LOTUS: Multilinear Decomposition for LLM Hallucination Detection

LOTUS uses low-rank multilinear decomposition to extract compact, discriminative representations from LLM hidden states for hallucination detection. It captures peak activations and cross-layer changes to identify factual errors while reducing dependence on large labeled datasets.