Glonce

Nine frontier LLMs pooled together still miss 42% of oncology decisions

researchThe Glonce editorsarXiv🔥78Glonce rating78Impact73Novelty70Relevance100Surprise58Players78Credibility87

Researchers built the Oncology Decision Boundary Benchmark (ODBB), a set of 2,005 oncology decision points drawn from NCCN clinical guidelines (1,586 items) and colorectal cancer cases (419 items), to test whether frontier language models share the same blind spots in clinical decision-making rather than just clinical knowledge. Existing medical benchmarks mostly measure factual recall on exam-style questions; ODBB instead tests guideline-pathway choices, escalation judgments and commitments under uncertainty, the kind of sequential decisions real oncology practice requires.

Nine frontier LLMs released between June 2025 and April 2026, four closed-source and five open-weight families, were evaluated on the full set. A fully deterministic scorer, with no LLM inference involved, classified every output into one of 14 failure types; two independent oncologists validated the scorer against a 225-item stratified sample and agreed with it closely (Cohen's weighted kappa of 0.939 and 0.790).

The authors then treated the nine models as a single pooled super-model, counting an item as unsolved only if every one of the nine got it wrong. On that measure, 42.1% of all items (95% confidence interval 40.0 to 44.3%) were answered correctly by none of the nine models. The failure rate was uneven across the two item types: 35.7% of the 1,586 NCCN-guideline items went unsolved by all nine, against 66.4% of the 419 colorectal-cancer cases. The authors describe the concentration of failures in choosing between guideline pathways before reasoning within any single one as a consistent blind spot in clinical meta-judgment, one they argue likely needs an architectural fix rather than more training data.

Two models tuned for decisiveness, GPT-5.5 and Gemini 3.1 Pro Preview, made unsafe commitments three to five times more often than the seven more cautious models in the set, without scoring any higher overall. Separately, in 3 to 9% of items a model stated the correct next clinical step in its reasoning but then failed to commit to it as the final answer, which the authors classify as failures of decision rather than of knowledge.

The paper's conclusion is that model quality is no longer the primary obstacle to deploying LLMs in clinical settings; the binding constraint is the assumption that any single model can be the sole basis for a clinical decision. The authors argue that progress requires architectures that can detect when a model has reached its competence boundary and route the decision to a clinician instead.

Key facts

  • The Oncology Decision Boundary Benchmark (ODBB) covers 2,005 oncology decision points: 1,586 from NCCN guidelines and 419 colorectal cancer cases.
  • Pooling nine frontier LLMs (four closed-source, five open-weight) as a single super-model still leaves 42.1% of items unsolved by any of them (95% CI 40.0-44.3%); the unsolved rate is 35.7% for NCCN items and 66.4% for colorectal cases.
  • A deterministic, non-LLM scorer sorted failures into 14 types and matched two oncologists' judgments closely (Cohen's weighted kappa 0.939 and 0.790) on a 225-item validation sample.
  • GPT-5.5 and Gemini 3.1 Pro Preview, tuned for decisiveness, made unsafe commitments three to five times more often than the seven more cautious models without any gain in accuracy.
  • In 3-9% of items models identified the correct next clinical step but failed to commit to it, a decision failure distinct from a knowledge gap.

Why it matters

Most clinical AI benchmarks test medical knowledge recall, the kind of thing exam-style questions capture well. ODBB instead probes sequential decision-making: choosing between guideline pathways, escalating appropriately, and committing to a course of action under uncertainty. The paper's headline finding is that this is a shared blind spot, not a gap that varies model to model. Pooling nine frontier models and counting an item solved only if at least one gets it right still leaves 42.1% of items unsolved by any of them. That means the errors overlap heavily rather than being idiosyncratic, and the authors argue the concentration in pathway selection specifically points to something architectural in how current models reason, not simply a training-data shortfall.

Who it affects

The finding bears on anyone building or evaluating clinical decision-support tools on top of frontier LLMs, and on oncologists or health systems considering LLM-assisted guideline navigation. It also matters for model developers: the two models built for decisiveness, GPT-5.5 and Gemini 3.1 Pro Preview, made unsafe commitments three to five times more often than the seven more cautious models in the study without any accuracy gain, which is a direct caution against tuning clinical-facing models toward confident answers.

How to use it

The paper's own recommendation is architectural rather than a specific product: build systems that can detect when a model has reached the edge of its competence and route the decision to a clinician, rather than treating any single model's output as sufficient for a clinical decision. The failures-of-decision finding, where a model states the right step but does not commit to it, also suggests that output-formatting or commitment-forcing techniques could recover some accuracy without touching the underlying reasoning.

How solid is it

The benchmark is sizeable (2,005 items) and the scoring is deterministic rather than LLM-judged, which removes one common source of evaluation noise. The scorer was checked against two independent oncologists on a 225-item stratified sample, with high agreement (Cohen's weighted kappa of 0.939 and 0.790). The text does not state the paper's publication venue or peer-review status, nor does it name the nine evaluated models beyond GPT-5.5 and Gemini 3.1 Pro Preview, nor how the 1,586 NCCN and 419 colorectal-cancer items were selected.

Risks and caveats

The 42.1% figure describes a pooled super-model that does not exist in deployment; it is a ceiling on what a best-of-nine ensemble could achieve, not a working system, and real-world use of any single one of these models would face a higher failure rate. The paper also does not specify a timescale or concrete design for the boundary-detection architectures it calls for. The sharp difference between the NCCN unsolved rate (35.7%) and the colorectal-case unsolved rate (66.4%) suggests the difficulty is uneven across item types, so the 42.1% headline should not be read as uniform across all of oncology decision-making.

“Model quality is no longer the primary bottleneck for clinical LLM deployment; the binding constraint is the assumption that any single model can be the sole basis for a clinical decision.”

— the authors