediate states, while better aligning training with testing and incentivizing non-greedy, forward-looking generation. Through extensive experiments, we show this approach outperforms the Tiny Recursion Model (TRM) on ARC-AGI, where it recently achieved breakthrough performance.

Large Language Models (LLMs) remain vulnerable to optimization-based jailbreak attacks that exploit internal gradient structure. While Sparse Autoencoders (SAEs) are widely used for interpretability, their robustness implications remain underexplored. We present a study of integrating pretrained SAEs into transformer residual streams at inference time, without modifying model weights or blocking gradients. Across four model families (Gemma, LLaMA, Mistral, Qwen) and two strong white-box attacks (GCG, BEAST) plus three black-box benchmarks, SAE-augmented models achieve up to a 5x reduction in jailbreak success rate relative to the undefended baseline and reduce cross-model attack transferability. Parametric ablations reveal (i) a monotonic dose-response relationship between L0 sparsity and attack success rate, and (ii) a layer-dependent defense-utility tradeoff, where intermediate layers balance robustness and clean performance. These findings are consistent with a representational bottleneck hypothesis: sparse projection reshapes the optimization geometry exploited by jailbreak attacks.

Transformer-based embedding models suffer from quadratic computational and linear memory complexity, limiting their utility for long sequences. We propose recurrent architectures as an efficient alternative, introducing a vertically chunked inference strategy that enables fast embedding generation with memory usage that becomes constant in the input length once it exceeds the vertical chunk size. By fine-tuning Mamba2 models, we demonstrate their viability as general-purpose text embedders, achieving competitive performance across a range of benchmarks while maintaining a substantially smaller memory footprint compared to transformer-based counterparts. We empirically validate the applicability of our inference strategy to Mamba2, RWKV, and xLSTM models, confirming consistent runtime-memory trade-offs across architectures and establishing recurrent models as a compelling alternative to transformers for efficient embedding generation.

We study post-training W4A4 quantization in a controlled 300M-parameter SwiGLU decoder-only language model trained on 5B tokens of FineWeb-Edu, and ask which input-activation sites dominate the error. Naive round-to-nearest W4A4 collapses validation perplexity from FP16 23.6 to 1727. A simple residual-axis training-time intervention -- Depth Registers with a register-magnitude hinge loss (DR+sink) -- reduces this to 119 (about 14x) at matched FP16 PPL and matched zero-shot capacity, and composes with SmoothQuant to 39.9 PPL. The residual ~2 PPL gap to FP16 is the diagnostic core. We decompose W4A4 damage by input-activation site: the five trainable linears in a SwiGLU block split into residual-axis readers (qkv, w1, w3) and block-internal generators (o_proj, w2). Elementary norm arguments show residual-axis magnitude control bounds readers tightly but leaves w2's bilinear input bounded only by the trivial product of factor bounds; empirically, DR+sink collapses reader kurtosis while leaving generators essentially unchanged, and the reader-rescued W4A4 residue is flat at ~0.28 nats across three matched checkpoints with Delta-remove(w2) dominating. We present DR+sink as a training-time probe rather than a deployment proposal: a post-hoc alternative (Per-Linear QuaRot) nearly matches it on the reader axis. Full QuaRot -- adding online per-head value Hadamard plus online w2-input rotation -- does not close the gap either, directly testing the prediction that orthogonal rotation cannot bound the bilinear SwiGLU tail. Claims are specific to our 300M, 5B-token, single-seed setting, and our experiments do not isolate the partition from the hinge.

Vision Transformers (ViTs) have become a dominant architecture in computer vision, yet their prediction process remains difficult to interpret because information is propagated through complex interactions across layers and attention heads. Existing attention based explanation methods provide an intuitive way to trace information flow. However, they rely mainly on raw attention weights, which do not explicitly reflect the final decision and often lead to explanations with limited class discriminability. In contrast, gradient based localization methods are more effective at highlighting class specific evidence, but they do not fully exploit the hierarchical attention propagation mechanism of transformers. To address this limitation, we propose Decision-Aware Attention Propagation (DAP), an attribution method that injects decision-relevant priors into transformer attention propagation. By estimating token importance through gradient based localization and integrating it into layer wise attention rollout, the method captures both the structural flow of attention and the evidence most relevant to the final prediction. Consequently, DAP produces attribution maps that are more class sensitive, compact, and faithful than those generated by conventional attention based methods. Extensive experiments across Vision Transformer variants of different model scales show that DAP consistently outperforms existing baselines in both quantitative metrics and qualitative visualizations, indicating that decision aware propagation is an effective direction for improving ViT interpretability.

The key-value (KV) cache is the dominant memory bottleneck during Transformer inference, yet little is known theoretically about how aggressively it can be compressed before multi-step reasoning degrades. We study this through $k$-hop pointer chasing on $n$ tokens under a shared KV cache of size $s$, attention dimension $m$, $H$ heads, $p$-bit precision, and a locality-respecting cache controller (satisfied by all standard KV-compression methods). We give three results. (1) Product depth lower bound (conjectured). We conjecture that any such Transformer ($n \geq 4k$, $s \leq \sqrt{n}/4$) requires depth $L = Ω(\lceil k/s \rceil \cdot \lceil \log_2 n/(Hmp) \rceil)$, and isolate the sole remaining gap as a probabilistic step on the joint distribution of cache trace and pointer chain. Unconditionally, we prove a matching upper bound $L = O(\min(k, \lceil k/s \rceil \log s) \cdot \log n/(mp))$ via windowed pointer doubling, and a max-bound $L = Ω(\max(\lceil k/s \rceil, \log n/(Hmp)))$. Closing the conjecture amounts to upgrading max to product. (2) Bandwidth barrier. The product bound binds only when $Hmp \lesssim \log n$. Any lower bound provable via per-window distinguishability counting -- including reachability, bandwidth, and combinations -- cannot exceed $\lceil k/s \rceil$ once $Hmp \geq \log_2 n$. Breaking this requires lifting unconditional communication-complexity bounds for pointer chasing to Cache-Transformer depth. (3) Adaptive vs oblivious error scaling. Under random cache over $T = \lceil \log_2 k \rceil$ doubling stages, oblivious caches give $\Pr[\mathcal{E}] \leq (s/(n-T))^T + 2T^3/n$ (exponential in $T$), while adaptive locality-respecting caches achieve $\Pr[\mathcal{E}] = s/n$ exactly, independent of $T$. The $Ω((n/s)^{T-1})$ separation explains why heavy-hitter eviction empirically dominates random eviction for multi-hop reasoning.

Vision-Language Models(VLMs) excel at autoregressive text generation, yet end-to-end autonomous driving requires multi-task learning with structured outputs and heterogeneous decoding behaviors, such as autoregressive language generation, parallel object detection and trajectory regression. To accommodate these differences, existing systems typically introduce separate or cascaded decoders, resulting in architectural fragmentation and limited backbone reuse. In this work, we present a unified autonomous driving framework built upon a pretrained VLM, where heterogeneous decoding behaviors are reconciled within a single transformer decoder. We demonstrate that pretrained VLM attention exhibits strong transferability beyond pure language modeling. By organizing visual and structured query tokens within a single causal decoder, structured queries can naturally condition on visual context through the original attention mechanism. Textual and structured outputs share a common attention backbone, enabling stable joint optimization across heterogeneous tasks. Trajectory planning is realized within the same causal LLM decoder by introducing structured trajectory queries. This unified formulation enables planning to share the pretrained attention backbone with images and perception tokens. Extensive experiments on end-to-end autonomous driving benchmarks demonstrate state-of-the-art performance, including 0.28 L2 and 0.18 collision rate on nuScenes open-loop evaluation and competitive results (86.8 PDMS) on NAVSIM closed-loop evaluation. The full model preserves multi-modal generation capability, while an efficient inference mode achieves approximately 40% lower latency. Code and models are available at https://github.com/Z1zyw/OneDrive

While advanced foundation models like ModernBERT significantly outperform older architectures in dense retrieval, they surprisingly lag behind the aging BERT-base baseline in learned sparse retrieval (LSR). We identify the root cause as the \textit{Vocabulary Gap}: modern tokenizers utilize raw, case-sensitive vocabularies designed for lossless reconstruction, which map single semantic units to redundant surface forms, wasting model capacity on morphological noise and hindering lexical matching. We formalize this intuition through a theoretical framework, demonstrating that appropriate vocabulary coarse-graining can tighten the generalization bounds by reducing complexity of the hypothesis class, provided that semantic integrity is preserved. To resolve this, we propose \textbf{Vocabulary Transfer (VT)}, a model-agnostic framework that migrates advanced encoders to sparse-friendly, normalized vocabularies with minimal computational cost. VT utilizes a novel \textbf{Semantic Initialization} via spatial topology to preserve geometric structure and an \textbf{Activation Potential Calibration (APC)} mechanism to align pre-trained manifolds with sparsity constraints, preventing the dead neuron and dense collapse observed in standard fine-tuning. Empirically, VT is universally effective: it enables ModernBERT to achieve state-of-the-art performance on the BEIR benchmark (\textbf{52.4} nDCG, a \textbf{+4.7} improvement), resuscitates failing models like RoBERTa-large, and generalizes seamlessly to inference-free architectures and specialized domains. These results confirm that the performance lag is not an architectural deficiency but a solvable vocabulary mismatch. We've released our code and models.\footnote{https://anonymous.4open.science/r/vocab-transfer/. All details included.}

In today's day and age, we face a challenge in detecting deepfake images because of the fast evolution of modern generative models and the poor generalization capability of existing methods. In this paper, we use an ensemble of fine-tuned vision transformers like DINOv2, AIMv2 and OpenCLIP's ViT-L/14 to create generalizable method to detect deepfakes. We use the DF-Wild dataset released as part of the IEEE SP Cup 2025, because it uses a challenging and diverse set of manipulations and generation techniques. We started our experiments with CNN classifiers trained on spatial features. Experimental results show that our ensemble outperforms individual models and strong CNN baselines, achieving an AUC of 96.77% and an Equal Error Rate (EER) of just 9% on the DF-Wild test set, beating the state-of-the-art deepfake detection algorithm Effort by 7.05% and 8% in AUC and EER respectively. This was the winning solution for SP Cup, presented at ICASSP 2025.

Global self-attention drives modern graph transformers, yet the softmax at its core imposes a structural constraint rarely examined directly: every attention row is non-negative and sums to one, so each per-head output is a mass-conserving convex combination of value vectors. A node can never "attend to nothing." We argue this conservation constraint is a single root cause behind three pathologies usually studied in isolation: the collapse of node representations with depth (over-smoothing), a low-rank bottleneck on per-head outputs, and brittle optimization in deep stacks. Drawing on how sigmoid gating removes analogous attention sinks in language models, we introduce SigGate-GT, a graph transformer that applies a learned, per-head, input-conditioned sigmoid gate to the attention output inside the GraphGPS framework. The gate is a smooth, per-dimension "volume control" that can drive head outputs toward zero, relaxing the constraint without abandoning attention's probabilistic interpretation. Analytically and through synthetic experiments, we show the gate strictly increases the stable rank of per-head outputs, and connect this rank gain to all three manifestations. On five molecular and long-range benchmarks, SigGate-GT matches the prior best on ZINC (0.059 MAE), records the strongest result among the graph-transformer baselines we evaluate on ogbg-molhiv (82.47% ROC-AUC), and is competitive on ogbg-molpcba and the Long-Range Graph Benchmark, with statistically significant gains over GraphGPS on all five datasets (p < 0.05). Mechanism analyses confirm the diagnosis: gating slows over-smoothing (a 30% mean relative gain in representation diversity across 4-16 layers), keeps attention entropy from collapsing, and stabilizes training across a 10x learning-rate range, at about 1% parameter overhead on OGB and under 3% wall-clock cost.

Rowhammer on GPU DRAM has enabled adversarial bit flips in model weights; shared KV-cache blocks in LLM serving systems present an analogous but previously unexamined target. In vLLM's Prefix Caching, these blocks exist as a single physical copy without integrity protection. Using software fault injection under ideal bit targeting, we characterize worst-case severity and identify three properties: (1) Silent divergence - 13 of 16 BF16 bit positions produce coherent but altered outputs, indistinguishable from legitimate responses without a clean baseline. (2) Selective propagation - only requests sharing the targeted prefix are affected. (3) Persistent accumulation - no temporal decay occurs, so cumulative damage grows linearly with subsequent requests. Together, these constitute a threat profile distinct from weight corruption: silent divergence and selective propagation enable detection evasion; persistent accumulation then proceeds unchecked, yielding damage amplification bounded only by how long the block remains cached. A checksum-based countermeasure detects any single-bit corruption at scheduling time, bounding cumulative damage to one batch independent of the block's cache lifetime, with negligible overhead. These results argue for integrity protection of prefix blocks before end-to-end exploitation is demonstrated.

Fine-tuning adapts pretrained networks to new objectives. Whether the resulting depth profile of representational change reflects an intrinsic property of the model or the magnitude of gradient flow has not been tested directly. We measure this profile across 240 fine-tuning runs spanning 15 models in four architecture families (encoder and decoder transformers, a state-space model, and an RNN) at scales from 125M to 6.9B parameters. Representational change concentrates in output-proximal layers in every standard-training run except one. We apply a per-layer control that equalizes $\|ΔW\|/\|W\|$ across layers after each optimizer step. Under this control, the profile persists in some conditions and collapses in others. At 125M--350M, sequential-block architectures (BERT, OPT, GPT-2) retain the slope across tested objectives while parallel-block architectures (Pythia, CodeGen) retain it only for causal-language-modeling objectives. This architectural distinction narrows at 1.3B--1.4B, where both block types show positive equal-step slopes for CausalLM. Under standard training, profile shape is described by two additional axes: steepness tracks a training-free objective distance at initialization, and profile width is dominated by architecture. We treat the locality gradient, the depthwise slope of representational change, as a composite phenomenon whose components are scale-dependent.

Path planning is usually solved by addressing either the (high-level) route planning problem (waypoint sequencing to achieve the final goal) or the (low-level) path planning problem (trajectory prediction between two waypoints avoiding collisions). However, real-world problems usually require simultaneous solutions to the route and path planning subproblems with a holistic and efficient approach. In this paper, we introduce NaviFormer, a deep reinforcement learning model based on a Transformer architecture that solves the global navigation problem by predicting both high-level routes and low-level trajectories. To evaluate NaviFormer, several experiments have been conducted, including comparisons with other algorithms. Results show competitive accuracy from NaviFormer since it can understand the constraints and difficulties of each subproblem and act consequently to improve performance. Moreover, its superior computation speed proves its suitability for real-time missions.

We present Open-TQ-Metal, the first implementation of fused compressed-domain attention on Apple Silicon, enabling 128K-context inference for Llama 3.1 70B on a single 64GB consumer Mac -- a configuration impossible with all existing inference frameworks. Open-TQ-Metal quantizes the KV cache to int4 on the fly and computes attention directly on the compressed representation via custom Metal compute shaders, eliminating all intermediate dequantization matrices. Across 330 experiments spanning two model families (Gemma 4 31B and Llama 3.1 70B), the fused sdpa_int4 kernel achieves 48x attention speedup at 128K context over the dequantize-then-attend baseline, reduces KV cache memory from 40 GB to 12.5 GB (3.2x compression), and maintains identical top-1 token predictions to FP16 inference. We further provide the first cross-architecture analysis of KV cache quantization methods, revealing that the attention scale factor -- not model size -- determines whether angular quantization schemes like PolarQuant succeed or fail, with Gemma 4's attn_scale=1.0 amplifying directional error 25-100x more than Llama's standard 1/sqrt(d) scaling.

Shanghai Composite Index prediction has become a hot issue for many investors and academic researchers. Deep learning models are widely applied in multivariate time series forecasting, including recurrent neural networks (RNN), convolutional neural networks (CNN), and transformers. Specifically, the Transformer encoder, with its unique attention mechanism and parallel processing capabilities, has become an important tool in time series prediction, and has an advantage in dealing with long sequence dependencies and multivariate data correlations. Drawing on the strengths of various models, we propose the CNN-Transformer-LSTM Networks (CTLNet). This paper explores the application of CTLNet for Shanghai Composite Index prediction and the comparative experiments show that the proposed model outperforms state-of-the-art baselines.

Standard transformer architectures apply the same number of layers to every token regardless of contextual difficulty. We present Token-Selective Attention (TSA), a learned per-token gate on residual updates between consecutive transformer blocks. Each gate is a lightweight two-layer multi-layer perceptron (MLP) that produces a continuous halting probability, making the mechanism end-to-end differentiable with 1.7% parameter overhead and no changes to the base architecture. Notably, TSA learns difficulty-proportional routing without any explicit depth pressure: even at $λ=0$ (no depth regularisation), the task-loss gradient alone drives the router to skip 20% of token-layer operations. On character-level language modeling, TSA saved 14-23% of token-layer operations (TLOps) across Tiny-Shakespeare and enwik8 at <0.5% quality loss. At matched efficiency, TSA achieved 0.7% lower validation loss than early exit, and the learned routing transfers directly to inference-time sparse execution for real wall-clock speedup.

Modeling single-cell gene expression across diverse biological and technical conditions is crucial for characterizing cellular states and simulating unseen scenarios. Existing methods often treat genes as independent tokens, overlooking their high-level biological relationships and leading to poor performance. We introduce SAVE, a unified generative framework based on conditional Transformers for multi-condition single-cell modeling. SAVE leverages a coarse-grained representation by grouping semantically related genes into blocks, capturing higher-order dependencies among gene modules. A Flow Matching mechanism and condition-masking strategy further enhance flexible simulation and enable generalization to unseen condition combinations. We evaluate SAVE on a range of benchmarks, including conditional generation, batch effect correction, and perturbation prediction. SAVE consistently outperforms state-of-the-art methods in generation fidelity and extrapolative generalization, especially in low-resource or combinatorially held-out settings. Overall, SAVE offers a scalable and generalizable solution for modeling complex single-cell data, with broad utility in virtual cell synthesis and biological interpretation. Our code is publicly available at https://github.com/fdu-wangfeilab/sc-save

Spiking transformers achieve competitive accuracy with conventional transformers while offering $38$-$57\times$ energy efficiency on neuromorphic hardware, yet no theoretical framework guides their design. This paper establishes the first comprehensive expressivity theory for spiking self-attention. We prove that spiking attention with Leaky Integrate-and-Fire neurons is a universal approximator of continuous permutation-equivariant functions, providing explicit spike circuit constructions including a novel lateral inhibition network for softmax normalization with proven $O(1/\sqrt{T})$ convergence. We derive tight spike-count lower bounds via rate-distortion theory: $\varepsilon$-approximation requires $Ω(L_f^2 nd/\varepsilon^2)$ spikes, with rigorous information-theoretic derivation. Our key insight is input-dependent bounds using measured effective dimensions ($d_{\text{eff}}=47$--$89$ for CIFAR/ImageNet), explaining why $T=4$ timesteps suffice despite worst-case $T \geq 10{,}000$ predictions. We provide concrete design rules with calibrated constants ($C=2.3$, 95% CI: $[1.9, 2.7]$). Experiments on Spikformer, QKFormer, and SpikingResformer across vision and language benchmarks validate predictions with $R^2=0.97$ ($p<0.001$). Our framework provides the first principled foundation for neuromorphic transformer design.

Current large language models reason in isolation. Although it is common to sample multiple reasoning paths in parallel, these trajectories do not interact, and often fail in the same redundant ways. We introduce LACE, a framework that transforms reasoning from a collection of independent trials into a coordinated, parallel process. By repurposing the model architecture to enable cross-thread attention, LACE allows concurrent reasoning paths to share intermediate insights and correct one another during inference. A central challenge is the absence of natural training data that exhibits such collaborative behavior. We address this gap with a synthetic data pipeline that explicitly teaches models to communicate and error-correct across threads. Experiments show that this unified exploration substantially outperforms standard parallel search, improving reasoning accuracy by over 7 points. Our results suggest that large language models can be more effective when parallel reasoning paths are allowed to interact.

Looped transformers promise test-time compute scaling by spending more iterations on harder problems, but it remains unclear which architectural choices let them extrapolate to harder problems at test time rather than memorize training-specific solutions. We introduce a fixed-point based framework for analyzing looped architectures along three axes of stability -- reachability, input-dependence, and geometry -- and use it to characterize when fixed-point iteration yields meaningful predictions. Theoretically, we prove that looped networks without recall have countable fixed points and cannot achieve strong input-dependence at any spectral regime, while recall combined with outer normalization reliably produces a regime in which fixed points are simultaneously reachable, locally smooth in the input, and supported by stable backpropagation. Empirically, we train single-layer looped transformers on chess, sudoku, and prefix-sums and find that downstream performance tracks the framework's predictions across tasks and architectural configurations. We additionally introduce internal recall, a novel recall placement variant, and show that it becomes competitive with -- and on sudoku, substantially better than -- standard recall placement once outer normalization is applied.

Sparse attention has been proposed as a way to alleviate the quadratic cost of transformers, a central bottleneck in long-context training. A promising line of work is $α$-entmax attention, a differentiable sparse alternative to softmax that enables input-dependent sparsity yet has lagged behind softmax due to the computational overhead necessary to compute the normalizer $τ$. In this paper, we introduce AdaSplash-2, which addresses this limitation through a novel histogram-based initialization that reduces the number of iterations needed to compute $τ$ to typically 1--2. The key idea is to compute a coarse histogram of attention scores on the fly and store it in on-chip SRAM, yielding a more accurate initialization that enables fast forward and backward computation. Combined with a sparsity-aware GPU implementation that skips zero blocks with low overhead, AdaSplash-2 matches or improves per-step training time relative to FlashAttention-2 when block sparsity is moderate-to-high (e.g., $>$60%), which often occurs at long-context lengths. On downstream tasks, models trained with our efficient $α$-entmax attention match softmax baselines at short-context lengths and achieve substantial gains in long-context settings.

KV caching is a ubiquitous optimization in autoregressive transformer inference, long presumed to be numerically equivalent to cache-free computation. This assumption fails under standard FP16 precision: cache-ON and cache-OFF execution paths employ different floating-point accumulation orderings which, due to FP16 non-associativity, produce a deterministic divergence in decoded token sequences. Across three open-weight models (LLaMA-2-7B, Mistral-7B-v0.3, Gemma-2-2B) evaluated on GSM8K, we observe a 100% token divergence rate across all sampling strategies, including greedy decoding, which rules out sampling randomness as a cause, and also with cache-ON yielding higher accuracy in 8 of 9 conditions, where the accuracy difference serves as an indicator that the divergence direction is systematic rather than random. Controlled FP32 falsification reduces divergence by eight orders of magnitude, eliminates token flips, and drops the flip rate to exactly 0.0%, confirming FP16 non-associativity as the sole causal driver. Layer-wise drift profiling reveals architecturally predictable propagation patterns: models using Grouped-Query Attention exhibit sharp divergence at the first layer, while Gemma's larger head dimension and sliding window attention produce uniform accumulation across all layers. Finally, activation patching of the entire residual stream fails to recover the cache-free trajectory, localizing the causal variable to the stateful KV cache. These findings establish that FP16 KV cache inference is fundamentally non-equivalent to recomputation and provide a mechanistic framework for understanding numerical instability in modern LLM inference systems.

Token pruning methods for Vision Transformers (ViTs) promise quadratic reductions in attention FLOPs by dropping uninformative patches. Yet standard variable-length attention APIs -- including FlashAttention-2's varlen and PyTorch's NestedTensor SDPA -- fail to translate these savings into proportional wall-clock gains at the short post-pruning sequence lengths typical of ViTs ($\leq$197 tokens). We identify a dispatch-overhead bottleneck: at these lengths, host-side kernel dispatch consumes ${\sim}$50,$μ$s regardless of workload, exceeding the actual GPU compute time at moderate-to-high pruning rates. We present a lightweight bidirectional Triton attention kernel whose dispatch floor is ${\sim}$24,$μ$s -- roughly 2.17$\times$ lower than FlashAttention-2 varlen -- allowing pruning savings to become visible in wall-clock time. Integrated into a complete pack-attend-unpack pipeline and evaluated on an NVIDIA RTX 4000 Ada Generation GPU, our system achieves 1.88$\times$ end-to-end throughput over padded PyTorch SDPA at standard 224$\times$224 inputs, scaling to 2.51$\times$ at 384$\times$384. Against FlashAttention-2 varlen -- the strongest baseline -- our kernel delivers 9-12% higher throughput at serving batch sizes (BS=1-4), and 2.17$\times$ lower kernel latency at 80% token pruning. Numerical correctness is verified with max absolute logit difference $<$0.004 and bit-exact top-1 predictions.

When do transformers commit to a decision, and what prevents them from correcting it? We introduce \textbf{prolepsis}: a transformer commits early, task-specific attention heads sustain the commitment, and no layer corrects it. Replicating \citeauthor{lindsey2025biology}'s (\citeyear{lindsey2025biology}) planning-site finding on open models (Gemma2 2B, Llama3.2 1B), we ask five questions. (Q1)~Planning is invisible to six residual-stream methods; CLTs are necessary. (Q2)~The planning-site spike replicates with identical geometry. (Q3)~Specific attention heads route the decision to the output, filling a gap flagged as invisible to attribution graphs. (Q4)~Search requires ${\leq}16$ layers; commitment requires more. (Q5)~Factual recall shows the same motif at a different network depth, with zero overlap between recurring planning heads and the factual top-10. Prolepsis is architectural: the template is shared, the routing substrates differ. All experiments run on a single consumer GPU (16,GB VRAM).

Recently, sparse autoencoders (SAEs) have emerged as a promising technique for interpreting activations in foundation models by disentangling features into a sparse set of concepts. However, identifying the optimal level of sparsity for each neuron remains challenging in practice: excessive sparsity can lead to poor reconstruction, whereas insufficient sparsity may harm interpretability. While existing activation functions such as ReLU and TopK provide certain sparsity guarantees, they typically require additional sparsity regularization or cherry-picked hyperparameters. We show in this paper that dynamically sparse attention mechanisms using sparsemax can bridge this trade-off, due to their ability to determine the activation numbers in a data-dependent manner. Specifically, we first explore a new class of SAEs based on the cross-attention architecture with the latent features as queries and the learnable dictionary as the key and value matrices. To encourage sparse pattern learning, we employ a sparsemax-based attention strategy that automatically infers a sparse set of elements according to the complexity of each neuron, resulting in a more flexible and general activation function. Through comprehensive evaluation and visualization, we show that our approach successfully achieves lower reconstruction loss while producing high-quality concepts, particularly in top-n classification tasks.

To quantify the geometric expressivity of transformers, we introduce a tropical geometry framework to characterize their exact spatial partitioning capabilities. By modeling self-attention as a vector-valued tropical rational map, we prove it evaluates exactly to a Power Voronoi Diagram in the zero-temperature limit. Building on this equivalence, we establish a combinatorial rationale for Multi-Head Self-Attention (MHSA): via the Minkowski sum of Newton polytopes, multi-head aggregation expands the polyhedral complexity to $\mathcal{O}(N^H)$, overcoming the $\mathcal{O}(N)$ bottleneck of single heads. Extending this to deep architectures, we derive the first tight asymptotic bounds on the number of linear regions in transformers ($Θ(N^{d_{\text{model}}L})$), demonstrating a combinatorial explosion driven intrinsically by sequence length $N$, ambient embedding dimension $d_{\text{model}}$, and network depth $L$. Importantly, we guarantee that this idealized polyhedral skeleton is geometrically stable: finite-temperature soft attention preserves these topological partitions via exponentially tight differential approximation bounds.

Existing single-modal RGB trackers often face performance bottlenecks in complex dynamic scenes, while the introduction of event sensors offers new potential for enhancing tracking capabilities. However, most current RGB-event fusion methods, primarily designed in the spatial domain using convolutional, Transformer, or Mamba architectures, fail to fully exploit the unique temporal response and high-frequency characteristics of event data. To address this, we1 propose FreqTrack, a frequency-aware RGBE tracking framework that establishes complementary inter-modal correlations through frequency-domain transformations for more robust feature fusion. We design a Spectral Enhancement Transformer (SET) layer that incorporates multi-head dynamic Fourier filtering to adaptively enhance and select frequency-domain features. Additionally, we develop a Wavelet Edge Refinement (WER) module, which leverages learnable wavelet transforms to explicitly extract multi-scale edge structures from event data, effectively improving modeling capability in high-speed and low-light scenarios. Extensive experiments on the COESOT and FE108 datasets demonstrate that FreqTrack achieves highly competitive performance, particularly attaining leading precision of 76.6% on the COESOT benchmark, validating the effectiveness of frequency-domain modeling for RGBE tracking.

Rank reduction discards dimensions; quantization keeps them at lower precision. Comparing the two requires a choice of what compression should preserve. For attention values, we study reconstruction of the attention output rather than reconstruction of the values alone. With fixed attention weights, the optimal orthogonal rank-$r$ projection uses the leading eigenvectors of $V^\topα^\topαV$, and its error is exactly the discarded eigenvalue sum. We extend this objective to calibration datasets and grouped query attention, and describe rank allocation under an additive local error budget. We also examine the limits of using local error to predict downstream loss. Historical weight-perturbation experiments favor coefficient rounding over the value projections tested, but do not establish a comparison at equal cache storage. The resulting distinction is practical: the projection objective has an exact solution, while a comparison with cache quantization requires separate activation-level experiments.

We investigate how looped transformers encode human preference, training lightweight evaluator heads on frozen Ouro-2.6B loop-iteration states on Anthropic HH-RLHF. v2: an erratum is prepended; the original manuscript is unchanged. A post-publication audit found the three headline results inflated by two independent evaluation errors. The 95.2% pairwise evaluator accuracy is a canonical-ordering artifact: the data were correctly split, but the evaluator learned to prefer the first-presented argument; its strict antisymmetrized accuracy on the full 8,552-pair test set is 63.9%. The 84.5% pairwise probe and the below-chance 21.75% pointwise probe were source-item leaks (orientation rows and pair partners crossing the train/test split); corrected pair-disjoint values are 56.5% and 54.2% -- above chance, so the "inverted polarity" finding is withdrawn. The central finding survives at much smaller magnitude: preference is decoded more accurately relationally than pointwise (paired +2.3 points, 95% CI [+1.3, +3.3]), and the antisymmetrized evaluator still beats the linear probe, but no corrected readout rivals end-to-end reward models. The methodological findings stand (constant-output degeneracy, flip test, swap-protocol metric deflation), with one correction: antisymmetrized accuracy, not antisymmetry correlation, certifies relational discrimination. The two errors are mutually invisible -- a split audit cannot see an ordering prior, antisymmetrization cannot see a leak -- so both checks are required. Full audit in the follow-up work (Kirin, 2026, in preparation).

We introduce the Large Sparse Reconstruction Model to study how scaling transformer context windows affects feed-forward 3D reconstruction. Although recent object-centric feed-forward methods produce robust, high-quality reconstructions, they still lag behind dense-view optimization in recovering fine-grained texture and appearance. We show that expanding the context window -- by substantially increasing the number of active object and image tokens -- narrows this gap and enables high-fidelity 3D object reconstruction and inverse rendering. To scale effectively, we adapt native sparse attention for 3D reconstruction with three key contributions: (1) an efficient coarse-to-fine pipeline that focuses computation on informative regions by predicting sparse high-resolution residuals; (2) a 3D-aware spatial routing mechanism that establishes accurate 2D-3D correspondences using explicit geometric distances rather than standard attention scores; and (3) a custom block-aware sequence-parallel strategy with an All-gather-KV protocol to balance dynamic, sparse workloads across GPUs. As a result, LSRM handles 20x more object tokens and >2x more image tokens than prior state-of-the-art (SOTA) methods. Extensive evaluations on standard novel-view synthesis benchmarks show substantial gains over the current SOTA, yielding >2.4dB higher PSNR and >40% lower LPIPS. Furthermore, when extending LSRM to inverse rendering, qualitative and quantitative evaluations on widely used benchmarks demonstrate consistent improvements in texture and geometry details, achieving an LPIPS that matches or exceeds that of SOTA dense-view optimization methods. Code and model weights are available on our project page.

Token-level sparse attention mechanisms, exemplified by DeepSeek Sparse Attention (DSA), achieve fine-grained key selection by scoring every historical key for each query through a lightweight indexer, then computing attention only on the selected subset. While the downstream sparse attention itself scales favorably, the indexer must still scan the entire prefix for every query, introducing an per-layer bottleneck that grows prohibitively with context length. We propose HISA (Hierarchical Indexed Sparse Attention), a plug-and-play replacement for the indexer that rewrites the search path from a flat token scan into a two-stage hierarchical procedure: (1) a block-level coarse filtering stage that scores pooled block representations to discard irrelevant regions, followed by (2) a token-level refinement stage that applies the original indexer exclusively within the retained candidate blocks. HISA preserves the identical token-level top-sparse pattern consumed by the downstream Sparse MLA operator and requires no additional training. On kernel-level benchmarks, HISA achieves up to speedup at 64K context. On Needle-in-a-Haystack and LongBench, we directly replace the indexer in DeepSeek-V3.2 and GLM-5 with our HISA indexer, without any finetuning. HISA closely matches the original DSA in quality, while substantially outperforming block-sparse baselines.

Mixture-of-Experts (MoE) models substantially improve performance by increasing the capacity of dense architectures. However, directly training MoE models requires considerable computational resources and introduces extra overhead in parameter storage and deployment. Therefore, it is critical to develop an approach that leverages the multi-expert capability of MoE to enhance performance while incurring minimal additional cost. To this end, we propose a novel pre-training approach, termed ExFusion, which improves the efficiency of Transformer training through multi-expert fusion. Specifically, during the initialization phase, ExFusion upcycles the feed-forward network (FFN) of the Transformer into a multi-expert configuration, where each expert is assigned a weight for later parameter fusion. During training, these weights allow multiple experts to be fused into a single unified expert equivalent to the original FFN, which is subsequently used for forward computation. As a result, ExFusion introduces multi-expert characteristics into the training process while incurring only marginal computational cost compared to standard dense training. After training, the learned weights are used to integrate multi-experts into a single unified expert, thereby eliminating additional overhead in storage and deployment. Extensive experiments on a variety of computer vision and natural language processing tasks demonstrate the effectiveness of the proposed method.

Most recent advances in 3D generative modeling rely on diffusion or flow-matching formulations. We instead explore a fully autoregressive alternative and introduce GaussianGPT, a transformer-based model that directly generates 3D Gaussians via next-token prediction, thus facilitating full 3D scene generation. We first compress Gaussian primitives into a discrete latent grid using a sparse 3D convolutional autoencoder with vector quantization. The resulting tokens are serialized and modeled using a causal transformer with 3D rotary positional embedding, enabling sequential generation of spatial structure and appearance. Unlike diffusion-based methods that refine scenes holistically, our formulation constructs scenes step-by-step, naturally supporting completion, outpainting, controllable sampling via temperature, and flexible generation horizons. This formulation leverages the compositional inductive biases and scalability of autoregressive modeling while operating on explicit representations compatible with modern neural rendering pipelines, positioning autoregressive transformers as a complementary paradigm for controllable and context-aware 3D generation.

Converting a pretrained Transformer into a more efficient hybrid model through distillation offers a promising approach to reducing inference costs. However, achieving high-quality generation in distilled models requires careful joint design of both the student architecture and the distillation process. Many prior distillation works evaluate downstream multiple-choice benchmarks by ranking candidate answers with log-likelihood rather than requiring autoregressive generation, which can obscure important differences in model quality. For example, on overlapping benchmarks, we show that a 7B distilled model that nearly matches its teacher to within 0.2 pp under log-likelihood scoring falls behind by 20.8 pp when it must generate answers autoregressively. We investigate this phenomenon with GenDistill, a multi-stage pipeline we designed for distilling a pretrained Transformer into an efficient Hybrid Kimi Delta Attention (Hybrid-KDA) student. Using it as a controlled testbed on Qwen3-0.6B, we systematically ablate six design axes (training objective, loss masking, training duration, dataset selection, parameter freezing, and architecture choice) and evaluate every choice under both log-likelihood and generation-based protocols. We find that log-likelihood-based evaluation consistently underestimates the gap between teacher and student, and can in some cases reverse the ranking of design choices, so conclusions drawn from perplexity-only evaluation may be misleading. Among the factors we study, dataset selection, completion-only masking, and freezing attention layers during post-training have the largest impact on generation quality. Our best distillation recipe, using a Hybrid-KDA model as the student, retains 86-90% of teacher accuracy on knowledge benchmarks while reducing KV cache memory by up to 75% and improving time-to-first-token by 2-4x at 128K-token contexts.

Transformers have shown inconsistent success in AI planning tasks, and theoretical understanding of when generalization should be expected has been limited. We take important steps towards addressing this gap by analyzing the ability of decoder-only models to verify whether a given plan correctly solves a given planning instance. To analyse the general setting where the number of objects -- and thus the effective input alphabet -- grows at test time, we introduce C*-RASP, an extension of C-RASP designed to establish length generalization guarantees for transformers under the simultaneous growth in sequence length and vocabulary size. Our results identify a large class of classical planning domains for which transformers can provably learn to verify long plans, and structural properties that significantly affects the learnability of length generalizable solutions. Empirical experiments corroborate our theory.

Understanding the internal mechanisms of transformer-based large language models (LLMs) is crucial for their reliable deployment and effective operation. While recent efforts have yielded a plethora of attribution methods attempting to balance faithfulness and computational efficiency, dense component attribution remains prohibitively expensive. In this work, we introduce Dual Path Attribution (DPA), a novel framework that faithfully traces information flow on the frozen transformer in one forward and one backward pass without requiring counterfactual examples. DPA analytically decomposes and linearizes the computational structure of the SwiGLU Transformers into distinct pathways along which it propagates a targeted unembedding vector to receive the effective representation at each residual position. This target-centric propagation achieves O(1) time complexity with respect to the number of model components, scaling to long input sequences and dense component attribution. Extensive experiments on standard interpretability benchmarks demonstrate that DPA achieves state-of-the-art faithfulness and unprecedented efficiency compared to existing baselines.

Language models struggle to generalize beyond pretraining context lengths, limiting long-horizon reasoning and retrieval. Continued pretraining on long-context data can help but is expensive due to the quadratic scaling of Attention. We observe that most tokens do not require (Global) Attention over the entire sequence and can rely on local context. Based on this, we propose L2A (Learning To Attend), a layer that enables conditional (token-wise) long-range memory access by deciding when to invoke global attention. We evaluate L2A on Qwen 2.5 and Qwen 3 models, extending their effective context length from 32K to 128K tokens. L2A matches the performance of standard long-context training to within 3% while skipping Global Attention for $\sim$80% of tokens, outperforming prior baselines. We also design custom Triton kernels to efficiently implement this token-wise conditional Attention on GPUs, achieving up to $\sim$2$\times$ improvements in training throughput and time-to-first-token over FlashAttention. Moreover, L2A enables post-training pruning of highly sparse Global Attention layers, reducing KV cache memory by up to 50% with negligible performance loss. Our code is released under Apache 2.0 at https://github.com/awslabs/hybrid-model-factory/tree/main/examples/research/L2A.

Transformer models have redefined sequence learning, yet dot-product self-attention introduces a quadratic token-mixing bottleneck for long-context time-series. We introduce the Phasor Transformer block, a phase-native alternative representing sequence states on the unit-circle manifold $S^1$. Each block combines lightweight trainable phase-shifts with parameter-free Discrete Fourier Transform (DFT) token coupling, achieving global $\mathcal{O}(N\log N)$ mixing without explicit attention maps. Stacking these blocks defines the Large Phasor Model (LPM). We validate LPM on autoregressive time-series prediction over synthetic multi-frequency benchmarks against honest baselines: it beats a zero-parameter persistence baseline and, with the corrected gradient path, improves monotonically with depth before saturating, while remaining competitive-but-not-superior to self-attention at a fraction of the parameter count. Our results establish an explicit efficiency--accuracy frontier, showing that scalable temporal modeling in oscillatory domains can emerge from geometry-constrained phase computation with deterministic global coupling.

Vision foundation models trained with self-supervised objectives achieve strong performance across diverse tasks and exhibit emergent object segmentation properties. However, their alignment with human object perception remains poorly understood. Here, we introduce a behavioral benchmark in which participants make same/different object judgments for dot pairs on naturalistic scenes, scaling up a classical psychophysics paradigm to over 1000 trials. We test a diverse set of vision models using a simple readout from their representations to predict subjects' reaction times. We observe a steady improvement across model generations, with both architecture and training objective contributing to alignment, and transformer-based models trained with the DINO self-supervised objective showing the strongest performance. To investigate the source of this improvement, we propose a novel metric to quantify the object-centric component of representations by measuring patch similarity within and between objects. Across models, stronger object-centric structure predicts human segmentation behavior more accurately. We further show that matching the Gram matrix of supervised transformer models, capturing similarity structure across image patches, with that of a self-supervised model through distillation improves their alignment with human behavior, converging with the prior finding that Gram anchoring improves DINOv3's feature quality. Together, these results demonstrate that self-supervised vision models capture object structure in a behaviorally human-like manner, and that Gram matrix structure plays a role in driving perceptual alignment.

Diffusion Transformers (DiTs) are a dominant backbone for high-fidelity text-to-image generation due to strong scalability and alignment at high resolutions. However, quadratic self-attention over dense spatial tokens leads to high inference latency and limits deployment. We observe that denoising is spatially non-uniform with respect to aesthetic descriptors in the prompt. Regions associated with aesthetic tokens receive concentrated cross-attention and show larger temporal variation, while low-affinity regions evolve smoothly with redundant computation. Based on this insight, we propose AccelAes, a training-free framework that accelerates DiTs through aesthetics-aware spatio-temporal reduction while improving perceptual aesthetics. AccelAes builds AesMask, a one-shot aesthetic focus mask derived from prompt semantics and cross-attention signals. When localized computation is feasible, SkipSparse reallocates computation and guidance to masked regions. We further reduce temporal redundancy using a lightweight step-level prediction cache that periodically replaces full Transformer evaluations. Experiments on representative DiT families show consistent acceleration and improved aesthetics-oriented quality. On Lumina-Next, AccelAes achieves a 2.11$\times$ speedup and improves ImageReward by +11.9% over the dense baseline. Code is available at https://github.com/xuanhuayin/AccelAes.

Vision Transformers require significant computational resources and memory bandwidth, severely limiting their deployment on resource-constraint hardware. Most structured pruning methods reduce theoretical cost effectively, yet they typically operate at a single structural granularity and depend on multi-stage pipelines with importance ranking, auxiliary solvers or post-hoc magnitude thresholding, followed by a separate fine-tuning phase to recover accuracy. We propose Hierarchical Auto-Pruning (HiAP), which casts ViT pruning as a single budget-aware learning problem and jointly allocates sparsity across four granularities in one end-to-end phase. HiAP introduces stochastic Gumbel-Sigmoid gates at macro level (attention heads and FFN blocks) and micro level (intra-head dimensions and FFN neurons), and optimizes them against the task loss together with an analytical MAC cost term. The budget coefficient steers the network to a target compute level while the gates gradually harden into a dense, smaller sub-network at convergence. It does not require importance heuristics, ranking metrics, auxillary solvers or secondary fine-tuning. On ImageNet with DeiT small, HiAP automatically discovers hetergenous architectures, pruning depths, heads, and width by different amount across layers, and reaches competitive accuracies against substantially more complex pruning pipelines at comparable compute from a single training run.

Event cameras provide sequential visual data with spatial sparsity and high temporal resolution, making them attractive for low-latency object detection. Existing asynchronous event-based neural networks exploit this low-latency advantage by updating predictions event by event, but still suffer from two bottlenecks: recurrent architectures are difficult to train efficiently on long sequences, and improving accuracy often increases per-event computation and latency. Linear attention is appealing because it enables parallel training and recurrent inference. However, its dense state updates make per-event computation scale with the state size, yielding a poor accuracy-efficiency trade-off for object detection, where accurate localization requires fine-grained spatial states. The key challenge is therefore to introduce sparse state activation that exploits the spatial sparsity of events while preserving efficient parallel training. We propose Spatially-Sparse Linear Attention (SSLA), which introduces a mixture-of-spaces state decomposition and a scatter-compute-gather training procedure, enabling state-level sparsity as well as training parallelism. Building on SSLA, we develop an end-to-end asynchronous linear attention model, SSLA-Det, for low-latency event-based object detection. On Gen1 and N-Caltech101, SSLA-Det achieves state-of-the-art accuracy among asynchronous methods, reaching 0.375 mAP and 0.515 mAP, respectively, while reducing per-event computation by over 20 times compared with the strongest prior asynchronous baseline, demonstrating the potential of linear attention for low-latency event-based vision.

Multimodal Transformers often produce predictions without clarifying how different modalities jointly support a decision. Most existing multimodal explainable AI (MXAI) methods extend unimodal saliency to multimodal backbones, highlighting important tokens or patches within each modality, but they rarely pinpoint which cross-modal feature pairs provide complementary evidence (synergy) or serve as reliable backups (redundancy). We present Feature-level I2MoE (FL-I2MoE), a structured Mixture-of-Experts layer that operates directly on token/patch sequences from frozen pretrained encoders and explicitly separates unique, synergistic, and redundant evidence at the feature level. We further develop an expert-wise explanation pipeline that combines attribution with top-K% masking to assess faithfulness, and we introduce Monte Carlo interaction probes to quantify pairwise behavior: the Shapley Interaction Index (SII) to score synergistic pairs and a redundancy-gap score to capture substitutable (redundant) pairs. Across three benchmarks (MMIMDb, ENRICO, and MMHS150K), FL-I2MoE yields more interactionspecific and concentrated importance patterns than a dense Transformer with the same encoders. Finally, pair-level masking shows that removing pairs ranked by SII or redundancy-gap degrades performance more than masking randomly chosen pairs under the same budget, supporting that the identified interactions are causally relevant. Code is available at https://github.com/dut0817/FL-I2MoE.

Transformers excel at in-context retrieval but suffer from quadratic complexity with sequence length, while State Space Models (SSMs) offer efficient linear-time processing but have limited retrieval capabilities. We investigate whether hybrid architectures combining Transformers and SSMs can achieve the best of both worlds on two synthetic in-context retrieval tasks. The first task, n-gram retrieval, requires the model to reproduce an n-gram that succeeds the query within the input sequence. The second task, position retrieval, presents the model with a query token and requires it to perform a two-hop lookup: first locating the corresponding element in the sequence, and then outputting its positional index. Under controlled conditions, we assess data efficiency, length generalization, robustness to out of domain training examples, and learned representations across Transformers, SSMs, and hybrid architectures. We find that hybrid models outperform SSMs and match or exceed Transformers in terms of data efficiency and extrapolation for tasks that require precise information retrieval from the input context. However, Transformers maintain superiority in position retrieval tasks. Through representation analysis, we discover that SSM-based models develop locality-aware embeddings where tokens representing adjacent positions become neighbors in embedding space, forming interpretable structures. This property is absent in Transformers as causal attention is sufficient for acquiring positional associations, and the introduction of positional encoding amplifies this behavior, leading to improvement in data efficiency. SSMs on the other hand update their internal representations incrementally and without positional encodings, are required to learn these associations. Our findings reveal fundamental differences in how Transformers and SSMs, and hybrid models learn positional associations.