AI
Teach Robots to Learn Like People Building versatile embodied AI (agents that see, read, and act) is hard because real-world data is scarce and training is expensive. We introduce Deliberate Practice Policy Optimization (DPPO), a metacognitive "metaloop" that alternates between: * Supervised fine-tuning to expand general competence * Reinforcement learning
AI
Smarter, faster explanations for what you’re recommended Big AI models can explain why you’re shown a product or movie—but wiring them directly into recommendation engines often slows everything down and muddles goals. "Prism" takes a cleaner route: it splits the system into two jobs. * Oracle:
AI
Meet CLOZE Clinical notes are full of real-world details, but they’re hard for computers to use. CLOZE is a new AI framework that reads these notes to expand medical ontologies—think of them as maps of diseases, symptoms, and their relationships. * Zero-shot: works without extra training or labeled data.
AI
NutriScreener is a new AI method to help screen for child malnutrition quickly and at scale—using ordinary photos captured from multiple poses. How it works: it combines CLIP-based visual features, a retrieval “knowledge base” that surfaces context from similar cases, and a multi-pose graph attention network that reasons across
Can AIs Be Conscious? A simple map of the debate Researchers Andres Campero, Derek Shiller, Jaan Aru, and Jonathan Simon offer a practical framework to sort arguments about whether digital AIs could be conscious. They organize challenges along two axes: * Level (in Marr's terms): what's being
AI
AI can be impressively accurate in medicine—but dangerously overconfident when cases are ambiguous. MedBayes‑Lite offers a fix: a lightweight add‑on that helps clinical language models say how sure they are. How it helps, without retraining: * Estimates what the model doesn’t know (epistemic uncertainty) using Monte Carlo
Robotics
Robots that learn by watching you What if multi-fingered robots could pick up everyday skills just by watching people in the real world? A new framework, AINA, takes a big step toward that. Using lightweight Aria Gen 2 smart glasses, anyone can record natural demonstrations at home, in the office,
AI
AI images that "think" as they form Most image generators plan before they draw or fix mistakes after. Thinking-while-Generating (TwiG) lets models do something new: interleave short bursts of textual reasoning during generation—like thinking out loud while painting. As pixels appear, the model explains what to add
Big vision models are now trained once and reused via simple "linear probes." This paper asks: can a tiny set of synthetic images replace massive real datasets for training those probes? Enter Linear Gradient Matching: it learns a handful of synthetic images so that, after a frozen feature
AI
Can we quickly tell if an image is AI-made — and even which generator produced it — without retraining huge models every month? This study says yes. The authors propose a two-stage system: * Learn a visual "fingerprint" space. A vision model trained with supervised contrastive learning separates real vs. synthetic
What if your app could not only pick what to recommend, but also explain why—clearly and fast? "The Oracle and The Prism" introduces Prism, a simple idea: separate the job of choosing items from the job of explaining them. Here’s the trick: a large "Oracle&
ASR
Not all transcription mistakes are equal. In clinic visits, the go-to score Word Error Rate (WER) can look fine even when errors change care-critical meaning. * Expert clinicians compared true utterances with ASR transcripts in two doctor-patient datasets, labeling the impact: No, Minimal, or Significant. * WER and many other metrics poorly