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:
- automating KB learning is an open problem — extends: gives the boiling-cauldron promotion loop a diagnosable acceptance interface, separating contamination control from clutter control
- oracle strength spectrum — grounds: both gates require oracles, but their oracle targets differ
- memory management policy is learnable but oracle-dependent — parallels: learned curation policies work only when the promotion signal is well defined
- Self-Training-LLM — evidenced-by: code-grounded review of the separate question-filtering and unknown-filtering pattern
- Self-training Large Language Models through Knowledge Detection — evidenced-by: paper-level account of consistency filtering and knowledge filtering as separate stages
- Into the Unknown: Self-Learning Large Language Models — evidenced-by: treats “what to learn” as the main problem for its loop, but operationalizes value only as selecting hallucination-defined unknown factual knowledge