Let Androids Dream of Electric Sheep: A Human-Inspired Image Implication Understanding and Reasoning Framework

TL;DR

Let Androids Dream introduces LAD, a three-stage framework for image implication understanding that mimics human cognition by Perception, Search, and Reasoning. By converting visuals to rich textual representations, enriching them with cross-domain knowledge via adaptive search, and outputting context-aligned implications through explicit reasoning, LAD achieves state-of-the-art performance on English image implication benchmarks and substantial gains in Chinese benchmarks, while also generalizing to broader VQA tasks. The work provides a detailed OSQ evaluation methodology with high human-model consistency and demonstrates robust improvements across multiple base models and benchmarks, highlighting the practical potential for nuanced multimodal reasoning and human-AI interaction. The framework's components—the perception of enriched textual representations, the adaptive cross-domain search (ModelSearch/WebSearch with Self-Judge routing), and the LAD-CoT reasoning—collectively address contextual gaps that limit traditional MLLMs in metaphor and implication tasks, offering a scalable path toward contextual-alignment vision-language reasoning.

Abstract

Metaphorical comprehension in images remains a critical challenge for AI systems, as existing models struggle to grasp the nuanced cultural, emotional, and contextual implications embedded in visual content. While multimodal large language models (MLLMs) excel in general Visual Question Answer (VQA) tasks, they struggle with a fundamental limitation on image implication tasks: contextual gaps that obscure the relationships between different visual elements and their abstract meanings. Inspired by the human cognitive process, we propose Let Androids Dream (LAD), a novel framework for image implication understanding and reasoning. LAD addresses contextual missing through the three-stage framework: (1) Perception: converting visual information into rich and multi-level textual representations, (2) Search: iteratively searching and integrating cross-domain knowledge to resolve ambiguity, and (3) Reasoning: generating context-alignment image implication via explicit reasoning. Our framework with the lightweight GPT-4o-mini model achieves SOTA performance compared to 15+ MLLMs on English image implication benchmark and a huge improvement on Chinese benchmark, performing comparable with the Gemini-3.0-pro model on Multiple-Choice Question (MCQ) and outperforms the GPT-4o model 36.7% on Open-Style Question (OSQ). Generalization experiments also show that our framework can effectively benefit general VQA and visual reasoning tasks. Additionally, our work provides new insights into how AI can more effectively interpret image implications, advancing the field of vision-language reasoning and human-AI interaction. Our project is publicly available at https://github.com/MING-ZCH/Let-Androids-Dream-of-Electric-Sheep.

Figures (9)

• Figure 1: An image is worth a thousand words: For the image implication understanding task, different elements' combination lead to different thinking paths, but the correct path needs all elements with multiple reasoning thoughts.
• Figure 2: The general framework of Let Androids Dream (LAD), which includes three stages: (1) Perception: converting raw visual information into rich and multi-level textual representations, (2) Search: iteratively searching and integrating cross-domain knowledge to resolve ambiguity, and (3) Reasoning: generating context-alignment image implication interpretations via explicit reasoning.
• Figure 3: A case study of different methods on Multiple-Choice Question. The End2End method shows superficial reasoning and the CoT method shows over-inference, while our LAD framework shows the correct path via more contextual alignment analysis. The full prompt is listed in Appendix \ref{['app:appendix_prompts']}.
• Figure 4: Evaluation metric and evaluation standard of Open-Style Question.
• Figure 5: The evaluation prompt of Open-Style Question (OSQ).