Integrating Local and Global Entropy for Uncertainty Quantification in LLMs

In three linesNew GLU method quantifies uncertainty in LLMs by combining local entropy (token-level) and global entropy (hidden-state geometry). Tested on 3 model families and 6 benchmarks. Detects confident-but-wrong hallucinations missed by existing methods.

Summary generated by Claude — human-verified