Attention
-
S$^3$R: Fast and Faithful RL Rollouts via Self-Sparse Speculative Decoding
2026-10-09 08:22:40Yuhang He, Hongyu Ge, Zihan Lin, Siyi Liu, Yongqi ZhangAbstract
Fast rollout generation for long-context reinforcement learning with verifiable rewards (RLVR) requires reducing KV-cache traffic without changing the response groups used for policy optimization. In our profiled RLVR workloads, attention accounts for 72--76\% of autoregressive decoding kernel time. Sparse attention reduces these reads, but sampling directly from a sparse policy changes the response groups used for learning. For group-relative optimization, we show that even exact response-wise reweighting does not generally recover the expected dense-policy update because the advantage of each response depends on the rewards of other group members. We propose Self-Sparse Speculative Rollouts (S$^3$R). It uses sparse attention to propose tokens, then generates responses through full-attention verification and exact speculative sampling. Drafting and verification share the same policy parameter, eliminating the need to train or update a separate drafter. The resulting response groups follow the dense-policy distribution. Under S$^3$R, we characterize speedup conditions and the costs of drafting and verification. We found that dense verification still requires the complete historical KV cache. A KV-aware scheduler therefore limits KV-cache repeated offloading and restoration while maintaining rollout parallelism. % OR thrashing In a matched ablation with an aggressive sparse drafting budget, S$^3$R maintains stable training while direct sparse sampling with training-side correction becomes unstable. On long-context mathematical reasoning, S$^3$R achieves $1.39\times~1.35\times$ the end-to-end training throughput of autoregressive rollouts on Qwen3 model, with comparable task performance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13202
PDF
-
High-Throughput Generative Video Super-Resolution on a Single Consumer GPU
2026-10-09 08:06:14Georgios Panaretou, Sotirios ChatzisAbstract
Generative video super-resolution (VSR) can recover visually compelling high-frequency detail by leveraging diffusion priors, but iterative sampling and heavy spatio-temporal coupling make practical deployments difficult. We study a different operating point: practical HD generative VSR on a single high-end consumer GPU without end-to-end retraining. To this end, we present a zero-shot framework built entirely from pretrained components. The proposed system decouples spatial enhancement from temporal stabilization. A time-distilled one-step diffusion backbone handles spatial upscaling, while temporal coherence is promoted by a lightweight pretrained flow-based interpolator that synthesizes intermediate frames between super-resolved keyframes. To further improve scalability, we introduce a resolution-aware hybrid attention scheme that, in a zero-shot manner, replaces selected high-resolution global-attention layers with neighborhood attention while preserving global attention in semantically critical low-resolution layers, while also improving temporal consistency. Experiments show that the proposed framework reaches up to ~32 FPS at 720p resolution under interleaved inference on a single NVIDIA RTX 4090. At 4K, the proposed attention substitution can yield up to a 44% speedup over the original attention implementation. Across standard VSR benchmarks, the method consistently shifts the speed-quality trade-off decisively toward runtime efficiency, trading reconstruction fidelity for strong no-reference perceptual quality, while remaining reasonably competitive against generative approaches in temporal consistency. It also exposes clear limitations in handling more challenging motion, which a simple motion gate partially mitigates. Code will be released upon acceptance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13179
PDF
-
Fix the Match, Not the Solver: STITCH -- a Self-Supervised Compatibility Measure for Eroded Visual Puzzles
2026-10-09 08:03:16Nadav Alali, Ohad Ben-ShaharAbstract
Unlike the clean and pixel-aligned complete pieces of standard jigsaw benchmarks, real-world puzzle fragments lose their boundary regions to erosion, destroying the signal that classical compatibility measures (CMs) rely on. We first re-establish, via oracle experiments, that the CM, not the solver, is the performance bottleneck: with a perfect CM, even a simple greedy solver achieves perfect reconstruction. This motivates STITCH, a self-supervised CM that sidesteps missing boundaries by comparing frozen ViT patch tokens from piece interiors via a lightweight cross-attention head (~249K trainable parameters), trained without manual annotation. Plugged into unchanged compatibility-based solvers, STITCH reaches .992/.954 Abs/Perf on JPwLEG-5 (+.629 Perf over the best prior method), nearly saturating the benchmark. On a new ImageNet 7x7 benchmark, it outperforms all alternatives, including inpaint-then-match pipelines built on modern generative models. Importantly, performance remains robust across erosion levels and generalizes to puzzles with over 40% missing pieces without retraining. Trained on the recent GAP datasets of irregular, heavily eroded fragments, the same CM recipe enables compatibility-based solvers to surpass recent learning-based solvers on GAP-3 and GAP-5. Since this leap in performance is achieved by a pluggable CM module, the takeaway is clear: fix the match, not the solver.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13176
PDF
-
GuidedNav: Shaping Vision-Language Navigation Representations through Pre-action Attention and Spatial Guidance
2026-10-09 07:18:03Jingyu Guo, Jiaxin Huang, Chengxing Lin, Xiaoyu Luo, Cheng Wen, Shanshan Ye, Tingjin Chu, Shaoli Huang, Mingming GongAbstract
Vision-and-language navigation (VLN) requires an agent to track progress through an extended instruction, identify landmarks relevant to its current objective, and understand the surrounding spatial structure to act. End-to-end navigation policies typically rely on action supervision to implicitly learn these capabilities. Without explicit guidance on the evidence underlying each decision, however, policies can overfit to correlations predictive of training actions rather than learn the task-relevant grounding and geometry needed for generalization. Our insight is to guide the representations that connect instructions to actions: identifying what remains to be done determines which visual evidence matters, while understanding the geometry around that evidence informs how to move. We introduce GuidedNav, a unified framework that explicitly shapes these representations through training-time guidance. Subtask and landmark guidance steer native pre-action attention toward the active instruction segment and relevant visual regions, grounding action prediction in the current objective. Spatial guidance aligns the policy's internal representations with geometric features that capture long-horizon context and local spatial structure. Together, these objectives strengthen navigation capabilities while avoiding the inference overhead of intermediate textual reasoning and online geometric encoding. We also introduce Guided-R2R and Guided-RxR, two annotation-enriched navigation datasets that support more interpretable VLN research. GuidedNav establishes state-of-the-art performance among the compared monocular RGB policies, improving success rates by 8.1% and 13.1% relative to the strongest prior results on R2R-CE and RxR-CE val-unseen, respectively, while achieving faster inference. Extensive experiments further demonstrate the effectiveness and real-world transferability of our approach.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13122
PDF
-
What Do Visual-Head Discovery Scores Actually Find? Attention Source, Necessity, and Random-Control Pitfalls in Vision-Language Models
2026-10-09 07:06:17Md Abdul Kadir, Daniel SonntagAbstract
Attention heads that attend to image regions relevant to a question are increasingly used to steer, ground and interpret vision-language models (VLMs) and to reduce their hallucinations. They are mostly found by ranking heads with an attention- pattern score and validated by steering or masking them. We ask how the choices inside such a discovery score shape the heads it returns, and which of the two validations identifies them. On Qwen3-VL-8B-Instruct we vary the query type, the discovery corpus and the token positions whose attention is read (the attention source) in a full 2 x 3 x 2 design, and test all 12 head-sets for steerability (CERS: Contrastive Evidence-Region Steering score) and necessity (NCS: Necessity Con- trast Score) against layer-matched random head-sets. (1) The attention source and the query type reshape the selected heads far more than the corpus: changing the source keeps a median of 49 of the top-100 heads, the query 55 and the cor- pus 84, while two halves of the same data share 96-99. (2) Necessity follows the attention source: head-sets discovered from query-span attention are more nec- essary than last-token ones (difference 0.083, p=6 x 10^-10), and the two re- ferring + query-span sets lie 17-19 standard deviations above 250 layer-matched random sets (Holm-adjusted p=0.048). The referring + query-span effect re- plicates on several recent VLMs. (3) Steerability is not specific to the discovered heads: no discovered head-set steers better than its layer-matched random sets (rank p >= 0.14), and random sets drawn from the same or from any layer span the discovered sets' scores. Necessity therefore separates the discovery settings, while steering at this strength does not. We turn these findings into a protocol for reporting visual-head discoveries.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13111
PDF
-
Attention Sees, MLP Refuses: Tracing the Attention-to-MLP Safety Pathway in LLMs
2026-10-09 06:34:26Zixuan Weng, Haoqi He, Weifei Jin, Yue Huang, Yang BaiAbstract
Safety-aligned large language models (LLMs) remain vulnerable to jailbreaks and can over-refuse benign requests. Understanding these failures requires distinguishing harmfulness recognition from the computation that writes refusal. We reframe this as a pathway question and make it measurable on residual-stream updates in decoder-only Transformers, separating the harmfulness information a module update exposes from the change it produces in readable refusal evidence. Across 18 Base-Instruct pairs spanning Qwen, Llama, and Gemma, harmfulness is decoded more readily from attention updates, particularly in early layers, whereas MLP updates contribute more strongly to harm-selective refusal writing in aligned models. Post-training strengthens MLP refusal writing in all 18 pairs without systematically improving harmfulness decodability, while positive MLP writing becomes more concentrated in deeper layers after post-training. Across six reasoning-enabled models, harmfulness remains decodable during reasoning while strong refusal writing concentrates at answer onset, a pattern that persists with reasoning disabled. Causal interventions further support this pathway: resetting downstream MLP updates to their clean values substantially reduces the refusal shift caused by upstream attention interventions. Building on these findings, we introduce Sentinel, which combines an attention-derived perception score P with an MLP-derived decision score D. Their joint profiles characterize jailbreaks and over-refusals, and conditional routing with both signals achieves a lower aggregate error rate than single-signal routing and evaluated fixed reminders across eight models.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13078
PDF
-
CliQ-RRG: Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation
2026-10-09 06:13:25Chandravardhan Singh Raghaw, Shahid Shafi Dar, Arnav Nishith Sharma, Shrey Modi, Nagendra KumarAbstract
Radiology reports are essential for diagnostic reasoning and patient care, yet their manual preparation is time-consuming and cognitively demanding. Automatic radiology report generation (RRG) offers a scalable alternative, but existing models often produce lengthy, unstructured narratives that overlook diagnostic cues and multi-view information. We present Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation (CliQ-RRG), a unified two-stage framework for interpretable and clinically grounded reporting. In Stage 1, CliQ-RRG employs a Disease-aware Visual-Textual Alignment module that aligns image and text representations using predicted disease embeddings, reinforced by a Prior-Guided Attention Module (PrAM) to capture multi-view dependencies across current and prior scans. In Stage 2, GraphRAG-guided retrieval augments intermediate textual representations with graph-grounded clinical evidence, and a large language model restructures them into concise, interpretable question-answer (QA) pairs with diagnostic summaries. Experiments on two public chest X-ray benchmarks demonstrate that CliQ-RRG consistently outperforms prior methods across both lexical and clinical metrics, generating accurate and clinically coherent QA-style radiology reports. Code is available at: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13059
PDF
-
Unified Generative Video Compression with Autoregressive and Bidirectional Video Diffusion
2026-10-09 05:25:27Tianyu Zhang, Shuai Huo, Dengchao Jin, Hewei Liu, Meng Lei, Tianwen Wu, Jiawen Gu, Chao Zhou, Li Li, Dong LiuAbstract
Video diffusion models can reconstruct visually compelling content from severely compressed representations at ultra-low bitrates, but they impose a rigid deployment trade-off. Bidirectional (BI) models achieve high quality through dense full-sequence attention, yet require the entire sequence before reconstruction and incur substantial latency. Autoregressive (AR) models enable low-delay streaming, but their restricted temporal context reduces compression efficiency. This tension is particularly relevant when a video stream must be decoded in real time yet is also stored for later, quality-oriented playback, making AR and BI models natural choices for the live and replay stages, respectively. To support both stages, we present UniGVC, a Unified Generative Video Codec that supports two decoding modes from a single bitstream. UniGVC-BI employs a BI-DiT with dense attention for quality-oriented random-access decoding, whereas UniGVC-AR uses windowed AR-DiT for low-delay streaming. Both paths share a hierarchical latent codec that recurrently produces the same conditions with ultra-low bitrates, allowing UniGVC-BI to serve as a conditional teacher for single-step distillation of UniGVC-AR. A tiny conditional video decoder further reduces streaming latency while preserving most of the reconstruction quality. On standard 1080p benchmarks, UniGVC-BI achieves the best overall rate-distortion-perception trade-off among the evaluated methods and produces temporally stable video at rates as low as 0.0005 bits per pixel. Meanwhile, UniGVC-AR remains competitive while decoding 1080p video at 36.85 FPS, over $250\times$ faster than UniGVC-BI. Code and models will be released later.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13023
PDF
-
Learning Features, Not Layouts: Feature-Axis Attention as a Transfer Prior for Medical Imaging
2026-10-09 05:16:26Aon Safdar, Mohamed SaadeldinAbstract
Natural-image-pretrained foundation models are increasingly adapted to medical imaging, yet transfer remains unreliable under domain shift, modality heterogeneity, and label scarcity. We trace part of this fragility to an overlooked design choice. Standard vision transformers place all global modeling on self-attention over spatial tokens, an object-layout prior well suited to natural scenes but not to medical images, where the discriminative signal lies in feature correlations and texture that recur across scanners while layout does not. We ask whether the axis of global attention, spatial versus feature, is itself a lever for transferable medical representations. Existing channel-aware hybrids confound this question, adding channel attention to a spatially-mixed backbone and leaving the axes coupled and the feature axis unmeasurable in isolation. We design FAD-ViT, a backbone that strictly decouples the two axes, making feature-axis attention the primary global operator and the attention axis a controlled variable. Evaluating comprehensively across 23 classification, segmentation, and low-data tasks under matched tokenization, capacity, and protocol, we find that shifting global attention from the spatial to the feature axis leaves from-scratch performance essentially unchanged, yet markedly strengthens transfer. Identical ImageNet pretraining yields up to roughly three times the transfer gain, with consistent improvements on every held-out segmentation task at up to a quarter the parameter count and far lower compute than standard and channel-attention baselines. The advantage is largest where texture rather than geometry is discriminative, tracking greater reuse of pretrained features. The axis of global attention is thus a design lever for medical transfer, and making it channel-primary is a simple, cheap, and principled choice.
Abstract (translated)
URL
https://arxiv.org/abs/2610.13012
PDF
-
Language Models Have Interpretable Narrative Mechanisms
2026-10-09 04:45:52David Yiyang Liu, Aditya Joshi, Paul Dawson, Andrew PiperAbstract
We investigate whether language models contain internal components that contribute to representing narrativity, the degree to which a text exhibits narrative properties. Across GPT-2-Small, Gemma-2-2B, and Llama-3.1-8B, we identify attention heads, MLP neurons, and sparse autoencoder features whose activations correlate with narrativity. Ablating selected neuron and feature sets increases next-token prediction loss more for high-narrativity than for low-narrativity texts, and reduces the narrativity of open-ended generations. On downstream benchmarks, the largest effects occur in story completion and narrative question answering, although they vary across models and languages; results on confabulation and hallucination benchmarks are mixed. These findings provide causal evidence that selected sets of narrative-associated components contribute to model behaviour, while their internal organisation, relation to narrative theory, and broader functions remain open questions.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12989
PDF
-
In-Context Learning Can Help Vision-Language Models Overcome Training Priors
2026-10-09 03:40:07Kun Wang, Xindi Wu, Sanghyuk Chun, Olga Russakovsky, Esin TureciAbstract
Vision-language models (VLMs) learn strong statistical regularities during training, which can make them fail to properly respond to visual evidence when input images violate those regularities. It is challenging to disambiguate these prior-induced errors from errors caused by the model missing the underlying visual capabilities. In this work, we use controlled visual in-context learning to encourage the model to focus on the visual evidence, and demonstrate that this alone is frequently sufficient to correct the errors -- suggesting that the models in fact do possess the relevant visual capabilities. Specifically, we show that across five vision-centric tasks, demonstrations yield large gains on prior-conflicting counterfactual examples ($+27.2\%$ for Qwen3-VL 235B and $+15.1\%$ for Qwen3-VL 32B) while leaving accuracy on prior-aligned examples nearly unchanged. These findings generalize across five VLM families with gains from $+9.7\%$ to $+30.3\%$. We analyze visual representations and find that demonstrations indeed redirect attention toward grounding cues while reducing reliance on cues that support the learned prior. Finally, we find that in-context learning is limited when the underlying visual capability is weak or absent. Together, these results show that current VLMs possess visual grounding capabilities that standard evaluations may overlook, underscoring the need for controlled evaluation design in diagnosing and improving model capability.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12942
PDF
-
What Egocentric Video Embeddings Encode and How to Combine Them with Inertial Sensors for Human Activity Recognition
2026-10-09 03:15:39Kansuke HabanoAbstract
Egocentric video and body-worn inertial sensors are increasingly recorded together, and the video is usually encoded by pretrained models. What these embeddings represent when the camera sits on the head, and how they should be combined with inertial sensing, has received little attention. Across VideoMAEv2, CLIP, DINOv2 and the egocentric VC-1, the leading principal directions of the embedding are dominated by where the camera looks. Measured head orientation explains them on Ego-Exo4D for egocentric (R^2 0.49-0.70) but not exocentric cameras, and the scene layout explains them on WEAR. The part that the limbs' inertial signals predict separates activities across people. In a set-level probe, the rest separates exercise sets within a participant but not activities across people. This is consistent with separate roles: the video finds the clips of the same set, and the inertial classifier labels them. On WEAR, with temporal order removed and one random limb per one-second clip, we propagate inertial predictions over a within-participant video graph whose leading components are removed. This raises subject-wise cross-validated macro-F1 from 0.648 to 0.852. In our comparisons, propagation helps only over the video graph, and fusing the modalities in one classifier, or adding inertial edges to the graph, performs worse.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12926
PDF
-
Win-Win Fusion: Multi-Modal Image Fusion with Visual-Semantic Collaboration
2026-10-09 02:22:40Hao Zhang, Xuhui Zuo, Jiayi MaAbstract
Existing multi-modal image fusion (MMIF) methods generally face an optimization gap between the visual and semantic domains, limiting the enhancement of perception capabilities. To overcome this long-standing barrier, this work proposes W2F, a new visual-semantic mutual reinforcing paradigm for MMIF. First, we reinforce the foundation for visual-semantic collaboration by amplifying feature complementarity through two targeted designs. Specifically, the segmentation-guided fusion loss introduces object-aware structural guidance into visual fusion, while the embedded information restoration prior strengthens the ability of visual fusion to provide reliable perceptual evidence for the semantic branch under degraded conditions. Then, we introduce a strong multi-modal information interaction mechanism composed of interactive gated mixed attention (IGM-Att) and progressive coupling attention (PC-Att). These modules ensure comprehensive scene representation from both local and global perspectives, providing informational support for visual-semantic collaboration. Finally, we construct a collaborative mechanism to couple visual fusion and semantic fusion. It employs a feature search module to extract complementary key features from both tasks, enriching the representational potential of visual and semantic information. Meanwhile, a gradient correction strategy ensures collaboration from the optimization perspective, enhancing performance for both visual and semantic fusion through consistency learning. Extensive experiments demonstrate that W2F surpasses state-of-the-art methods in both visual appeal and decision accuracy. The code is publicly available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12887
PDF
-
LeapBot-JAC: Joint Action Chains for World Action Models
2026-10-09 02:21:19Chenhao Zhu, Baibing Liu, Pei Liu, Rui Yang, Lang Zhang, Mingyue Feng, Yaonong Wang, Jun MaAbstract
Action-chunking policies predict future controls but execute only a prefix before replanning. How should a fixed action-token budget be organized to support execution? We propose LeapBot-JAC, which allocates sparse future-control references and dense execution targets within one shared action expert. Both chains are supervised by demonstrated controls and jointly denoised at aligned physical-time positions. Directed attention allows execution tokens to read reference tokens subject to the host's visibility constraints. JAC adds neither action tokens nor parameters and retains the host's world-modeling process; both chains participate in inference, while only the execution chain drives the robot. On LIBERO, JAC reaches 98.3% success with FastWAM and LeapBot-WA, up from 97.0% and 97.3%. On zero-shot LIBERO-Plus, their success rates improve from 53.5% to 70.4% and from 73.1% to 77.3%, respectively. RoboTwin and two real-world tasks provide additional deployment evaluations. Single-training-seed comparisons and FastWAM ablations support action-role organization as a practical design choice under a fixed action-token budget.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12886
PDF
-
Time Delay System for Sequence Modeling
2026-10-09 01:43:37Youshan Zhang, Jialu LiAbstract
Existing sequence models share a critical limitation: they lack an explicit, interpretable mechanism for content-aware temporal recall. LSTMs compress everything into a single hidden state, thereby losing information about specific past events. Transformers attend to all tokens indiscriminately at quadratic cost, while state-space models like Mamba compress history into a fixed-dimensional state without explicit content-addressable memory. To address this gap, we introduce a novel framework grounded in nonlinear time-delay systems, in which the effective delay is dynamically estimated based on the semantic similarity between current and historical inputs. Our approach features three key innovations: (i) a differentiable semantic lag estimator that computes content-aware soft attention weights over history, producing a weighted historical context and similarity score; (ii) a diagonal state space with closed-form cumulative sum solution, achieving O(1) per-step complexity with no sequential loops; and (iii) an interpretable fuzzy semantic modulation mechanism that learns context-dependent rules for balancing instantaneous processing versus historical recall, where rule centers correspond to prototypical semantic contexts. Experimental results on text classification and generation outperform state-of-the-art methods.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12864
PDF
-
MS-GLA: Multi-Scale Gated Linear Attention for Addressing Representational Bottlenecks via Multi-Temporal Resolution
2026-10-09 00:44:36Prasoon Dev, Anirudh Sankar, Vasudeva VarmaAbstract
Gated Linear Attention (GLA) Transformers advance linear recurrent models through data-dependent gating, but face a core limitation: the fixed-capacity memory matrices across all heads operate at a single temporal resolution, where each token is processed individually, forcing them to simultaneously encode local syntactic patterns and long-range semantic structure, creating a representational bottleneck that gating alone is insufficient to resolve. We introduce Multi-Scale Gated Linear Attention (MS-GLA), which addresses this by distributing attention heads across multiple temporal resolutions. Coarser resolutions pool longer token spans naturally specializing toward long-range dependencies, while finer head groups retain sensitivity to local syntactic structure. A learnable, input-dependent fusion layer dynamically recombines head group outputs at each timestep, expanding effective memory capacity without increasing per-head state size. This multi-resolution decomposition draws on principles from Multi-Scale State-Space Models (MS-SSM), adapting them to the gated linear attention setting. We evaluate MS-GLA on language modeling, recall-intensive tasks, and long-context generalization. Across all settings, MS-GLA consistently achieves higher accuracy and lower perplexity than GLA at matched parameter counts, with up to 18.9% improvement on recall-intensive tasks and 9.5% lower average perplexity on language modeling benchmarks, validating multi-temporal resolution decomposition as a principled and effective extension of Gated Linear Attention.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35664
PDF
-
KITE: A Knowledge-Injected Temporal Encoder for Time-Series Event Forecasting
2026-10-08 21:35:42Jonas Petersen, Pratik Jawahar, Riccardo Maggioni, Gian-Alessandro Lombardi, Matteo Cozzi, Marek Gajewski, Federico Martelli, Camilla Mazzoleni, Philipp Christian PetersenAbstract
Forecasting rare, high-cost events in multivariate time-series systems is hard because their precursors are seldom observed, so a purely data-driven model has little to learn from. The engineers who run these systems, however, hold knowledge that precedes such events, information absent from the scarce labels. We call this a semantic prior and inject it into a self-supervised event forecaster through two switchable mechanisms, a covariate mechanism and an attention-bias mechanism, each the identity at its null, so one architecture serves every prior type with no per-dataset architecture search. A prior helps when it supplies information the observed window does not already contain: in a short context an informative prior beats no prior in-domain (pooled h-AUROC over the no-prior baseline: a periodicity clock +14%, p<0.001; a conservation residual +47% at a 0.25 label fraction), and a working prior transfers out-of-domain (+53%), substituting for feature transfer, while priors the window already recovers stay inert and the attention-bias mechanism shows no measurable gain. KITE is a switchable, attributable route to contextual conditioning, and a substrate for scaling toward cross-domain, pretrained time-series foundation world models that carry their governing structure explicitly.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12738
PDF
-
Zebra3D: Efficient Hybrid Foundation Models For Scalable 3D Reconstruction
2026-10-08 21:09:21Utkarsh Saxena, Ruijing Yang, Mehdi Rezagholizadeh, Akash Haridas, Vikram Appia, Emad BarsoumAbstract
Feed-forward 3D foundation models reconstruct scenes from unposed images, but the quadratic cost of global attention limits their scalability to large view sets. We present Zebra3D, a layerwise sparse-linear hybrid attention framework that improves the efficiency of pretrained 3D foundation models without training from scratch. Across VGGT, HunyuanWorld-Mirror, and $\pi^3$, we observe a broadly consistent depth-wise pattern: early stages are relatively tolerant to linear replacement, while sensitive middle-to-late stages tend to exhibit more concentrated attention. Zebra3D exploits this structure by assigning linear-centered attention to tolerant stages and tile-sparse attention to sensitive stages, using landmark summaries to preserve unselected context. Closed-form affine fitting enables training-free conversion, while optional post-conversion training combines task supervision with output and relational feature distillation. Across all three backbones, Zebra3D achieves strong reconstruction and camera-pose accuracy at high attention sparsity, with training-free variants remaining close to their dense counterparts. End-to-end speedups increase with input view count, reaching up to $7.1\times$ at 1,024 views. After post-conversion training, Zebra3D-VGGT reaches 80.5\% RealEstate10K AUC@30 at 94\% mean attention sparsity, within 1.1 percentage points of dense VGGT fine-tuned on the same data and schedule.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12727
PDF
-
Bridging Human and Model Attention: Explainability Analysis of CNN, Mamba, and ViT Architectures with Gaze-Based Validation
2026-10-08 20:33:38Marco Stefani, Edoardo Parpaiola, Matteo Bino, Gianluca Caregnato, Francesco Roncolato, Daniel Fusaro, Loris Nanni, Alberto PrettoAbstract
High-risk applications of deep learning require precise characterization of model decision-making. We conduct a systematic explainability evaluation across Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Vision Mamba (ViM). Unlike predominantly qualitative studies, we quantify the alignment between model attention and human visual perception using eye-tracking ground truth. We formalize this comparison via a standardized normalization pipeline mapping post-hoc attribution maps and gaze heatmaps onto spatial probability distributions. Alignment is evaluated across fine-grained classification, medical imaging, and emotion recognition using architecture-specific XAI methods. We measure fidelity using Jensen-Shannon similarity, Chi-square similarity, and Pearson correlation, three complementary estimators capturing support overlap, mass discrepancy, and structural co-variation. Our results reveal distinct attention dynamics: CNNs tend to exhibit a more global focus, with spatially more diffuse saliency maps that correlate well with human perception in general contexts, while ViTs demonstrate fine-grained, patch-level attention, often focused on discriminative microfeatures. The Mamba architecture shows biased behavior: it achieves strong explainability in medical tasks but exhibits divergent intensity distributions in natural images; our analytical characterization of the metrics makes this dissociation explicit, since near-unit support overlap is attained simultaneously with a vanishing linear correlation. We further analyze the effect of backbone freezing and ensemble strategies on model interpretability. The results highlight how architectural and training choices influence the alignment between model saliency and human visual attention.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12703
PDF
-
Kronecker Product Attention: Memory-Efficient KV Compression via Structured Factorization
2026-10-08 19:23:56Bahareh Nikpour, Hossein Rajabzadeh, Maryam Dialameh, Vahid Partovi Nia, Parsa OmidiAbstract
Efficient autoregressive generation is increasingly limited by the memory footprint of the key-value (KV) cache, which grows linearly with sequence length, batch size, and model depth. Existing approaches such as Multi-Query Attention (MQA), Grouped-Query Attention (GQA), Multi-Head Latent Attention (MLA), and Tensor Product Attention (TPA) reduce this cost through different trade-offs, but often sacrifice expressivity, require additional positional-encoding machinery, or impose restrictive low-rank structure. We propose Kronecker Product Attention (KPA), which represents key and value activations as sums of Kronecker products of compact, token-dependent matrix factors. KPA strictly generalizes TPA: rank-one vector-outer-product factorization arises as a degenerate Kronecker split, while non-degenerate matrix factors can represent richer block-structured patterns. We further show that a structured rotary embedding operator matched to the Kronecker representation preserves the relative-position property of RoPE while rotating only the compact factors, removing the need for a decoupled RoPE branch as in MLA. Empirically, KPA matches or exceeds TPA, MHA, MQA, GQA, and MLA across four model scales and nine downstream benchmarks under identical training conditions, while achieving up to 10x KV cache compression. The Kronecker head split is a tuneable hyperparameter that allows practitioners to navigate the compression-accuracy trade-off at design time without architectural changes.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12657
PDF
