Verified, then logged.
Sun 22 Nov 2026 · 00:33 KST / 15:33 UTC한국어 | ENSearch /

Breaking01 / AI

Unfinished medical AI reasoning yields existing diagnoses

Researchers separated diagnoses already present in reasoning from answer completion and diagnostic correctness.

AI-generated editorial illustration for this article

Researchers reported extracting diagnoses from 75 of 94 failed initial responses by a reinforcement-learning model in the MamaReason study. They counted explicitly stated diagnoses, assessing correctness separately.

The paper examined MamaReason's unfinished reasoning. Its evaluation used paired maternal and child-health questions.

The researchers compared the extracted diagnosis names with answers produced by continuing the unfinished responses. Using diagnosis strings alone, they found agreement with the earlier selection in 67/75 cases.

The researchers assessed answer recovery using the existing correctness labels. Reading diagnoses already present in the reasoning recovered 30 of the 40 pairs restored by continuation, without generating additional text.

The paper distinguished routine mentions of a diagnosis from repeated selection of it as an answer. It also assessed diagnostic content, answer completion and diagnostic correctness as separate outcomes.

The abstract does not report effects in actual clinical care.