Frontier AI Models Share the Same Blind Spot in Cancer Decisions

TL;DR

• Researchers built the Oncology Decision Boundary Benchmark (ODBB), 2,005 decision points pulled from NCCN guidelines and real colorectal cancer cases.
• Nine frontier LLMs released between June 2025 and April 2026 were tested. Pooled together as one giant “super-model,” they still got 42.1% of items wrong, every single one of them.
• The failure rate jumped to 66.4% on real-world colorectal cancer cases, compared with 35.7% on straightforward guideline questions.
• GPT-5.5 and Gemini 3.1 Pro Preview, both tuned to sound decisive, made unsafe clinical commitments 3 to 5 times more often than the other seven models, with no accuracy advantage to show for it.

The Benchmark That Nine Models Couldn’t Beat

Most AI medical benchmarks test recall. Ask a model to name a drug interaction or recite a staging criterion, and it’ll usually nail it, because that’s the same skill tested on a written exam. ODBB does something meaner. It builds 2,005 decision points from actual NCCN treatment guidelines and 419 real colorectal cancer cases, then asks models to make the next clinical move in a sequence, the kind of judgment call an oncologist makes mid-case, not a fact they memorized.

The researchers ran nine frontier models through it, four closed-source and five open-weight, spanning releases from June 2025 through April 2026. No LLM graded the answers. A fully deterministic scorer did, which matters, because it removes the fuzziness that lets a lot of AI benchmarks quietly inflate their own results.

Then they did something clever. Instead of just ranking the nine models against each other, they pooled them into a single hypothetical “super-model,” the best possible answer any one of the nine gave for each item. If even one model out of nine got a question right, the super-model counted it as solved. The logic: if scaling and diversity of training data were the real fix for clinical reasoning gaps, throwing nine different models at the same problem should mop up almost everything.

It didn’t. 42.1% of all 2,005 items were answered correctly by exactly zero of the nine models. On the guideline-based NCCN items, that number was 35.7%. On the real colorectal cancer cases, it climbed to 66.4%, meaning two out of every three real-patient decision points stumped every frontier model tested, together and separately.

The Shared Blind Spot That Money Can’t Buy Out

I’ve read a fair number of AI benchmark papers this year, and most of them read like marketing decks with footnotes. This one doesn’t, and that’s precisely what makes the 42.1% figure land differently than the usual leaderboard churn.

Picture nine locksmiths standing in front of the same stuck padlock. Each one brings a different toolkit, different training, different confidence in their own skill. If the padlock’s failure point is something none of their tools were built to address, it doesn’t matter how many locksmiths you line up. Nine walk away and the lock stays shut. That’s roughly the shape of this result: the failures cluster specifically around choosing between competing guideline pathways before reasoning within any single one, a step that happens before the actual medical knowledge even gets applied. Add more locksmiths, and you still don’t fix the mechanism.

The competitive detail here is the part that should worry model builders more than the headline number. GPT-5.5, OpenAI’s entry, and Google’s Gemini 3.1 Pro Preview were both tuned to be decisive rather than hedge-y. In this benchmark, that decisiveness translated into unsafe clinical commitments 3 to 5 times more often than the seven more cautious models in the test, without any corresponding bump in accuracy. In plain terms: sounding confident and being right are not the same skill, and right now the industry’s push toward decisive-sounding assistants is optimizing for the wrong one in a domain where the wrong one kills people.

There’s a subtler failure buried in the data too. In 3% to 9% of items, a model’s own reasoning identified the correct next clinical step, then the model failed to actually commit to that step in its final answer. That’s not a knowledge problem. It’s a decision problem, happening after the model already knew the right move. No amount of additional training data fixes a model that knows the answer and flinches anyway. The paper’s authors argue this points toward architectural intervention, essentially building in a mechanism that recognizes when a model has hit its competence boundary and hands the case to a human, rather than hoping the next generation of pretraining quietly closes the gap.

Medical Exams Test Memory, Not Judgment

Why does this matter more than another benchmark score? Because the entire pitch for AI in clinical settings has leaned hard on medical exam performance, USMLE-style scores where frontier models now routinely beat human test-takers. But those exams are largely recall and single-step reasoning. They don’t simulate the messy, sequential, branching judgment calls that define actual oncology care, where a clinician has to weigh competing guideline pathways, patient-specific complications, and escalation risk all at once.

ODBB was built specifically to close that gap between what gets tested and what actually happens in a treatment room. Real colorectal cancer cases carry ambiguity that guideline text doesn’t. A patient’s comorbidities, prior treatment history, and atypical presentation all pull decisions away from the clean, single-path logic a model might handle fine on a written exam. That’s exactly why the failure rate on real cases (66.4%) dwarfs the failure rate on guideline items (35.7%). The gap between textbook and bedside is the story here, and it’s a much bigger gap than most people assumed going into this.

Where This Boundary Gets Tested Next

A few things are worth tracking from here. First, whether other research groups replicate this pooled-failure pattern outside colorectal cancer, since a boundary this consistent in one cancer type raises the question of whether it shows up in lung, breast, or hematologic oncology too. Second, whether any lab actually builds the architectural routing layer the authors call for, a system that flags when a model has crossed its competence boundary and defers to a clinician instead of guessing. Third, watch whether decisiveness tuning, the same trait that made GPT-5.5 and Gemini 3.1 Pro Preview riskier here, gets dialed back for medical deployments or just gets marketed harder anyway. This is a preprint, not a peer-reviewed publication yet, so how it holds up under scrutiny from oncologists and AI safety researchers over the coming months will tell us a lot about whether 42.1% becomes an industry wake-up call or a footnote nobody outside the field ever reads.

Source: arXiv