Loud and Clear helps generators use retrieved page evidence

When a document system retrieves the page containing an answer, what makes the generator miss evidence that is already there? The authors ask whether the way pages are presented, including their image detail and the visibility of relevant text, can change how fully the generator answers.

The training-free Loud and Clear method scales retrieved pages and adds relevant OCR text, improving fully correct answers in three tested document question sets.

Can a document-answering generator use evidence already present on the pages it retrieves? The authors tested this first with pages known to contain the answer, asking whether more visual input and clearer evidence would improve fully correct answers.

They then combined scaled pages with text from relevant regions. In end-to-end tests using the same retrieved pages and generator as ColQwen3, LoC scored higher on three document question sets under a judge that credited only fully correct answers. This supports page presentation as a contributor to answer accuracy in those settings.

The evidence comes from English document question-answering tests, Qwen-family generators, and one reference-based answer judge. Larger image inputs also trade accuracy against inference latency, and the method depends on document parsing and OCR quality.