Self-Referential Prompts Increase Response Instability in LLMs

A recent investigation published on arXiv (2608.13258) examines the instability of responses in large language models (LLMs) when faced with self-referential inquiries that prompt first-person accounts akin to subjective experiences. This preliminary research, yet to undergo peer review, quantifies instability by calculating one minus the average pairwise cosine similarity of sentence embeddings derived from a condensed core claim in each reply. The analysis contrasts three categories of questions: self-referential prompts, philosophical dilemmas without clear answers, and queries with definitive correct responses. Utilizing 30 independent replies per question (360 in total) through the Gemini API at a temperature of 0.7, the authors observe that self-referential prompts demonstrate the greatest instability (0.343 ± 0.047), suggesting these questions lead to less consistent outputs compared to other open-ended inquiries. This outcome raises concerns regarding the reliability of such responses and enhances our understanding of LLM behavior in introspective-like scenarios, with significant implications for AI alignment and the interpretation of generated subjective states.