The intuitive workflow for AI-assisted code generation: a reasoning model plans the approach, then a code model implements the plan. Specification flows downward from abstract to concrete. The reasoning model is the architect; the code model is the builder.
Miller found that reversing the order works better.
When a code model generates first and a reasoning model reviews after, the combined system achieves 90.2% pass@1 on HumanEval+ — outperforming GPT-4o (87.2%) and O1 Preview (89.0%). This runs on commodity hardware at roughly $2/hour. The improvement scales 4x more on richly-specified problems, where the review model has more specification to check against.
The structural reason: planning is bottlenecked by specification. A reasoning model working from a natural-language specification must translate intent into code structure — and this translation is where errors enter. The specification is ambiguous, incomplete, or subtly wrong, and the planning model propagates those problems into its plan.
Review operates on concrete artifacts. The code exists. The specification exists. The review model's job is comparison: does this code satisfy this specification? This is a strictly easier cognitive task than generation-from-specification. The review model doesn't need to invent structure — it needs to check alignment between two already-existing objects.
The broader implication: the bottleneck in code synthesis is not generation but critique. Improving the upstream planning model provides diminishing returns because the planning step introduces errors that downstream implementation cannot correct. Improving the downstream review step catches errors that upstream planning let through, and the catch rate increases with specification richness.
This inverts the assumed value chain. In human software engineering, we invest heavily in requirements and design (upstream) and treat testing and review (downstream) as verification. The AI workflow reverses the leverage: generation is cheap and fast, but generation-plus-review outperforms expensive-generation-alone. The review is not a check. It is the mechanism that converts a rough draft into a correct program.
Planning is overrated. Critique does the work.