Referring uncertain answers by decision-model confidence: Jev, OpenAI Decisions and Clef on Korean and US medical licensing examination questions

In the rapidly evolving landscape of artificial intelligence, a new class of tools has emerged to tackle a specific problem: how to make machines choose between fixed options without writing out long explanations. Traditionally, large language models answer questions by generating text word by word, a process that is flexible but can be slow and expensive. A newer approach, known as decision modeling, skips the text generation entirely. Instead, the model looks at a list of possible answers provided by the user and assigns a probability score to each one, essentially saying, "I am 90 percent sure this is the right choice, and 10 percent sure it is that one." This method is faster and cheaper, but it raises a critical question for high-stakes fields like medicine: can these machines know when they are wrong? If a model is confident but incorrect, it could lead to dangerous mistakes. The challenge is to find a way for the system to recognize its own uncertainty and hand over difficult cases to a human expert or a more powerful computer before an error occurs.

Researchers at Inje University College of Medicine set out to test whether these decision models could reliably spot their own mistakes on medical licensing exams. They focused on three different models from different companies, testing them on hundreds of questions from Korean and United States medical board examinations. The goal was not just to see how many questions the machines got right, but to see if the probability scores they generated could act as a warning system. The researchers asked: if a model says it is only slightly confident in an answer, does that actually mean the answer is likely to be wrong? And if the system is told to set aside its least confident guesses and send them to a stronger, more expensive model for a second look, does that improve the overall accuracy without costing a fortune?

To find the answers, the team ran a series of computer simulations using real exam questions. They fed thousands of medical questions to three different decision models: Jev, OpenAI Decisions, and Clef. Each model was asked to pick the correct answer from a list of options and provide a confidence score for its choice. The researchers then compared these scores against the official answer keys to see if low confidence truly correlated with errors. They discovered that the models were indeed quite good at this self-assessment. When the models were unsure, they were often wrong. Specifically, the models correctly identified the right answer for about 83 to 92 percent of the Korean questions and 85 to 88 percent of the English questions. More importantly, the probability scores successfully flagged the errors. In the Korean tests, the models' uncertainty scores were able to distinguish wrong answers from right ones with a high degree of reliability, and the same held true for the English tests.

The study then moved to a practical application: a referral system. The researchers simulated a scenario where the least confident 20 percent of answers from each model were automatically sent to a much more powerful and expensive model, GPT-6.1 Sol, for a final decision. The results showed that this simple strategy worked remarkably well. By referring just one-fifth of the answers, the system caught between 56 and 91 percent of the errors made by the initial models. This process boosted the overall accuracy of the system by 6 to 9 percentage points. The cost of this improvement was a fraction of what it would have cost to run the powerful model on every single question. The referral approach achieved nearly half to almost all of the possible accuracy gains that would be available if a perfect "oracle" knew exactly which answers were wrong in advance, all while spending only about a quarter to a third of the cost of using the powerful model exclusively.

However, the researchers also uncovered important limitations that suggest these tools are not yet ready for prime time in a hospital. While the models were good at spotting their own uncertainty, they were not perfect. Some errors were made with high confidence, meaning the machine was sure of the wrong answer. These confident errors were often stable, meaning they did not change even when the order of the answer choices was shuffled, and they sometimes matched the wrong answers given by other advanced models. This suggests that when these models are confident but wrong, they are often wrong for the same deep reasons, such as missing information in the question or relying on outdated regulations. In the Korean exams, for instance, many errors involved specific local laws and administrative rules that the models struggled with, even when they expressed low confidence. In the English exams, some of the questions themselves were disputed by human doctors, making it difficult for any machine to get the "right" answer based on historical keys.

The study also examined how sensitive these systems are to the way questions are presented. When the researchers rotated the order of the answer choices, the models' decisions changed in about 7 to 16 percent of cases. This means that a model might refer a question for a second look simply because the options were listed in a different order, even if the medical content remained the same. While the overall accuracy of the system remained stable despite these small shifts, the individual decisions were not entirely consistent. The researchers noted that this instability, combined with the existence of confident errors, means that these models cannot yet be trusted to work alone. They function best as a first reader that handles the easy cases and flags the difficult ones for a human doctor to review.

Ultimately, the research demonstrates that decision models offer a promising middle ground between speed and accuracy. They can process medical questions quickly and cheaply, and their internal confidence scores provide a useful mechanism for filtering out the most likely errors. By referring the uncertain 20 percent of answers to a stronger system, it is possible to recover most of the accuracy of a top-tier model at a much lower cost. Yet, the path to clinical use requires more than just these simulations. The presence of confident errors and the sensitivity to question formatting indicate that these tools need careful validation in real-world settings. Before they can be used to guide patient care, they must be tested in actual clinical workflows where human doctors can review the referred cases and audit the confident answers that might still be wrong. The technology shows great potential, but it remains a tool that needs a human hand to guide it.