A recent arXiv preprint takes aim at a popular trend in document embedding: training models on rendered page images instead of the underlying text. The authors argue that while these document-as-image representations have become common, they do not hold up in scientific retrieval settings.
The paper points to a mismatch between the approach and the task. Scientific documents contain dense structure, equations, and precise terminology that may be lost or distorted when pages are treated purely as images. The abstract suggests that existing retrieval benchmarks fail to capture this shortfall, making the problem easy to overlook.
Because the source is only an abstract, the specific experiments and results are not yet available in detail. Still, the claim is a useful caution for researchers who rely on image-based embeddings for scientific search and discovery.