A reasoning trace can show the right sequence of steps while still failing to teach any difficult step well. Separating the procedure from its subtask demonstrations lets each component receive targeted examples, turning prompt design into a modular composition problem.
Key Findings
- Separate prompts teach decomposition and subtask execution independently; a difficult primitive no longer has to be learned inside a full reasoning trace.
- Recursive handlers break difficult operations or long inputs into smaller instances, supporting composition beyond a single prompt.
- Symbolic retrieval can replace an LLM handler. GPT-3 experiments report gains over prior few-shot approaches on symbolic reasoning and multi-hop QA.
Who Should Read It
Great fit if you design multi-step LLM workflows and can define reusable subtasks with their own demonstrations. Look elsewhere if you need evidence for unrestricted autonomous task discovery or current-model performance: the reported results use GPT-3, and effective decomposition still depends on the chosen handlers and examples.