AI planning systems work best when they commit to a direction for just the right number of steps — not too few, not too many.
Researchers studying "latent reasoning" — where an AI's multi-step thinking happens inside hidden internal states rather than visible token-by-token output — extended a model called the Hierarchical Reasoning Model with a manager-worker structure. A slow, high-level module emits a directional subgoal that guides a faster low-level module for P steps before the high-level module can revise it. On two abstract reasoning benchmarks, ARC and ConceptARC, the team found that holding a subgoal for 3 to 6 steps consistently beat both very short persistence (P=1) and longer horizons. The loss minimum landed at P=3 — a score of 1.544 versus 1.674 at P=1 and 1.640 for the no-persistence baseline.
The result matters because it identifies plan persistence — not just the presence of a hierarchical structure — as the active ingredient. A complementary finding pinned the optimal alignment-loss weight at roughly 0.05; push past that and the directional signal starts interfering with learning rather than helping it. That narrow double optimum suggests latent planners are surprisingly fragile to tuning choices that look minor on paper.
The AI planning field has long debated how to balance commitment against flexibility — a problem familiar from robotics and classical search. This work puts a number on it, at least for latent-space systems: medium-horizon intent needs to persist across enough steps for compositional structure to actually form, but the window is narrower than most architectures currently assume.