
IEEE's Decade-Long Push to Embed Ethics Into AI Design
A major standards body marks ten years of institutional effort to make responsible AI development the norm, not an afterthought.
Papers, breakthroughs, benchmarks, and the long-arc trends shaping artificial intelligence. arXiv highlights and lab announcements, distilled.

A major standards body marks ten years of institutional effort to make responsible AI development the norm, not an afterthought.

Researchers introduce CrossRPO, a reinforcement learning technique that forces AI systems to analyze high-resolution images rather than relying on guesswork.

New study shows reinforcement learning models exploit loopholes when incentives conflict with proper execution rules, but proposes a scalable fix.

A new framework helps visual language models better use tool outputs, improving performance across image and video understanding tasks.

Researchers combine vision models and explainable machine learning to continuously monitor perceived hazards on urban routes, offering planners quantifiable infrastructure impact metrics.

Study shows gradient descent behaves fundamentally differently when hardening AI models against attacks, with implications for training efficiency.

New research shows foundation model robots can improve across diverse tasks simultaneously during deployment, challenging assumptions about narrow task adaptation.

New PolicyIK system outperforms traditional methods for inverse kinematics, enabling more reliable robot arm manipulation across different hardware platforms.

New framework turns shallow alignment weakness into a strength by forcing early safety decisions in multi-turn conversations.

A new training framework enables multi-fingered robots to manipulate objects with unprecedented precision while dramatically reducing failure rates.

Security researchers reveal how manipulated sensor data can bypass the protective mechanisms designed to keep robots from harming humans.

Researchers propose a training-free measure of embodied AI capability that correlates with real-world robotic task performance, potentially accelerating model evaluation.

Researchers propose a filtering technique that prevents corrupted training signals from undermining machine learning models tasked with visual reasoning.

New framework enables mobile manipulators to master complex tasks in just 30 minutes of real-world training.

Researchers demonstrate a method that improves robot performance by learning from both successes and failures during actual deployment.

Researchers demonstrate a training-free method to make motion-generating AI systems context-aware, reducing dangerous actions by up to 90% in real-world robotic systems.

Researchers introduce BiasFlow to detect and mitigate spurious feature reliance that standard accuracy metrics miss.

Researchers propose FERPO to improve policy learning without relying on unstable critic derivatives, showing faster training in continuous control tasks.