Measuring and Mitigating Silent Narrative Switching in Large Language Models through Epistemic Mediation
Open Access
- 1 January 2026
- journal article
- Published by Walter de Gruyter GmbH in Applied Computer Systems
- Vol. 31 (1) , 206-213
Abstract
The paper defines silent narrative switching, the covert change of a large language model’s dominant position across user contexts without disclosing that alternatives exist, and measures it with a Silent Narrative Switching Index (SNSI). Across four models from different providers, SNSI ranged from 0.73 to 0.80. An execution-level epistemic mediation layer over unmodified models reduced the pooled SNSI from 0.82 to 0.09 without retraining and without a detectable loss of contextual relevance; an ablation study locates most of the effect in the interpretation-generation stage.Keywords
This publication has 11 references indexed in Scilit:
- A Survey on Evaluation of Large Language ModelsACM Transactions on Intelligent Systems and Technology, 2024
- Modular Pluralism: Pluralistic Alignment via Multi-LLM Collaboration2024
- The Butterfly Effect of Altering Prompts: How Small Changes and Jailbreaks Affect Large Language Model Performance2024
- Generative Agents: Interactive Simulacra of Human Behavior2023
- ChatGPT for good? On opportunities and challenges of large language models for educationLearning and Individual Differences, 2023
- Out of One, Many: Using Language Models to Simulate Human SamplesPolitical Analysis, 2023
- G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment2023
- Taxonomy of Risks posed by Language Models2022
- Are We There Yet? - A Systematic Literature Review on Chatbots in EducationFrontiers in Artificial Intelligence, 2021
- RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models2020