MHeatmap (Mosaic Heatmap) is a visualization technique for weighted rectangular matrices that improves readability by reordering rows and columns using a Two-Walk spectral seriation approach and by scaling cell dimensions proportionally to row and column totals in a Mosaic layout. The Two-Walk method extends traditional bipartite ordering by incorporating length‑two walks, creating similarity links between rows sharing columns and columns sharing rows, and then sorting a Fiedler vector of the resulting graph Laplacian to obtain coordinated orders. The Mosaic layout partitions a unit square so that cell height and width reflect row and column totals, while color encodes the matrix entries, and the authors provide theoretical proofs and efficient sparse computation for the method, demonstrating superior performance on synthetic and real datasets.
By Juntang Wang, Sheng'en Li, Liming Zhu, Yihan Wang, Siyuan Cao, Hao Wu, Man Zhang, Shixin Xu, Dongmian Zou
The paper introduces a progressive permutation alignment technique for implicit neural representations, using layer-wise correlation estimation, Hungarian matching, and temperature-controlled spectral annealing to gradually sharpen the alignment toward a permutation matrix. This deterministic, lightweight method avoids costly gradient optimization and is followed by a final refinement stage that optimizes layer-by-layer weight matching. Experiments on MLPs, CNNs, and NeRFs demonstrate reliable neuron correspondence recovery and reduced loss barriers during weight interpolation.
By Yuta Horioka, Naoya Chiba, Yuki Uranishi
ExGrad introduces explicit clean‑gradient subtraction in adversarial training by constructing clean‑input, adversarial‑input, and adversarial‑weight‑perturbed gradients and solving a globally solvable three‑variable problem to select coefficients. Across CIFAR‑10, CIFAR‑100, and ImageNet‑100, ExGrad consistently outperforms AWP and PGD‑AT, with the strongest gains at larger affine radii. The method demonstrates that negative clean coefficients, when properly tuned, improve robust accuracy without sacrificing clean performance.
By Zhiyuan Ye, Xiangyu Zhou, Ji Qi, Hao Zhang, Yi Zhou
The paper examines post‑hoc debiasing techniques for large pretrained models like CLIP, noting that all such methods rely on identifying and expressing bias features in text. It shows that this reliance can cause unexpected performance drops, such as on the UrbanCars benchmark where methods underperform a vanilla classifier even with ground‑truth bias labels. The authors propose a new approach that identifies cross‑class training pairs based on frozen encoder geometry and learns a single linear map for debiasing, achieving competitive results without bias annotations or large language models.
By Divyaansh Mertia, Chirag Bhutra, Himanshu Shekhar, Shubham Bagwari, Pratik Mazumder
The paper investigates how biases present in in-context examples used by Large Language Models (LLMs) for synthetic tabular data generation can propagate to the generated data, causing global statistical distortions even when the biases are mild. It also explores an adversarial scenario where a malicious actor can inject bias through a subset of examples, thereby compromising the fairness of downstream classifiers for protected subgroups. The authors evaluate prompt-side mitigation strategies, such as preprocessing examples and instruction-based defenses, noting that while these can reduce disparity, LLMs remain highly sensitive to adversarial prompts, revealing a significant vulnerability in data generation pipelines for sensitive domains.
By Pol G. Recasens, Alberto Gutierrez, Jordi Torres, Josep. Ll Berral, Javier Carnerero-Cano, Anisa Halimi, Kieran Fraser
The paper investigates how large language models can accurately copy long repeated strings, a task that standard transformers struggle with due to overrun or truncation. It proposes a two‑layer transformer using RoPE‑style rotations (VO‑rotations) with small prime periods to encode distances modulo primes, enabling the model to distinguish all necessary positions via the Chinese Remainder Theorem and correctly terminate copying. Experiments show that baseline transformers fail on long repeats, while the VO‑based model generalizes to lengths up to about 5005 with near‑perfect accuracy.
By Anxin Guo, Tianlong Huang, Chenxiao Yang, Zhiyuan Li
The paper introduces Multiview Masked Autoencoder (MVMAE), a self‑supervised framework that uses the natural multi‑view structure of radiology studies to learn view‑invariant, disease‑relevant representations by combining masked image reconstruction with cross‑view alignment. An extended version, MVMAE‑V2T, adds radiology reports as an auxiliary text signal to improve semantic grounding while keeping inference vision‑only. Experiments on MIMIC‑CXR, CheXpert, and PadChest show that MVMAE outperforms supervised and vision‑language baselines, and MVMAE‑V2T yields further gains, especially when labeled data are scarce.
By Sonia Laguna, Andrea Agostini, Alain Ryser, Samuel Ruiperez-Campillo, Irene Cannistraci, Moritz Vandenhirtz, Stephan Mandt, Nicolas Deperrois, Farhad Nooralahzadeh, Michael Krauthammer, Thomas M. Sutter, Julia E. Vogt
GLUScope is an open‑source tool designed to analyze neurons in Transformer language models that use GLU‑based activation functions such as SwiGLU. It visualizes the four possible sign combinations—gate and in activation being positive or negative—by providing text examples for each quadrant and indicating their frequency. The resource aims to uncover interpretable behaviors that other tools might miss and is available for researchers to extend with additional models, corpora, or neuron pages.
By Sebastian Gerstner, Leonor Veloso, Hinrich Sch\"utze
arXiv:2605.20197v2 Announce Type: replace
Abstract: Medical concept extraction from electronic health records underpins many downstream applications, yet remains challenging because medically meaning...
By Zhichao Yang, Gregory D. Lyng, Sanjit Singh Batra, Robert E. Tillman
SAIL (Spatial Autocorrelation Index Learner) is a differentiable framework that extends classical spatial autocorrelation statistics, such as Moran's I and Geary's C, to high‑dimensional graph‑structured data. It learns feature projections and inter‑node neighbourhood affinities, producing interpretable local and global spatial coherence descriptors for both supervised prediction and unsupervised discovery of spatial autocorrelation modes and hotspots. Experiments on synthetic data and large‑scale biomedical graphs demonstrate that SAIL recovers traditional statistics under specific settings and can approximate Bayes‑optimal predictors for tasks driven by latent spatial autocorrelation.
By Piotr Keller, Fayyaz Minhas
arXiv:2610.29158v1 Announce Type: new
Abstract: Event-based optical flow estimation achieves high temporal resolution and dynamic range by exploiting asynchronous events, while self-supervised learni...
By Shuangming Yang, Qing He, Shangqi Guo, Badong Chen
arXiv:2610.29244v1 Announce Type: new
Abstract: Thermal cameras enable robots to perceive human activities in darkness while revealing less personally identifiable appearance information and fewer de...
By Han Hu, He Chen
arXiv:2610.28880v1 Announce Type: cross
Abstract: Human video scales easily and does not need robot hardware, but existing methods require depth sensing, overlay the robot onto 2D video, or generate...
By Matthew Strong, Ankush K. Dhawan, Julien Poffet, Giuse Pham, Jinho So, Monroe Kennedy III
arXiv:2602.09850v3 Announce Type: replace
Abstract: Industrial anomaly detection demands precise reasoning over fine-grained defect patterns. However, existing multimodal large language models (MLLMs...
By Peng Chen, Wenqiang Wang, Yunkang Cao, Chengliang Liu, Wei Wang, Mingbo Yang, Li Shen, Wenqi Ren, Xiaochun Cao, Chao Huang
arXiv:2605.08802v4 Announce Type: replace
Abstract: Due to the potential for exploratory reasoning of Latent Visual Reasoning, recent works tend to enable MLLMs (Multimodal Large Language Models) to...
By Ziyang Ding, Linjian Meng, Yiming Wu, Yuhan Li, Yuhao Liu, Zhen Zhao
arXiv:2606.14504v2 Announce Type: replace
Abstract: Physical adversarial attacks on vision systems are typically studied via scene manipulation, such as adversarial patches or projections. A smaller...
By Qinlin He (Brian), Zeming Zhuang (Brian), Yongji Wu (Brian), Lan Zhang (Brian), Xiaoyong (Brian), Yuan
arXiv:2610.26538v2 Announce Type: replace-cross
Abstract: Photo-based estimation of food energy could ease dietary assessment, and general-purpose multimodal large language models (LLMs) make such es...
By Xiaomeng Li
arXiv:2610.30121v1 Announce Type: cross
Abstract: Geometric algebra (GA) unifies geometric primitives and transformations in a single algebraic framework. Neural networks designed around this framewo...
By Tao Zhong, Kaison Fong, Keqin Wang, Christine Allen-Blanchette
arXiv:2512.11761v2 Announce Type: replace-cross
Abstract: Data integration often requires aligning datasets without unique identifiers; for network data this is commonly addressed through graph match...
By Trisha Dawn, Jes\'us Arroyo
The paper analyzes gradient flow in a two‑layer linear neural network as the input, hidden, and output dimensions grow proportionally. Using a closed‑form solution combined with random matrix theory, it derives exact asymptotic spectral distributions for the weight matrices at all times and tracks how the learned subspace evolves. In the limit of small initialization, the study reveals a silent alignment phase where singular directions align quickly while singular values and loss stay unchanged, followed by a visible learning phase on a logarithmic timescale where singular values increase and loss decreases through a saddle‑to‑saddle dynamic that becomes continuous in the high‑dimensional limit.
By Simon Martin (DI-ENS, LPENS, SIERRA), Francis Bach (DI-ENS, SIERRA), Giulio Biroli (LPENS)