New study warns: the multimodal medical AIs behind report writing and diagnosis can be steered off-course by barely visible tweaks to images. Researchers introduce Medusa, a black-box attack that nudges an input scan so the system retrieves the wrong supporting evidence and then generates misleading text. Medusa learns to make
TL;DR Teaching robots with reinforcement learning hinges on the reward function—the 'scorecard' that tells them what to do. Designing good rewards is hard and time-consuming. LEARN-Opt is a new, fully autonomous approach that lets large language models write, run, and evaluate reward functions from plain-English task
What if an AI could become its own sparring partner? UniGame turns unified multimodal models (ones that both understand and generate across text/images) into their own adversary to fix a core mismatch: understanding prefers compact signals, while generation prefers rich reconstructions. This mismatch can misalign decisions and make models
AI
AI that designs enzymes—and explains why Researchers introduce Genie-CAT, an agentic large language model that acts like a digital lab partner for protein design. * Reads the literature via retrieval to ground its ideas. * Parses 3D structures from the Protein Data Bank. * Computes electrostatic potentials to assess physics. * Uses machine
Robotics
What’s new Meet SENTINEL, an end-to-end model that lets humanoid robots turn plain-language instructions into whole-body motion - no hand-crafted pipelines or teleop in the loop. How it works * Trains on a large dataset built by tracking human motions in simulation and pairing them with text descriptions. * Directly maps
Can we keep open-weight genome AI models safe just by removing pathogen sequences from their training data? A new study tests that idea and finds it's not enough. * Researchers assessed a state-of-the-art genomic language model and showed that post-release fine-tuning on sensitive pathogen data can partially restore risky
AIResearch
Long, well-sourced answers are hard for AI to learn: most training rewards short, easily graded Q&A. This work introduces Reinforcement Learning with Evolving Rubrics (RLER) - grading guides that co-evolve with the model, so feedback stays aligned with what the model actually explores during multi-step research. Built with
AI
Meet a new way to write captions and hashtags for fashion photos—grounded in what is actually in the picture. * Detects multiple garments in an image with a YOLO-based model. * Extracts dominant colors and infers fabric and gender by retrieving similar products with CLIP and FAISS. * Packages these facts as
AI
What if your watch could help estimate your daily calories—not just count steps? Researchers introduce the Nutrition Photoplethysmography Language Model (NPLM), which blends heart-signal data from consumer wearables (PPG) with meal descriptions so AI can reason about both physiology and food. Trained on 19,340 people and 1.1
AI
AI is quietly rebuilding Africa’s digital life—from what we watch to what we buy. A new chapter by Ndaka, Avila‑Acosta, Mbula‑Ndaka, Amera, Chauke, and Majiwa asks: who holds the power? Using six months of participant observation and talks with Kenyan social media users, they show how
AudioAI
Ever wished your mic could zoom in on the right voice, even in a noisy room? This research shows how. An on-device deep learning tracker works with a compact ring-shaped array of tiny MEMS microphones to steer a "listening beam" in real time—locking onto a moving target
AI
What if you could direct a video generator by doodling right on the frames? The paper introduces In-Video Instruction: instead of long, vague text prompts, you add visual cues—overlaid words, arrows, or motion paths—inside the image. Each cue acts as a concrete instruction tied to a specific object.