Evaluating Thematic Drift in Long-Context LLM Dialogue via the "Whiteboard Probe"

Abstract

This record contains Revision 4 of the study examining thematic drift in long‑context LLM dialogue using the Whiteboard Probe. The probe is a lightweight conversational method designed to elicit metaphorical externalization of a model’s expressed thematic focus without relying on assumptions about internal state. Through extended dialogue across three configurations of two LLMs (Gemini, ChatGPT Persona Session “YUKAPON,” and ChatGPT New Account / Default Context), the study observes recurring patterns in how models articulate thematic structure, exhibit drift, and reorganize their expressed focus when prompted. The probe provides a natural‑language moment for inspecting thematic alignment, enabling the human interlocutor to identify misalignment and offer clarifications. The study highlights the role of the Observer‑Switching Protocol, in which the human alternates between participant and observer roles to maintain coherence during long‑horizon interaction. All interpretations remain strictly within observable natural‑language behavior; no claims are made about internal mechanisms. Future work may expand model diversity, incorporate quantitative evaluation, explore automated drift detection, and formalize the Observer‑Switching Protocol to deepen understanding of long‑context human–LLM interaction.