Integrating Local and Global Entropy for Uncertainty Quantification in LLMs
Researchers have developed a new method called Global-Local Uncertainty (GLU) to improve how large language models quantify their uncertainty. This approach combines token-level entropy with a novel measure of global uncertainty derived from the geometric complexity of the models' hidden states. By fusing these two signals, GLU can identify confident but incorrect outputs that traditional methods often miss, demonstrating improved reliability across various benchmarks and model architectures with a single forward pass. AI
IMPACT Enhances LLM reliability by identifying confident but incorrect outputs, crucial for safe deployment.