Taxonomy of Latent Space Pathologies: Identifying, Classifying, and Mitigating Failure Modes in Large Language Models

TL;DR: Let’s make a “field guide” to all the weird and dangerous ways latent spaces can go wrong in big language models, and invent new tools to find and fix them.

Research Question: What are the major classes of pathological behaviors or “failure modes” that emerge within the latent spaces of large language models, and how can we systematically detect and address them?

Hypothesis: Latent spaces in LLMs exhibit identifiable classes of pathologies—such as representational collapse, anomalous attractors, or “jailbreak” subspaces—that can be detected via unsupervised probing and mitigated by targeted interventions.

Experiment Plan: Collect a large suite of trained LLMs and analyze their latent activations during standard and adversarial inputs. Develop an unsupervised taxonomy of observed pathologies using clustering, anomaly detection, and information-theoretic measures. Correlate these pathologies with model failures in safety, reasoning, or factuality. Test mitigation strategies (e.g., adversarial training, latent space regularization, dynamic recalibration). Release an open-source toolkit for latent space pathology detection and repair.

References:

• Yu, X. et al. (2026). The Latent Space: Foundation, Evolution, Mechanism, Ability, and Outlook.