Large language models that ace medical exams are failing miserably in real-world clinical settings. A study by Krishan Rajaratnam and Keno Bressem from the Technical University of Munich has exposed a critical flaw: even top-tier proprietary models struggle with agentic clinical reasoning. In a hematology-oncology simulation, the best of 32 tested systems scraped together a mere 68% accuracy. The issue isn't a lack of knowledge, but a fatal inability to operate under uncertainty.
There is a massive gulf between medical "knowledge" and clinical "action." Researchers observed a phenomenon of cognitive laziness: diagnostic accuracy depends directly on a model's ability to proactively request patient data. However, as the process reached the final stages, the AI's drive to seek information effectively collapsed. Instead of extracting critical molecular and cytogenetic markers, the agents demonstrated "premature closure"—issuing a verdict based on incomplete patient profiles.
AI behaves like a struggling student who is too embarrassed to admit they don't know something and simply guesses the diagnosis.
For MedTech founders and Chief Innovation Officers, this is a major red flag: current benchmarks are deceptive regarding AI's readiness for autonomous work. High scores in clinical reasoning have proven to be entirely decoupled from actual diagnostic precision. We are seeing a dangerous trend where a system can sound like a brilliant oncologist while making elementary errors. Without implementing rigorous active data-seeking mechanisms, deploying these agents in a clinic isn't innovation—it's a game of Russian roulette with patient lives.
Key Findings of the Study
Diagnostic accuracy in complex cases does not exceed 68%, even among market leaders. Models suffer from "cognitive laziness," ceasing data collection at critical decision-making points. Existing medical benchmarks fail to reflect the real-world treatment capabilities of AI. The skill of active information gathering is more vital to the outcome than the model's theoretical knowledge base.
Modern LLMs are essentially interns with a professor's vocabulary but none of their clinical grit. If you are building diagnostic tools, the "information retrieval" metric must become more important than raw theoretical scores. Until models learn to fill gaps in a patient's history with targeted follow-up questions, their utility in complex clinical cases remains doubtful and their risks unacceptable.