Agentic data science with the Data Agent Kit
cloud.google.com·8h ago
AI Research
arXiv CS.AI·10d ago·5 min read
Verbalized and Internal Probabilities Are Coupled in Large Language Models
TL;DR
Large language models (LLMs) exhibit a connection between their internal probabilities and verbalized uncertainties, revealing that both are influenced by training data. This alignment suggests that verbalized probabilities can effectively probe a model's internal distribution.
✦ Why It Matters
Utilize verbalized probabilities to gain insights into a model's internal decision-making process today.
Key Takeaways
- Explore how LLMs express uncertainty through verbalized probabilities.
- Use verbalized probabilities to understand model behavior better.
- Investigate the impact of training data on probability readouts.
- Align internal and verbalized probabilities for improved interpretability.
How It Works
LLMs generate text by sampling from a probability distribution that reflects their uncertainty. This study shows that both internal probabilities and verbalized uncertainties are influenced by the same sources of uncertainty in the training data, indicating a deeper connection between the two.