Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Sep 3·5 min read
Door-in-the-Face Requests and Refusal Behaviour in Large Language Models
TL;DR
Research investigates the effectiveness of the door-in-the-face technique on large language models (LLMs). While Anthropic's Opus 5 shows a significant increase in compliance after a refusal, other models from OpenAI and Google experience a decrease in compliance.
✦ Why It Matters
Consider testing the door-in-the-face technique with your LLMs to optimize request compliance.
Key Takeaways
- Test the door-in-the-face technique with your LLMs for better results.
- Be aware of model-specific responses to refusal techniques.
- Context matters; tailor requests to improve compliance.
- Experiment with request types to find optimal strategies.
How It Works
The door-in-the-face technique leverages the psychological principle that a large initial request, when refused, can make a smaller follow-up request seem more reasonable. In this study, LLMs were tested by first refusing a large request and then being asked a smaller version of the same request.
The varying responses across different model families highlight the importance of understanding how each model interprets and reacts to refusals.