When an AI plans its approach before executing — instead of improvising both at once — its accuracy on multi-step problems measurably improves.

This explores why splitting the 'what steps do I take' planner from the 'do this step' executor makes multi-step reasoning more reliable — rather than having a single model improvise plan and answer in one breath. The clearest evidence is direct: when you train a separate decomposer and solver instead of one monolithic model, accuracy goes up, and — more surprisingly — the *decomposition* skill transfers across domains while solving does not Does separating planning from execution improve reasoning accuracy?. That asymmetry is the tell. Knowing how to break a problem apart is a general skill; grinding out each sub-step is local. When one model carries both, the two interfere, and the general skill gets contaminated by the specific one.

The mechanism behind the gain seems to be *isolation* — keeping each step from being polluted by everything around it. LLM Programs make this explicit: wrap the model inside an algorithm that hands each call only the context relevant to that one step, hiding the rest Can algorithms control LLM reasoning better than LLMs alone?. Cognitive tools do the same by implementing each reasoning operation as its own sandboxed call, and just enforcing that separation lifted GPT-4.1 on competition math from 27% to 43% with no additional training Can modular cognitive tools unlock reasoning without training?. The capability was already there; what unlocked it was refusing to let steps bleed into each other. Decoupling reasoning from tool observations lands in the same place from a different angle — plan the whole chain before executing, and you kill the redundant re-reading and quadratic prompt bloat that creep in when planning and doing are interleaved Can reasoning and tool execution be truly decoupled?.

Why does interleaving hurt so much? Because reasoning models left to plan-and-execute in one pass fail *structurally*, not for lack of compute. They wander down invalid paths and abandon promising ones prematurely — the 'underthinking' failure — even when a viable solution was within reach Why do reasoning models abandon promising solution paths?. A separate decomposer acts as a commitment device: it lays out the shape of the search before the solver can impulsively bail. RLAD makes this concrete by generating explicit abstractions first, which force breadth-first exploration and prevent exactly that depth-only collapse Can abstractions guide exploration better than depth alone?.

There's a memory angle worth noticing too. Atom of Thoughts decomposes a problem into a dependency graph and then contracts it so each state depends only on the current sub-problem — not the full accumulated history — which strips out the 'historical baggage' that bloats and destabilizes long reasoning while keeping the answer equivalent Can reasoning systems forget history without losing coherence?. Separation isn't only about who plans versus who executes; it's about not dragging the entire trace through every step. That's the same instinct as feeding each call only step-specific context — accuracy is protected by controlling what each stage is allowed to see.

The thing you might not have expected: the benefit isn't that decomposition adds reasoning power. Several notes suggest the raw capability is already latent, and the models fail mostly through disorganization — abandoning good paths, re-consuming irrelevant context, imitating reasoning form without holding the thread Can modular cognitive tools unlock reasoning without training? Why do reasoning models abandon promising solution paths?. Separating decomposition from execution is less about making the model smarter and more about giving structure a place to live outside the model, so its existing intelligence stops tripping over itself.