The AI knowledge 99.9% of people never discover
New arXiv preprint shows Transformer training on inductive reasoning can be confined to a low-dimensional manifold for automatic circuit detection.
Reasoning scaffold lifts GPT-4.1-mini by 0.21 but degrades GPT-5-mini by 0.63, arXiv study finds, exposing an architecture-dependent split.
A new arXiv survey of 60+ papers finds federated learning for vehicle intrusion detection relies on artificial data splits and weak attack tests.
New arXiv preprint introduces AHA, a system using one AI agent to auto-discover reusable vulnerabilities in production agents like Claude Code and Codex.
FactorDiff decomposes diffusion samples into pixel-level factors and routes each to the best expert, beating global weighting on ARC-AGI reasoning tasks.
AMT-X multi-turn red-teaming framework hit up to 100% attack success on six frontier AI models, exposing critical gaps in current LLM safety testing.
A field experiment across six companies found GPT-5's playful rewriting lifted emotional tone, with engagement gains flowing through an indirect path.
PromptGraph models LLM prompts as graphs to catch privacy leaks that existing sanitisers miss, affecting anyone sending sensitive data to cloud AI.
ARMOR-IMC, an arXiv preprint, hardens in-memory computing chips against manufacturing faults and power attacks without retraining the AI model.
