AI Safety & Robustness

2 summarised stories about AI Safety & Robustness, each linking back to the original source. Browse all topics →

Thursday, 16 July 2026

PC-Diffuser: Path-Consistent Capsule CBF Safety Filtering for Diffusion-Based Trajectory Planner

arXiv cs.AI 18 hours ago

Researchers developed PC-Diffuser, a safety framework that embeds barrier functions directly into diffusion-based trajectory planning for autonomous driving to prevent collisions. The method applies safety corrections at each denoising step using capsule-distance functions and kinematic constraints while preserving the learned behavior distribution. This approach enables diffusion models to generate certified safe trajectories without post-hoc correction that may distort the learned planning patterns.

AI & Autonomous Vehicles AI Safety & Robustness AI Trajectory Planning Diffusion Models

Ask Before You Diagnose: Safe-Psych, a Sequential Evaluation Benchmark for LLMs in Psychiatry

arXiv cs.AI 18 hours ago

Researchers introduced Safe-Psych, a benchmark for evaluating how large language models handle incomplete clinical information in psychiatry. The benchmark contains over 1,000 real-world psychiatric notes segmented to simulate evidence disclosure over time, with psychiatrist-labeled actions (DIAGNOSE, CLARIFY, or ABSTAIN) at each stage. Testing multiple state-of-the-art models revealed that under-abstention exceeded 60% for most models, with premature diagnoses being less accurate than those made with sufficient evidence, indicating that models struggle to recognize when clinical information is insufficient and additional clarification is needed.

AI Benchmarking AI Decision Making AI & Healthcare AI Safety & Robustness