Geometric Latent Biopsy: Zero-Shot Anomaly Detection in LLM Residual Streams

Fuente: Zenodo
Salvato in:
Dettagli Bibliografici
Autore principale: Llorente-Saguer, Isaac
Natura: Recurso digital
Pubblicazione: Zenodo 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866901492005339136
author Llorente-Saguer, Isaac
author_facet Llorente-Saguer, Isaac
contents A zero-shot geometric anomaly detector for LLM residual streams. The method identifies harmful or anomalous prompts by measuring angular deviations from a normative (safe) manifold in hidden representations, using PCA-derived reference directions and Gaussian Mixture Model scoring. No harmful examples are required at fit time under the normative reference strategy.
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19294977
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Geometric Latent Biopsy: Zero-Shot Anomaly Detection in LLM Residual Streams
Llorente-Saguer, Isaac
large-language-models
anomaly-detection
interpretability
residual-stream
safety
geometric-methods
PCA
gaussian-mixture-model
A zero-shot geometric anomaly detector for LLM residual streams. The method identifies harmful or anomalous prompts by measuring angular deviations from a normative (safe) manifold in hidden representations, using PCA-derived reference directions and Gaussian Mixture Model scoring. No harmful examples are required at fit time under the normative reference strategy.
title Geometric Latent Biopsy: Zero-Shot Anomaly Detection in LLM Residual Streams
topic large-language-models
anomaly-detection
interpretability
residual-stream
safety
geometric-methods
PCA
gaussian-mixture-model
url https://doi.org/10.5281/zenodo.19294977