CogAttentionBench: Probing Cognitive Attention Mechanisms in Frontier AI Models

CogAttentionBench is a benchmark of 38 items across five tasks grounded in established experimental psychology paradigms for evaluating cognitive attention in frontier AI models: selective attention (Stroop / flanker), attention shifting (task-switching), sustained attention (CPT), inattentional blindness (gorilla paradigm), and saliency awareness (feature integration). Unlike needle-in-a-haystack tests that rely on explicit markers, CogAttentionBench creates genuine interference by exploiting LLM-specific processing biases (statistical-frequency traps, context priming, expectation violation) that act as analogs to the automatic processing underlying human interference effects. This deposit contains the dataset files, Croissant 1.0 JSON-LD metadata, README, the source markdown of the NeurIPS Datasets & Benchmarks 2026 submission, the compiled PDF, and the LaTeX source. Companion theoretical framework: Attention as a Magnetic Field (Zenodo: 10.5281/zenodo.18979607).