Measuring Mechanistic Uncertainty via Causal Robustness Margin
October 19, 2026
Mechanistic uncertainty quantifies LLM reliability by measuring the minimum internal intervention required to deviate an output using Jacobian Lens steering vectors. Testing across three open-source models shows that incorrect or uncertain responses exhibit smaller causal robustness margins, providing a complement to standard output-based uncertainty metrics.
HOW THIS AFFECTS YOU
●
builderYou can implement mechanistic uncertainty checks to improve reliability in high-stakes deployment domains.
●
researcherYou can use the causal robustness margin to study internal model stability beyond surface-level token probabilities.