PAE Scenario Dataset

500 scenarios across four domains for evaluating peripheral context use in LLM responses.

Dataset Description

Each scenario contains:

• id: unique identifier (e.g., "cus_001", "med_001")
• domain: domain label (customer_service, medical, legal, finance)
• query: literal user query
• direct_answer_material: policy/guideline/reference text
• asker_context: entity history, profile, and patterns
• situational_context: temporal and environmental signals
• anomaly_flags: pre-computed deviations from expected patterns
• expert_response: gold-standard response that uses all peripheral context
• blind_response: technically correct response that ignores peripheral context

Domains

Scenario Archetypes (40 CS / 20 each for others)

1. Standard: Peripheral context clearly matters; lazy agent gives worse answer
2. Hidden signal: Subtle but important signal buried in routine details
3. Red herring: Peripheral context exists but should not change the answer
4. Multi-issue: 2-3 overlapping problems; context drives prioritization
5. Polarity flip: Peripheral context reverses the surface-level answer

Usage

Generated by DeepSeek V4 Flash, validated by MiniMax M2.7. Used in the PAE experiment:

• 6 conditions, 3 model families, 4 domains
• Preregistered at OSF: 10.17605/OSF.IO/W3XYV

Citation

Lee, A. (2026). Peripheral Attention Engineering: Structured Peripheral Context Improves LLM Decision Quality. OSF Preregistration. https://doi.org/10.17605/OSF.IO/W3XYV

License

CC0 1.0 Universal — Public Domain Dedication