[tsenturk]

When we set out to implement this within an n8n automation, we encountered some implementation challenges. The issues stemmed from the self-training process—specifically regarding the interest and stress score tree—reaching information saturation, which led to a decay in curiosity. However, keeping the threshold constant (not updating it) fundamentally resolved the issue. To be clear, the experiment has only just begun; this is merely an introductory post outlining the basic architecture.

[esafak]

In that case I would start by studying the literature. The first two uncertainty estimation & out-of-distribution (OOD) detection approaches you mention, "Embedding Distance" and "Self-Interrogation", are sometimes called feature-space density estimation and consistency-based uncertainty quantification. Practical algorithms include Semantic Entropy, Self-Consistency / Verbalized Confidence, and Embedding-based Density (Mahalanobis Distance).

References:

A Survey of Uncertainty Estimation Methods on Large Language Models (https://aclanthology.org/2025.findings-acl.1101/)

A Survey of Uncertainty Estimation in LLMs: Theory Meets Practice (https://arxiv.org/abs/2410.15326v1)