Least-to-Most Prompting

anchor

Findings

• IC-489State-of-the-art MLLMs fail at multi-step visual analogical reasoning, with best accuracy at 13% (Llama 3.2) on VOILA-WD and 29% (GPT-4o) on VOILA-ND, far below human performance of 71% and 70% [primary]
• IC-490GPT-4o can identify visual relationships at 97% accuracy when given ground-truth descriptions but drops to 17% when asked to apply known relationships to new visuals, revealing a specific bottleneck in relational transfer [primary]
• IC-491Presenting three images as a single collage rather than sequentially reduces MLLM accuracy by approximately 40% on the relationship application step [primary]
• IC-492Least-to-most prompting consistently improves MLLM accuracy on the relationship application step compared to direct answering, with GPT-4o improving from 0.9% to 6.44% on VOILA-WD [builds-on]
• IC-492Least-to-most prompting consistently improves MLLM accuracy on the relationship application step compared to direct answering, with GPT-4o improving from 0.9% to 6.44% on VOILA-WD [primary]