The serving system identifies eligible repeated tokens and avoids recomputing some of their intermediate state. Cache rules, lifetime, and exact-prefix requirements vary, so applications should treat caching as an optimization rather than correctness logic.