A new study from arXiv proposes BettiSafe, a topology-guided split selection strategy for privacy-aware split learning that improves resistance to feature inversion attacks by 2 to 5 times compared to depth-based heurist…
Researchers propose CausAdv, a causal framework for detecting adversarial examples in convolutional neural networks (CNNs) by learning causal and non-causal features and quantifying counterfactual information (CI) per fi…
Researchers introduce EgoPlay, an event-triggered video-to-video editor for egocentric streams that jointly learns event recognition, temporal restraint, and pixel-level editing in a single end-to-end model. On the Ego4D…
Researchers introduced the Multimodal Interactive Motion Encoder (MIME), the first dedicated multimodal encoder for two-person interactive motion, achieving a 12.8% relative improvement in text-to-motion R@1 on the Inter…
A new approach to Hardware Trojan detection uses And-Inverter Graphs (AIGs) and Knowledge Graph Embedding (KGE) to scale to modern System-on-Chips with billions of gates, achieving linear complexity and clear geometric s…
Researchers propose Delta-InvFormer, a transformer-based surrogate model that uses visible-light video frames to predict two-dimensional light intensity distributions in tokamak divertors, accelerating traditional method…
Researchers show for the first time that visual token pruning enhances the robustness of Multimodal Large Language Models (MLLMs), achieving an average improvement of 13.29% in defending jailbreak attacks while also redu…
Researchers introduce MegaSlide-DiT, a prototype that adapts a pre-trained 105-billion-parameter Diffusion Transformer (DiT) for high-resolution video generation on a single Nvidia H200 GPU with 1.5 TB of host RAM, by ke…
Researchers introduce FogDrive, a multi-modal synthetic driving dataset for autonomous driving under fog, containing 660 scenes (~133k frames) with four synchronized cameras, LiDAR, and radar across three calibrated fog …
Researchers developed the Language-guided Segmentation-assisted Diagnostic Transformer (LSDT) model, which uses zero-shot segmentation and multi-modal slice fusion to classify prostate cancer into four clinically meaning…
Researchers introduce γ-Bridge, a look-parametric diffusion bridge for SAR despeckling that uses a schedule L(t) connecting noisy observations at L_obs to the clean limit via exact multiplicative Gamma marginals. The met…
Researchers have introduced CAVSR, a benchmark of 605 real-world AIGC videos from multiple platforms, and a hierarchical risk taxonomy with 6 top-level categories and 26 fine-grained labels to systematically evaluate chi…
Researchers introduce TOM-GS, an editable video representation that uses regular 3D Gaussians with a continuous temporal opacity formulation, avoiding complex deformations. The method outperforms prior editable video rep…
Researchers propose DAMamba-UNet3D, a parameter-efficient state-space model U-Net for 3D medical image segmentation that achieves a mean Dice of 0.815 on BraTS 2020 with ~5.3 million parameters, roughly 13 times fewer th…
Researchers propose an automated framework using Vision-Language Models to verify actions in cognitive remediation tasks for schizophrenia patients, aiming to replace subjective manual observation by clinicians. The syst…
A new classification-guided large vision-language model (LVLM) framework for visual information extraction (VIE) from documents achieves 86.43% F1-score in zero-shot inference on a real-world bidding dataset with 16 cert…
Researchers introduced LC-SEPLM (Long-range Contact-supervised ESM Protein Language Model), which adapts ESM2 with LoRA and long-range residue-pair contact supervision while retaining sequence-only downstream inference. …
Semalith v1.4, a 184M-parameter DeBERTa-v3-base safety classifier, achieves state-of-the-art prompt-injection detection at 44x fewer parameters than Llama-Guard-3-8B, winning all 7 prompt-injection evaluations and 11 of …
Researchers introduced CausalGate, an intervention-guided framework for transformer module pruning that measures semantic damage via Kullback-Leibler divergence, outperforming dynamic routing baselines on TinyLlama-1.1B,…
Researchers introduced EMG-CrossFormer, a hybrid convolutional-transformer model for multimodal surface EMG hand gesture recognition, achieving mean accuracies of up to 92.79% on NinaPro datasets when fusing sEMG with in…