Choosing what to learn requires both validity and learning-value gates

Type: kb/types/note.md · Tags: learning-theory, deploy-time-learning

Learning loops need two different checks before they promote a candidate into durable memory, weights, or knowledge artifacts. The candidate must be valid enough to learn from, and it must be worth learning for the current system.

A validity gate asks whether the candidate is grounded, coherent, and trustworthy: whether an answer is source-supported, a proposed note is faithful to its evidence, or a link relationship is real rather than a keyword accident. Without this gate, the loop accumulates contamination.

A learning-value gate asks whether promoting the candidate would improve the system: whether the item covers a real gap, exposes unreliable model behavior, or helps future tasks enough to justify the added maintenance burden. Without this gate, the loop accumulates clutter.

The self-training case

Self-Training-LLM makes the split concrete. The corresponding paper, Self-training Large Language Models through Knowledge Detection, uses the sequence instruction generation → SFT (supervised fine-tuning) → preference labeling → knowledge filtering. It filters the preference data before DPO (preference optimization) in two stages: consistency filtering removes low-confidence responses, and knowledge filtering removes samples on which the SFT model is considered knowledgeable. In this factual-QA pipeline, the first is a proxy for validity and the second is a proxy for model-relative learning value. The source does not establish the broader gates; the contamination and clutter arguments above do that work.

The KB case

For KB learning, the same split should govern note, link, and synthesis promotion. A proposed note can be faithful but redundant. A proposed link can be plausible but not navigationally useful. A proposed synthesis can be valid but too narrow to change future answers.

This refines automating KB learning is an open problem. The open problem is not one generic oracle for "should we learn this?" It is at least two oracles: one for groundedness and one for value relative to the current system's gaps. A single score can combine them for ranking, but the underlying checks should remain separate so failures are diagnosable.


Relevant Notes: