This project quantifies the geometric and structural "break-point" of LLM representations under data corruption. Using Intrinsic Dimension and CKA mapping, it identifies the exact transformer layers where noise destroys semantic alignment and triggers Uncertainty Calibration collapse.
machine-learning transformers pytorch representation-learning manifold-learning information-bottleneck data-corruption adversarial-attack llm-robustness intrinsic-dimensionality multimodal-ai semantic-drift bayesian-uncertainty clip-vlm cka-analysis
-
Updated
Mar 31, 2026 - Python