Token-Efficient Multimodal Reasoning via Image Prompt Packaging

Researchers introduce Image Prompt Packaging (IPPg), a technique that embeds text directly into images to reduce multimodal AI inference costs by 35.8-91.0% while maintaining competitive accuracy. The method shows significant promise for cost optimization in large multimodal language models, though effectiveness varies by model and task type.

• →Image Prompt Packaging achieves 35.8-91.0% inference cost reductions across GPT-4.1, GPT-4o, and Claude 3.5 Sonnet models.
• →Despite token compression of up to 96%, accuracy remains competitive in many settings with highly model- and task-dependent outcomes.
• →The technique works best on schema-structured tasks but struggles with spatial reasoning, non-English inputs, and character-sensitive operations.
• →Visual encoding choices can cause accuracy shifts of 10-30 percentage points, making them critical variables in multimodal system design.
• →GPT-4.1 showed simultaneous accuracy and cost gains on CoSQL while Claude 3.5 incurred cost increases on several VQA benchmarks.