AQuA preprint: fixed evaluator can still be gamed by agent loops

AQuA preprint distinguishes fixed evaluator from adaptive agent: base LM and evaluator stay fixed, but agent can still adapt to repeated feedback. Reddit discussion highlights this nuance.