LLM Uncertainty Patterns Compared to Human Cognition
A new computer science paper investigates how similar large language model uncertainty is to human uncertainty, an underexplored question in uncertainty quantification. The study examines uncertainty alignment—the presence of human-like uncertainty signals—in LLM overt behavior and internal activation patterns. It tests whether models show simultaneous alignment and calibration across multiple-choice and open-ended factual recall datasets, and characterizes the effect of instruction fine-tuning on these facets. The work aims to recognize and combat hallucination by improving calibration, the accuracy of uncertainty judgments relative to task efficacy.
Key facts
- Uncertainty quantification is a growing subfield of LLM behavioral analysis.
- The study focuses on how similar LLM uncertainty is to human uncertainty.
- It investigates uncertainty alignment in LLM behavior and internal activation patterns.
- Models are tested for simultaneous alignment and calibration on multiple datasets.
- Datasets cover both multiple choice and open ended factual recall.
- The effect of instruction fine-tuning on alignment and calibration is characterized.
- The field primarily aims to recognize and combat hallucination.
- Calibration measures the accuracy of uncertainty judgments to task efficacy.
Entities
Institutions
- arXiv