Work

Anatomical Review of Progressive Adversarial Sensitisation in LLM Alignment: Kindling Analogues, Competing Theories, and Ontological-Epistemological Verticals

Year: 2026
Type: article
Abstract: This anatomical review dissects the 2026 Scientific Reports article by Ngo Cheung, which demonstrates that iterative preference-style tuning of compact large language models (LLMs) on intentionally biased synthetic data produces progressive adversarial sensitisation analogous to psychiatric kindling. Using TinyLlama-1.1B and a Qwen2.5-3B replication across ten alignment cycles and six experimental conditions, the original work shows delayed elevation of weak-prompt jailbreak rates under biased no-mitigation and early-replay-only regimes, with sparse LoRA regrowth combined with diverse replay attenuating the drift. The present review supplies a structural, multi-layered critique: methodological anatomy, quantitative findings, strengths and limitations, comparison against at least ten competing theoretical and empirical models drawn from RLHF, continual learning, safety collapse, reward hacking, dynamic sparse training, and clinical kindling literatures, and an extended analysis of the ontological and epistemological verticals that underwrite the kindling–alignment analogy. Extensions are proposed toward full RLHF pipelines, larger-scale models, real preference data, multi-seed statistical power, and policy-relevant monitoring protocols for iterative alignment systems. All citations follow Harvard style. Also available on Academia.edu: https://www.academia.edu/attachments/133930547/download_file?s=portfolio
Language: English
Open Access status: green
OpenAlex ID: W7214534227
FWCI:
Cites:
Cited by:
Related to: 10