Longer reasoning traces do not necessarily teach a model to solve harder problems. The useful shift here is to make the dependency structure explicit: demonstrate how to decompose a task, then let each solved subproblem become context for the next. This targets easy-to-hard generalization rather than merely matching the difficulty of prompt examples.
Key Findings
- On SCAN, code-davinci-002 achieves at least 99% accuracy across splits with 14 exemplars; the paper contrasts this with 16% for chain-of-thought prompting. The result concerns this model and benchmark setup, rather than a universal advantage.
- Experiments cover symbolic manipulation, compositional generalization, and mathematical reasoning, testing whether simple demonstrations can support more complex inputs.
- Both decomposition and solving use few-shot prompts without training or fine-tuning, making the intervention a change in inference procedure rather than model weights.
Who Should Read It
Great fit if you study compositional generalization or design workflows whose intermediate answers can be reused. Look elsewhere if you need evidence for open-ended agent reliability: these are controlled reasoning tasks, and sequential solving adds model calls while making later answers depend on earlier ones.