Long-Horizon Planning and Goal Consistency

A core challenge for large models acting as agents, referring to the ability of a model to maintain its initial goal and avoid policy drift across many rounds of prolonged interaction. Because autoregressive generation lacks an explicit goal representation, this is widely recognized as a weakness.