Researchers introduce Loud and Clear (LoC), a method for helping AI systems use evidence in pages retrieved for document question-answering. It combines higher-resolution page images with text from query-relevant regions, without extra training or ground-truth annotations.
In the paper’s reported tests, LoC outscored eight prior methods on three benchmarks, including ColQwen3 by 8.88 points on SciEGQA. The authors also found that increasing the visual-token budget improved accuracy even when the correct pages were already retrieved.
Why it matters
For teams building document-answering systems, the results suggest that improving how retrieved pages are shown to a model may help even when retrieval already finds the right pages; the paper does not establish how well LoC works in real-world deployments.
