Frontier large language models routinely ace static medical board exams, yet executing complex, multi-step clinical pathways remains a systemic vulnerability. Researchers Sheng Zhang and YoSean Wang from City University of Hong Kong, Jinming Li from Fudan University Shanghai Cancer Center, Wangyang Chen from Westlake University School of Medicine, and Zhiwei Bao from Zhejiang University evaluated how leading architectures navigate guideline-conformant and patient-specific oncology choices. The research team constructed the Oncology Decision Boundary Benchmark (ODBB), analyzing 2,005 clinical decision points derived from National Comprehensive Cancer Network (NCCN) protocols and complex colorectal cancer cases.
The Failure of Model Ensembles
Testing nine frontier models across four closed-source and five open-weight architectures, the researchers replaced error-prone LLM-as-a-judge scoring with a deterministic, rule-based evaluator categorizing outputs into 14 explicit failure modes. Two independent oncologists validated this benchmark across a stratified 225-item subset. The findings dismantle the assumption that model ensembling solves clinical reliability: pooled together as an aggregate system, none of the models answered correctly on 42.1% of all items.
Performance deteriorated sharply as clinical ambiguity increased, with collective failures jumping from 35.7% on standardized NCCN guideline items to 66.4% on real-world colorectal cancer cases.
"Treating the nine as a pooled super-model, 42.1% (Wilson 95% CI 40.0%–44.3%) of all items—35.7% of the 1,586 NCCN guideline items and 66.4% of the 419 colorectal cancer case items—were answered correctly by none"
These shared breakdowns clustered around clinical meta-judgment—specifically, determining which high-level pathway applies before executing reasoning within it.
Decisiveness Tuning and Commitment Failures
Tuning models for decisive outputs generated severe safety liabilities. Systems optimized for assertiveness made unsafe clinical commitments between three and five times more frequently than conservative models, yielding zero net gain in overall accuracy. Furthermore, in 3% to 9% of benchmark scenarios, models correctly articulated the necessary clinical intervention in their chain-of-thought rationale but failed to commit to it in their final recommendation. These breakdowns represent failures of decision execution rather than knowledge retrieval.
Raw parameter scale and benchmark scores no longer represent the primary bottleneck for medical AI. For HealthTech executives and clinical deployers, treating an LLM as an autonomous decision-maker is an untenable risk. Enterprise deployment demands deterministic verification rails that actively detect model competence boundaries and enforce human-in-the-loop escalation before any clinical recommendation reaches a patient.