Ask a multimodal AI model to review a long recording or document in one shot, and it will quietly skip about a third of it.
New research tests what happens when a multimodal model reviews long sources in a single pass versus two. The one-pass approach drops roughly a third of the content and pads the rest with embellishment, and it is not a perception problem: the same dropped material reappears when the model is simply asked to transcribe the source. The bottleneck is capacity: one pass can't perceive, reason, and write a faithful long review at the same time, and even a much larger reasoning budget doesn't help, because the model spends it planning a review instead of transcribing the source. Splitting the job into two passes with the same model weights, first transcribing then reviewing the transcript, improved both faithfulness and coverage across a 21-source test suite.
This is a structural fix for a problem you can't prompt your way out of, since more reasoning time just makes the model plan a review of content it hasn't written down yet. It matters for anyone using AI to summarize meetings, contracts, or scanned documents, where losing a third of the material is the difference between a useful summary and a liability. It also suggests the industry's habit of throwing bigger reasoning budgets at problems is solving the wrong one here.
Still, the researchers call this a mitigation, not a cure: the review pass can run out of room on very long sources, and once the transcript is out of context, the model can start confabulating from memory anyway.