Ingest: Into the Unknown: Self-Learning Large Language Models

Type: kb/sources/types/ingest-report.md

Source: into-the-unknown-self-learning-large-language-models.md Captured: 2026-04-16 From: https://arxiv.org/html/2402.09147v4

Classification

Type: scientific-paper -- arXiv preprint with a formal concept, methods for identifying unknown knowledge, proposed metrics, experiments across model families, and a training-loop evaluation. Domains: learning-theory, continual-learning, evaluation, hallucination Author: Teddy Ferdinan, Jan Kocon, and Przemyslaw Kazienko, Department of Artificial Intelligence, Wroclaw Tech. Academic NLP/AI affiliation and reproducibility artifacts are useful credibility signals, but treat the claims as preprint-level.

Summary

The paper proposes a self-learning LLM loop organized around "Points in the Unknown" (PiUs): atomic factual questions that a model does not know and can identify through hallucination scoring. The system first performs self-questioning through one of four PiU-discovery methods: external prompts from trending topics, open model-generated questions, induced 5W+1H questions, or oracle-selected embedding-space topic samples. It then filters hallucinated questions, searches for answers, constructs training data, and updates the model so the former unknowns become known. The authors also propose self-learning capability metrics: Curiosity Score for question diversity, Knowledge-Limit Awareness Score for generating questions the model actually cannot answer, Brevity Coefficient for question economy, and a combined SLC score. Their experiments suggest instruction-tuned models with at least roughly 3B parameters are better candidates for this loop, and a Mistral-Instruct self-learning run reduced hallucination scores on the selected questions while preserving performance on sampled Wiki and Alpaca evaluations.

Connections Found

The connect report places this source in the KB's continual-learning, substrate, and oracle-construction cluster. It contrasts with Treat continual learning as substrate coevolution because it is a clean opaque-substrate loop: identify unknowns, search for facts, train weights. It qualifies Continual learning's open problem is behaviour, not knowledge because the paper's successful training would be durable capacity change, but the KB's criterion is durability rather than weights. It extends Automating KB learning is an open problem by offering a concrete "what to learn" operator, while also showing why factual hallucination scoring does not directly solve judgment-heavy KB mutations. The closest oracle comparison is Memory management policy is learnable but oracle-dependent: AgeMem uses task completion as the oracle; this paper uses hallucination/self-checking as an unknown detector and external search as supervision. The source also connects to Oracle strength spectrum, The boundary of automation is the boundary of verification, and Letta's Continual Learning in Token Space.

Extractable Value

  1. "What to learn" is the central missing operator in self-learning loops -- The paper's strongest contribution is not the fine-tuning recipe but the explicit problem framing: a system cannot self-learn until it can select unknowns worth acquiring. This transfers directly to KB learning, where "what note/link/synthesis is missing?" is harder than writing a candidate once selected. [quick-win]
  2. Hallucination can serve as an unknown-selection oracle only in narrow factual regimes -- PiU treats hallucination on simple questions as evidence of missing knowledge. That is operationally useful for factual QA, but it is a soft oracle rather than ground truth. The limitation is valuable because it clarifies where self-assessment can and cannot replace external verification. [experiment]
  3. Self-learning capability decomposes into exploration and self-knowledge -- Curiosity Score and Knowledge-Limit Awareness Score separate two properties often blurred together: generating diverse candidate questions and identifying questions the model genuinely cannot answer. This decomposition could inform agent-memory and KB-maintenance agent evaluation. [experiment]
  4. The loop shape mirrors artifact-learning loops despite targeting weights -- Self-questioning, knowledge searching, and model training map loosely to candidate generation, evidence gathering, and promotion. The analogy is useful, but the promotion target differs: the paper promotes into weights; Commonplace-style learning promotes into inspectable artifacts. [just-a-reference]
  5. The paper is an opaque-loop counterpoint to substrate coevolution -- It makes the mainstream bet more concrete: freeze non-opaque artifacts, automate unknown discovery, and improve model weights. That gives the KB a sharper contrast case for why non-opaque substrates still need their own loops. [quick-win]
  6. Embedding-space "unknown regions" are useful vocabulary but weak mechanism -- Human Knowledge Space and PiU boundaries make the learning problem discussable, but the actual mechanism is question generation plus hallucination scoring. Treat the geometric framing as a metaphor unless future work makes the space and boundary estimation testable. [deep-dive]

Limitations (our opinion)

Hallucination is not uniquely caused by missing knowledge. It can also arise from prompt ambiguity, conflicting evidence, reasoning failure, decoding variance, retrieval mistakes, or instruction-following quirks. That makes PiU a noisy unknown detector and weakens any broad claim that hallucination scoring finds true knowledge gaps.

The "atomic knowledge in Human Knowledge Space" framing is under-specified. It helps visualize known and unknown regions, but the paper does not show that factual knowledge decomposes into stable atomic points or that the embedding-space boundary has the precision implied by the diagrams.

The learning loop still depends on external sources and human-quality control. In the self-learning experiment, the authors used gpt-4o-2024-05-13 to answer the selected questions and manually verified answer quality. That is not a fully autonomous closed loop, and it moves the oracle burden into source quality and answer verification.

The evaluation is factual and bounded. It does not show transfer to judgment-heavy learning, synthesis, policy learning, or curation. This is exactly where KB learning is hardest: unknownness is not just "the model cannot answer a fact" but "the knowledge base lacks a useful abstraction, connection, or durable procedure."

The catastrophic-forgetting result is local to the chosen method, model, and evaluation slices. Perplexity and ROUGE/Judge scores on sampled Wiki/Alpaca are useful smoke tests, but they do not establish general safety for continual fine-tuning.

Write a note titled "Hallucination is an unknown-selection oracle only for factual learning" connecting to Automating KB learning is an open problem, Oracle strength spectrum, Memory management policy is learnable but oracle-dependent, and Treat continual learning as substrate coevolution. The note should argue that PiU is valuable as a concrete unknown-selection operator, but that KB learning needs different oracles for missing abstractions, stale links, and synthesis quality.