CARLAR Influence: A Structured Hallucination Risk in Multimodal Cheque Extraction and an Exploratory Compressed-Image Countermeasure
Multimodal large language models (LLMs) are increasingly used for structured extraction from financial documents, but their behavior under visual ambiguity is poorly understood.
ProofPaper ↗
Key points
- We study CARLAR (Courtesy Amount Recognition and Legal Amount Recognition) influence: a model returns the same amount for both cheque fields even though their annotations differ, including cases with a blank annotation encoded as zero.
- Across ten visual prompting conditions, we evaluate GPT-4.1 and Claude 4.5 Sonnet on 243 primary cheques (124 with two observed unequal amounts) and 598 strict mismatches from two additional datasets.
- We also test a compressed-image few-shot intervention that presents examples as one visual table without model fine-tuning.
- These results motivate validation methods that account for visual context and dependencies between extracted fields.
Sources (1)
- [1]CARLAR Influence: A Structured Hallucination Risk in Multimodal Cheque Extraction and an Exploratory Compressed-Image CountermeasurearXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 07:29 PM
Multimodal large language models (LLMs) are increasingly used for structured extraction from financial documents, but their behavior under visual ambiguity is poorly understood.
We study CARLAR (Courtesy Amount Recognition and Legal Amount Recognition) influence: a model returns the same amount for both cheque fields even though their annotations differ, including cases with a blank annotation encoded as zero.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026Reasoning-Aware Multicalibrated Reliability Predictors for Language Models
- Oct 19, 2026Small Average Error, Wide Individual Error: General-Purpose LLMs Estimating Energy Intake from Meal Photos in a Feeding Trial
- Oct 19, 2026An Empirical Study of Parameter-Efficient Fine-Tuning for Mixture-of-Experts LLMs
- Oct 19, 2026TokenGate: Adapting LLMs for Fast and Memory-Efficient Prefill with Learned Token Pruning
- Oct 17, 2026Medical Fine-Tuning Reshapes, Not Reduces, Patient-Facing AI Risk: A Rubric-Based, LLM-Judge Evaluation Framework
- Oct 17, 2026EpochGo: Local-first multi-agent AutoML engine using offline Ollama LLMs and deterministic Scikit-Learn pipelines [P]