Pick the most reliable cause of persona drift across a long conversation

**Persona drift** is the slow shift in an assistant's voice, tone, or behavior over the course of a long conversation. The bot starts in character and gradually slips out of it. Four candidate causes sound reasonable, but only one is the dominant structural cause. **System prompt diluted by the long tail.** The persona lives in the system prompt at the top of the context. As the conversation grows, the user-assistant exchanges fill the middle and bottom. The persona instructions become a smaller and smaller share of what the model is reading. Their **competitive attention weight** drops, and the model's behavior shifts toward whatever the recent tail looks like. This is structural: it happens by construction whenever the conversation gets long enough, regardless of temperature or user intent. **Low-probability sampling.** Stochastic sampling can amplify drift but is not the dominant cause. Drift happens even at temperature zero in long conversations. **Temperature too high.** Temperature affects token-level variety, which can speed up drift, but again is not the structural driver. **Adversarial users.** Jailbreak attempts cause persona breaks, but those are sharp and visible, not the slow drift this question is about. The correct answer is the structural one because it is the cause that operates **on every conversation**, not just adversarial ones, and not just high-temperature ones. The fix is also structural: periodically restate the persona, mirror its key rules near the bottom of the user turn, or extract its load-bearing constraints into a small mid-conversation reminder block.

**Persona drift** is the slow shift in an assistant's voice, tone, or behavior over the course of a long conversation, distinct from the sharp persona breaks that adversarial inputs produce. The dominant cause is structural: **the system prompt sits in the privileged primacy slot at the top of the context, but the long user-assistant tail dilutes its influence**. The persona definition has a fixed token cost. The conversation history grows turn by turn. At turn five, the persona might be 30 percent of the context. At turn fifty, it is three percent. The model still reads the persona, but the competitive attention weight against the recent tail keeps dropping, and the model's behavior shifts toward whatever pattern the tail establishes. The other three options are real effects but are not the dominant cause of slow drift. **Low-probability sampling** (option 1) can amplify drift by occasionally selecting an off-persona token, but drift happens at temperature zero in long conversations too. Stochastic sampling is not the structural driver. **Temperature too high** (option 2) increases token-level variability and can speed up drift, but temperature alone does not cause drift if the persona's attention share stays high. A 4,000-token persona prompt at temperature 0.7 in a five-turn conversation does not drift; the same prompt at the same temperature in a fifty-turn conversation does. **Adversarial users** (option 3) cause sharp persona breaks, not slow drift. The two phenomena have different signatures: jailbreaks are visible in a single turn, drift accumulates over many. The operational fixes target the structural cause. **Mid-conversation persona reminders**: periodically re-inject a short version of the persona rules just before the most recent user turn. **Trailing critical rules**: mirror the most load-bearing persona constraints at the bottom of the user turn so they sit in the recency slot.