Why LLMs Stumble on Sustained Mental Attention

When you paste three years of blood panels and unstructured symptom logs into an automated assistant, you assume the system processes the final paragraph with the same precision as the first. A recent evaluation using standard neuropsychological attention tests demonstrates that this assumption is incorrect. While top-tier language models achieved above 90% accuracy on brief cognitive tasks such as basic color-word matching, their accuracy collapsed toward near-total failure as prompt sequence length and task complexity expanded.

In human psychology, this degradation mirrors executive exhaustion: the inability to maintain selective focus when competing signals accumulate. In neural networks, the failure stems from how transformer attention heads distribute probability weights over long context windows. Rather than isolating critical anomalies buried midway through an extended medical file, the systems dilute their focus across background noise, losing track of instructions they executed reliably seconds earlier.

The limits of long-context triage

This degradation directly impacts anyone building or using automated intake summaries. When an automated triage system evaluates an initial appointment questionnaire, it often performs adequately on concise summaries. Add six months of continuous glucose monitor logs, sleep architecture exports, and a chronological medication timeline, and the reliability of automated extraction deteriorates.

If you use consumer models to organize complex personal medical files, treat them as short-span reference aids rather than exhaustive synthesis engines. Feed your records in isolated fragments of one test or one clinical encounter at a time, check specific figures manually against source laboratory PDFs, and verify every synthetic summary with your primary clinician.