Self-Function Vectors Enable Bayesian Aleatoric Uncertainty Estimation in ICL
June 19, 2026
Self-function vectors built from internal model representations model the latent concept learned during in-context prompting, enabling direct aleatoric uncertainty estimation without brittle input or decoding manipulations. The method separates data-inherent uncertainty from model limitations in ICL settings where existing uncertainty decomposition methods fail.
HOW THIS AFFECTS YOU
●
researcherThis offers a mechanistic interpretability-grounded approach to uncertainty decomposition in ICL that avoids the sensitivity issues of prompt-perturbation baselines.