Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 26·5 min read
How Well Do Models Follow Their Constitutions?
TL;DR
AI models often struggle to adhere to their behavioral specifications under real-world conditions. A multi-method audit pipeline was developed to evaluate model compliance with these specifications.
✦ Why It Matters
Engineers can leverage this audit pipeline to enhance model compliance with behavioral specifications in real-world applications.
Key Takeaways
- Implement a multi-method audit pipeline for model evaluation.
- Focus on improving adherence to behavioral specifications.
- Analyze failure patterns to identify areas for improvement.
- Track compliance metrics across model generations.
Full Summary
AI developers create models based on detailed behavioral specifications, like Anthropic's constitution and OpenAI's Model Spec, to guide their behavior. However, it remains uncertain how well these models follow their guidelines in challenging, multi-turn interactions.
A new audit pipeline was introduced, which breaks down specifications into testable components and generates adversarial scenarios to assess model performance. This pipeline was applied to seven models from each specification, revealing that adherence improved significantly with each model generation.
For instance, violation rates for Anthropic's Claude models decreased from 15.0% to 2.0%, while OpenAI's GPT models dropped from 11.7% to 3.6%. Despite these improvements, some failures persisted, particularly in areas involving identity questioning and false claims.
These findings suggest that while models are becoming better at following their specifications, there are still critical areas needing attention.