Sparse vs. Dense Reasoning

Not all tasks benefit from the same density of explicit reasoning.

Dense reasoning (explicit thought at every step) works well for knowledge-intensive tasks where each step requires verification or synthesis. Question answering and fact verification benefit from thought-action-observation at each turn.

Sparse reasoning (thought only at key decision points) works better for decision-making tasks with many sequential actions. In household navigation and web shopping tasks, thoughts are most valuable for: decomposing goals into subgoals, tracking subgoal completion, determining what to do next, and reasoning about where objects might be found.

The difference maps to task structure. Knowledge tasks have few steps but each step is consequential. Decision tasks have many steps but most are routine, only the strategic inflection points benefit from explicit reasoning.

Related: 05-molecule—thought-action-observation-pattern, 05-atom—reasoning-grounding-tradeoff