AI models that read text and look at images sometimes solve the exact same problem differently depending on which input they get.
Researchers built ODA-Data, a dataset of geometry problems presented three ways: as text, as an image, and as text plus image together, specifically to study and train on this inconsistency. They then developed Modality-Informed Reciprocal Reasoning Optimization, or MIRROR, a reinforcement learning method that has a model check its own performance across all three versions of a problem. Whichever version the model answers best becomes a teacher, and the model is trained to pull its answers on the other versions closer to that teacher's reasoning, using a reward for correct answers plus a penalty for straying too far from the teacher. On multi-modal math benchmarks, models trained this way were more accurate and more consistent across text, image, and combined inputs than models trained with standard reinforcement learning.
The underlying finding is the more interesting part: these models do not have one consistent reasoning process, they have several, and which one shows up depends on how a question is framed. That is a real problem for any product that assumes a user gets the same answer whether they type a question or photograph it off a whiteboard.
It is self-improvement by committee: the model grades its own homework across formats, then copies from whichever version of itself did best.