Instruction Drift in Long Conversations

A conversation may begin with a persistent request—such as “use plain language,” “keep each answer under 100 words,” or “do not reveal a customer’s address”—then accumulate new messages, examples, corrections, and tool results. Instruction drift occurs when later responses stop honoring a relevant earlier constraint or reinterpret it inconsistently. This is a measurable behavior, not a universal fixed rate. Li et al.’s 2024 study introduced a benchmark based on self-chats between instructed chatbots and reported drift in tested systems within eight conversation rounds. The tested models and experimental setup define the claim; it should not be assumed to predict all current models or every kind of conversation. Other multi-turn research also finds that task performance and reliability can differ between fully specified one-shot instructions and instructions revealed gradually. Long interactions can combine several challenges: older constraints may compete with newer requests, context may be lengthy, and the user’s goal may evolve. Not every apparent drift is a model failure: the user may have changed the instruction, or a newer instruction may legitimately supersede an older one. Evaluation should specify which rules persist, which can change, and how conflicts are resolved. For important tasks, restate the active requirements in a concise checkpoint, ask the assistant to summarize them, and verify outputs against a checklist. In software or agent workflows, enforce critical format, access, and safety rules in code or tools rather than relying only on conversational instructions. Treat summaries and reminders as helpful controls, not guarantees.