long-horizon reasoning

Long-horizon reasoning is staying coherent and on-target across many steps, the kind of task that unfolds over dozens of moves rather than one reply. Proving a long theorem, debugging across a sprawling codebase, or running a multi-day research project all demand that the thread from goal to current action never snap. The challenge is less about any single step, which the model may handle easily, and more about not losing the plot, contradicting an earlier decision, or quietly forgetting the original objective along the way.

Two pressures make this hard. First, small errors compound: a step that is ninety-nine percent reliable still drifts badly over hundreds of steps, so reliability that looks fine on short tasks crumbles on long ones. Second, the relevant facts may scroll out of the context window, so the model literally cannot see decisions it made earlier. Remedies include external memory, periodic summarisation, checking against an explicit plan, and acting then re-grounding in real results rather than reasoning in one unbroken stretch. Long-horizon competence is a frontier precisely because it tests endurance, not just cleverness.