Neural networks can be fooled by adversarial images—but here’s a twist: those fake images are even more fragile than the real ones, especially when parts are hidden. Researchers tested nine popular attacks (like FGSM and PGD) on CIFAR-10 and slid a small mask across each image while watching
Most pathology AIs read slides patch by patch, often missing the subtle shapes of individual cells. JWTH (Joint-Weighted Token Hierarchy) changes that by fusing cell-level clues with whole-tissue context in one foundation model. * What’s new: a cell-centric post-tuning step and attention pooling that combine local and global tokens. * Why
As more networks switch to Software-Defined Networking (SDN), keeping them both safe and speedy is a challenge. SmartSecChain-SDN is a new framework that tackles both. It blends three pieces: * AI intrusion detection using Random Forest, XGBoost, CatBoost, and a CNN-BiLSTM to spot attacks in real time with high precision. * Tamper-resistant
AI that writes your architecture docs—from your code Architecture docs often lag behind reality. This research shows a semi-automated way to generate Software Architecture Descriptions (SADs) directly from source code. * Recover the system’s structure with reverse engineering to build a high-level component diagram. * Use a large language model
AI agents are starting to work in teams. That unlocks power—and new ways things can go wrong. TAMAS is a benchmark that stress‑tests multi‑agent LLM systems against adversarial tricks and coordination failures. * 5 realistic scenarios, 300 attack instances across 6 attack types * 211 tools, plus 100 harmless
AI
AI that sees, thinks—and uses tools Meet DeepEyesV2, a multimodal “agentic” model that doesn’t just read text and look at images—it can call external tools like code runners and web search, then weave the results into its reasoning. Key ideas: * Two-stage training: a cold-start phase teaches basic
AI
Meet your AI public-speaking coach Engineers need to explain complex ideas clearly, but personalized practice is hard to get. This research builds a multimodal AI trainer that analyzes both what you say and how you say it, then delivers targeted feedback at scale. * Verbal: pitch, loudness, pacing, intonation * Non-verbal: facial
AI assistants are not used in a vacuum—they operate inside hospitals, banks, classrooms, and brands, each with its own rules. This paper introduces Pluralistic Behavior Suite (PBSUITE), a testbed to see whether language models can stick to your custom policies over multi-turn conversations. * 300 realistic behavioral policies across 30
AI
Meet LiveStar — an AI co‑host for live streams that understands what’s happening on screen and speaks up at the right moment. Most video AIs do great offline, but stumble live: they process frames slowly and interrupt or fall behind. LiveStar fixes that with: * Adaptive streaming decoding: keeps a
AI
Predictive maintenance means fixing equipment before it fails. But in the auto industry, messy maintenance logs—full of typos, missing fields, near-duplicates, and wrong dates—can derail the machine-learning models that power those predictions. This study tests large language model (LLM) agents as smart cleaners for those logs. The team
AI
Criminal justice data often record charges and outcomes—but not what actually happened. Yet many continental European verdicts spell out the facts. This study shows we can automatically pull those fact sections from public Slovak court decisions. The team tested two tools: 1) smart pattern‑matching that recognizes how courts
AILiteracy
Artificial intelligence is now part of everyday tools. How should community colleges teach it to non-STEM students? Researchers tested AI User, a no-code, scenario-based online curriculum, with four focus groups of community college instructors. * Instructors loved exploratory, real-world tasks that let students experiment safely with AI. * They preferred interactive demonstrations