Paper 5 · Question 49

A: Reminders every few turns fight a 2,800-token rule block instead of replacing it.

B: A 2,800-token system prompt with declarative rules is vulnerable to drift because abstract rules require the model to reason about them on every turn. Replacing verbose rules with concrete few-shot examples that demonstrate correct proficiency-level adaptation gives the model clear behavioral patterns to match—this is more reliably followed across many turns than abstract instructions. Reminder injection (A) helps but addresses symptoms; end-placement (C) helps initially but not with turn-level drift; regeneration (D) is expensive and corrective.

C: Moving rules to the end of the system prompt does not make an abstract methodology easier to follow on turn 12.

D: A regenerate step catches a mismatch late. It does not change the rules the model is drifting from, and it adds a call when they fail.