KORG: Endogenous Self-Reinforcing Flow Capture in Longitudinal LLM Workflows

Large language models fail in many familiar ways: they hallucinate, become sycophantic, inherit earlier mistakes, falsely report execution, corrupt memory, over-rely on previous answers, and sometimes acknowledge corrections without durably changing later behavior. This paper does not propose another name for those failures. It proposes KORG as a process-level construct for a different condition: the point at which one or more familiar LLM failure modes cease to behave as isolated errors and become an endogenous, self-reinforcing process state that captures an increasing share of the workflow. KORG is defined here as an endogenously produced, self-reinforcing flow-capture state in an LLM-centered workflow in which model-generated unsupported content becomes operative process state, influences subsequent transitions, and persists despite corrective feedback sufficiently to redirect process resources from the original task toward maintaining, reconciling, or reproducing the contaminated state. The construct is motivated by longitudinal naturalistic observation of extended human-LLM work and is examined alongside recent findings on multi-turn unreliability, pigeonholing, memory hallucination, sycophancy hysteresis, failed intrinsic self-correction, and tool-use hallucination. The proposed novelty is not the individual failure modes. It is the higher-order process relation: KORG is identified by persistence of flow capture across changing failure modes. The paper introduces pseudo-trace, provenance laundering, repair resistance, cross-content regeneration, repair-induced amplification, and the context-salvage problem, and proposes an operational threshold distinguishing local error, error propagation, and KORG-like states. KORG is presented as a falsifiable working hypothesis rather than as an experimentally established new class of LLM failure.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24183402
Citations
3
Primary Topic
Natural Language Processing Techniques
Type
article
Field-Weighted Citation Impact
4.18
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

KORG: Endogenous Self-Reinforcing Flow Capture in Longitudinal LLM Workflows

Alen Širola
3 citations
Zenodo (CERN European Organization for Nuclear Research)
Natural Language Processing Techniques
4.18
article

KORG: Endogenous Self-Reinforcing Flow Capture in Longitudinal LLM Workflows

Alen Širola
article en
3 citations

Abstract

Large language models fail in many familiar ways: they hallucinate, become sycophantic, inherit earlier mistakes, falsely report execution, corrupt memory, over-rely on previous answers, and sometimes acknowledge corrections without durably changing later behavior. This paper does not propose another name for those failures. It proposes KORG as a process-level construct for a different condition: the point at which one or more familiar LLM failure modes cease to behave as isolated errors and become an endogenous, self-reinforcing process state that captures an increasing share of the workflow. KORG is defined here as an endogenously produced, self-reinforcing flow-capture state in an LLM-centered workflow in which model-generated unsupported content becomes operative process state, influences subsequent transitions, and persists despite corrective feedback sufficiently to redirect process resources from the original task toward maintaining, reconciling, or reproducing the contaminated state. The construct is motivated by longitudinal naturalistic observation of extended human-LLM work and is examined alongside recent findings on multi-turn unreliability, pigeonholing, memory hallucination, sycophancy hysteresis, failed intrinsic self-correction, and tool-use hallucination. The proposed novelty is not the individual failure modes. It is the higher-order process relation: KORG is identified by persistence of flow capture across changing failure modes. The paper introduces pseudo-trace, provenance laundering, repair resistance, cross-content regeneration, repair-induced amplification, and the context-salvage problem, and proposes an operational threshold distinguishing local error, error propagation, and KORG-like states. KORG is presented as a falsifiable working hypothesis rather than as an experimentally established new class of LLM failure.

Zenodo (CERN European Organization for Nuclear Research)
Openalex Percentile: Top 6%
Natural Language Processing Techniques
4.18
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon