A new arXiv preprint accepted at NeurIPS 2026 reports that multimodal AI models become significantly worse at refusing harmful requests when they are given tools such as zooming and tagging. The paper's authors propose two possible explanations, and the findings carry implications for anyone building or regulating agentic AI systems.
A new benchmark from two Berkeley researchers grades AI models on schematic transcription and device sizing using 273 topologically distinct open-source circuits from Tiny Tapeout manufacturing shuttles. The results reveal sharp limits in visual reasoning, and a refusal problem that has nothing to do with skill.
A NeurIPS 2026 paper argues that the dominant design of training-free image safety filters, which removes one reusable unsafe subspace across all prompts, faces a structural coverage-selectivity trade-off. Its proposed fix, CALM, routes each prompt locally. This analysis explains the argument for general readers, and flags which parts are demonstrated versus self-reported.