Parikh Vectors Characterize Cyclic Equalizability
A new result ties a card-cryptography property to simple letter-count vectors.
Research Digest
Preprints and papers, rewritten in plain English.
Optimal Block Withholding Attack on Pay-Per-Share Pools
A new analysis models how miners can maximize profits by withholding blocks in pay-per-share mining pools.
CorrCLIPv2 Uses Unlabeled Images to Boost Open-Vocabulary Segmentation
A new approach extends CLIP's open-vocabulary segmentation by exploiting unlabeled images, potentially reducing the need for costly pixel-level annotations.
SL-Ground: Sentence-Level Grounding for Open-Vocabulary Detection
A new arXiv paper proposes sentence-level grounding to improve fine-grained open-vocabulary object detection.
Wavelet-Based Pooling Method for Ultrasound Lesion Segmentation
A spatially adaptive pooling mechanism replaces standard downsampling in U-Net models for ultrasound lesion segmentation.
PLANE: A Plug-and-Play Fix for Vision-Language Hallucinations
A new method called PLANE targets hallucination in large vision-language models by adding fine-grained visual experts.
Study Questions Patch-Decoding Interpretability in Vision Transformers
A new arXiv paper argues that the common assumption that vision transformers can be explained by decoding individual patch tokens may be an illusion.
Packer identification tools lack semantic validation, study finds
New research highlights how incorrect packer labels can mislead malware analysis and proposes methods to characterize and repair these tools.
Cerium Framework Brings Terabyte-Scale Encrypted Inference to Multi-GPU Systems
A new multi-GPU framework aims to make fully homomorphic encryption practical for large-scale AI inference by tackling its heavy compute and memory demands.
5GDescrambler decodes 5G scheduling data to track users
A new tool locates, descrambles, and decodes 5G scheduling information, enabling real-time user identification and activity tracking via side-channels.
QKD Network Routing: Survey and Optimization Framework
A new paper surveys the unique routing challenges in quantum key distribution networks and proposes a conceptual framework for multi-criteria optimization.
Memory-Induced Tool-Drift: When LLM Personalization Backfires
A new study examines how personality-driven preferences stored in long-term memory can lead LLM agents to misuse their tools.
CHERI-D: An Architectural Extension for Temporal Memory Safety
A proposed extension to the CHERI architecture targets efficient temporal memory safety, a growing priority across software and hardware.
FedPix: A New Approach to Federated Learning That Minimizes Data Leakage
FedPix reframes federated learning by asking what information should leave the client, not just how models are trained.
Event-Driven Memory System Targets Long Video Understanding
A new arXiv paper proposes EMMA, an event-driven multimodal memory system with multi-agent retrieval, to overcome the context window limits of multimodal large language models.
Instance-Specific Guidance Balances Fidelity and Prompt Adherence in Image Editing
A new method adapts classifier-free guidance to each input image, addressing the core trade-off in prompt-based editing of diffusion and flow models.
Decoupling Attention Capacity from Depth in SegFormer Compression
A new study asks whether transformer compression should treat attention capacity and execution depth as separable, rather than coupled, quantities.
EAGLE Enhances Object Localization with Partial Polygon Supervision
A new arXiv preprint proposes an entropy-aware method that uses partial polygon supervision to sharpen object detection at high IoU thresholds.