Multimodal large language models are being applied to structured extraction from financial documents, yet their reliability when images are unclear remains largely unexamined. A new preprint on arXiv investigates this gap in the specific context of cheque processing, where models must read fields like the courtesy amount — the numeric value written on the cheque.
The authors identify a structured hallucination risk, abbreviated as CARLAR, that can arise during courtesy amount recognition. The abstract does not fully define the acronym, but it positions CARLAR as a failure mode tied to visual ambiguity, distinct from general free-form hallucination because it occurs within a fixed extraction schema.
As an exploratory countermeasure, the paper proposes using compressed images to reduce this risk. The authors do not claim a full solution; the work is presented as an initial investigation into whether image compression can make models more robust when extracting structured financial data from imperfect visual input.