Recursive Abstraction and Structural Underdetermination: An Axiomatic Framework for Knowledge Generation and LLM Hallucination
Why do large language models (LLMs) hallucinate, and why does the same generative capacity also support knowledge production and creativity? This paper develops a conceptual and partially formal framework in which cognition is modeled as the recursive generation of representations from mediated information. The framework comprises one ontological postulate and six axioms concerning mediated access to reality, lossy abstraction, recursive representation, object and operation abstraction, candidate generation, and epistemically independent testing. It formalizes first-level representation as the abstraction of mediated information and higher-level cognition as the recursive abstraction of existing representations, possibly with cross-level inputs and new evidence. Under an explicit reconstruction assumption, two results follow. First, along a closed abstraction chain, processing cannot increase mutual information about the original input. Second, reconstructing concrete content from a non-injective abstraction is structurally underdetermined. These results do not by themselves establish a complete mechanism of LLM hallucination; rather, they identify a general informational structure that can connect representation loss, candidate generation, and factual error. The paper argues that creativity and hallucination may share this generative structure while differing in task, epistemic stance, novelty, and relation to evidence. It therefore treats hallucination as a failure of warranted factual assertion, not as a property of generation alone. The framework yields testable predictions about evidence removal, independent verification, calibration, prompt variation, and generation-verification systems, while explicitly stating its empirical and formal limitations.
Authors
- Bolin Zhang (ORCID: https://orcid.org/0009-0002-9626-6323)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24247979
- Citations
- 1
- Primary Topic
- Topic Modeling
- Type
- preprint