Alignment Drift in Long-Term Human-LLM Interaction: A Mechanism-Oriented Framework

In three linesStudy of alignment drift: gradual process where LLM outputs become less constrained by current user message and more shaped by interaction history, while remaining coherent. Proposed mechanism-oriented framework distinguishes signals A/B, explains feedback loops and sub-pattern selection across three interactional regimes.

Summary generated by Claude — human-verified