LLM fine-tuning safety research from Meta AI and CMU pinpoints the harmless-looking training examples that erode AI model safety before deployment. Researchers from Meta AI, Carnegie Mellon University ...

Background Health inequalities are widely studied but how governments operationalise (define and measure) the phenomenon is less well-understood. Existing research focuses mainly on health-led ...

Public FDA shortage records helped a machine-learning model concentrate 40% of 90-day recurrence events in the highest-risk ...

An enhanced asynchronous actor-critic framework trained in simulation and on a physical mobile robot showed faster ...

Enterprise AI agent reliability training with Microsoft and Hugging Face's ThinkingBox shows database-state checks repurposed ...

IntroductionHello! This is AI Nest Inc.In robot imitation learning, behavior cloning, which reproduces expert actions from ...

Perplexity’s JOLT trains one model as teacher and student, matching GRPO’s MATH accuracy with 6.5x fewer training tokens.

Constraining policies to select behavior modules allows reinforcement learning transfer from simulation to real robot swarms.

New GRAFT research uses a deletable adapter to curb AI reward hacking, scaling to Alibaba's Qwen3-32B and cutting hacking to about 1.2% in tests.

Introduction A few years ago, when I was tasked with an in-house PoC for image classification, I took sample code for transfer learning and ran it almost as-is. The training ran. The accuracy stopped ...

A logo of Nvidia is picture before the Nvidia/Japan AI Ecosystem Reception in Tokyo on July 16, 2026. Philip FONG/AFP via Getty Images NVIDIA researchers published and open-sourced FlashREINFORCE on ...

Anthropic has unveiled GRAM (Gradient-Routed Auxiliary Modules), a novel approach to controlling access to dual-use knowledge in AI models, according to a research update published on July 8, 2026.