Why LLMs Lose Their Focus During Lengthy Health Consults

When you hand an LLM health copilot a single lab panel or ask it to define an isolated biomarker, it usually delivers crisp reasoning. When you keep feeding it a decade of clinical notes, complex medication schedules, and evolving symptom diaries in a single continuous session, its reliability drops off a cliff.

Cognitive psychologists recently subjected leading language models to the classic Stroop task, a standardized psychometric protocol used to measure selective attention and cognitive interference. In short trials, top-tier architectures achieved accuracy scores above 90%. As researchers steadily increased trial sequence length and interference complexity, model performance decayed rapidly, sliding near complete operational failure.

The mechanism comes down to self-attention mathematics. Transformer architectures distribute attention weights across all tokens in their working context window. As irrelevant tokens, conflicting temporal markers, and conversational tangents pile up, the attention weights dilute. The model begins favoring local syntactic patterns over distant semantic constraints, missing critical contradictions tucked away twenty prompts earlier.

If you feed your own blood panels or doctor notes into a language model, break your inquiry into isolated, modular questions. Clear the chat history between distinct medical topics, verify that the model retains core boundary constraints, and treat multi-hour conversational triage as prone to mathematical attention drift.