Properties2
| Type | Concept |
| Note created | Jun 16, 2025 |
Explicit Extraction and Reasoning (EER) is a structured prompting strategy that breaks down complex reasoning tasks into four explicit, sequential steps. Unlike Zero-Shot COT, EER guides the model through a predefined reasoning pipeline.
Four-Step Process
- Extraction of Regions - Identify relevant regions/areas for the question
- Extraction of Relevant Places - Extract specific locations within those regions
- Extraction of Values from Legend - Extract corresponding values from visual data
- Reasoning based on Extracted Values - Use extracted values to reach final answer
Performance Insights
From the MAPWise benchmark:
- Most models performed better with Zero-Shot COT
- Exception: Gemini 1.5 Flash performed better with EER (strong instruction-following)
- Smaller open-source models struggled with complex EER instructions
Use Cases
EER is particularly suited for:
- Visual question answering (maps, charts)
- Tasks requiring systematic data extraction
- Models with strong instruction-following capabilities