Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 31·5 min read
When Linguistic and Internal Confidence Diverge in Large Language Models
TL;DR
Large language models (LLMs) often misalign their reported linguistic confidence with their internal confidence levels. This study reveals significant divergence across various tasks and models, suggesting that linguistic confidence should be evaluated with multi-axis diagnostics before use in reliability assessments.
✦ Why It Matters
Evaluate linguistic confidence with multi-axis diagnostics to ensure reliability in your AI applications today.
Key Takeaways
- Assess linguistic confidence against internal metrics for accuracy.
- Use multi-axis diagnostics to evaluate confidence reliability.
- Be cautious of inflated confidence from attitude cues.
- Consider prompt design's impact on confidence reporting.
How It Works
The study compares linguistic confidence, which is how models express certainty in their responses, with internal confidence derived from logits (raw model outputs) and semantic entropy (uncertainty in generated text). By analyzing various tasks and models, it identifies patterns of divergence, particularly in how different models report confidence based on their training and prompt design.