Layer-wise Semantic Dynamics (LSD) is a model-agnostic framework for hallucination detection in Large Language Models (LLMs). It analyzes the geometric evolution of hidden-state semantics across transformer layers, using contrastive alignment between model activations and ground-truth embeddings to detect factual drift and semantic inconsistency.
nlp machine-learning deep-learning pytorch representation-learning language-model-evaluation contrastive-learning trustworthy-ai large-language-models interpretability-and-explainability geometric-learning hallucination-detection hallucination-mitigation factuality-evaluation semantic-geometry transformer-analysis semantic-drift-analysis
-
Updated
Oct 28, 2025 - Python