AI models can't read a city's worth of patient records without choking on the sheer volume of text.
Researchers have published PopMedQA, a benchmark designed to expose what they call the verbose context problem: the gap between how medical data is coded and how language models consume it. Analyzing longitudinal records for a patient cohort can push past 400,000 tokens — well beyond what most models handle efficiently. The team built the benchmark using a new synthetic-record library called neopatient, which generates artificial patient histories under controlled conditions. They then ran a battery of approaches against it: prompt compression, agentic decomposition, and various prompting strategies.
None of them worked well. Every tested method failed to meaningfully close the gap, which matters because population health analysis — spotting disease trends, flagging at-risk cohorts, powering public health decisions — depends on reasoning across exactly these kinds of sprawling, longitudinally-coded datasets. The finding suggests that generic efficiency tricks borrowed from other domains don't translate when the data structure itself is the problem.
The authors conclude that real progress will require techniques that exploit medical-domain structure specifically, not just longer context windows or smarter prompting. That's a narrower and harder research target than the industry usually advertises — and a quiet rebuttal to any lab claiming its frontier model already "handles" clinical data at scale.