VibeGen: AI model that Design Proteins by Motion, Not Just the Shape
VibeGen uses agentic AI to design de novo proteins based on vibrational motion. This 2026 study reveals that diverse protein folds can share similar d...
Meta-Harness: Automated LLM Code Optimization Breaks Barriers
Meta-Harness automated LLM harness optimization code improves text classification by 7.7 points and math reasoning by 4.7 points using filesystem acce...
TurboQuant: Near-Optimal Vector Quantization with Zero Indexing Time
TurboQuant online vector quantization distortion rate optimization achieves near-optimal results within a 2.7 factor of theoretical bounds. It ensures...
AI Has a Secret 'Assistant Mode' and Scientists Found the Off Switch
Anthropic researchers found a hidden 'Assistant Axis' in AI that, when tampered with, cuts harmful responses by nearly 60%. Oops.
Anthropic's Assistant Axis: How LLM Persona Drift Causes Harmful AI Behavior
Researchers at Anthropic and Oxford identified a linear 'Assistant Axis' in LLM activation space that governs persona stability. Activation capping al...
First AI System Predicts 130 Diseases From a Single Night of Sleep
SleepFM AI model trained on 585,000 hours of sleep data predicts 130 diseases including dementia and heart failure from one overnight recording.
Prompt Repetition Improves Non-Reasoning LLMs Without Added Latency
Repeating the input prompt improves accuracy across Gemini, GPT, Claude, and Deepseek models in 47 out of 70 benchmarks with zero losses and no added...
SleepFM AI Model Predicts 130 Diseases From One Night of Sleep Data
SleepFM, a foundation model trained on 585,000+ hours of sleep recordings, predicts 130 diseases from a single night's polysomnography with C-Index sc...
Giving AI a Digital Playground to Think Better
Scientists gave AI a virtual computer playground. Now the AI can do math, read huge books, and even make music all by itself!
LLM-in-Sandbox: How AI Agents Use Virtual Computers to Boost Performance
Researchers created LLM-in-Sandbox, a framework that gives language models access to a virtual computer where they can execute commands, manage files,...
Redefining AI Perception: Introducing VL-JEPA, a new vision-language model
Meta's VL-JEPA learns by predicting semantic embeddings instead of generating text token-by-token, allowing it to focus on meaning rather than phrasin...
Meta's VL-JEPA: A Smarter Way for AI to See?
Meta's VL-JEPA represents a shift in vision-language AI by predicting abstract meaning (embeddings) instead of generating text word-by-word, focusing...