推理 / NAVIN RESEARCH
Route Reasoning Effort by Task Difficulty
Build a simple difficulty router with escalation based on confidence and verification rather than one expensive default.
Build a simple difficulty router with escalation based on confidence and verification rather than one expensive default.
简要回答: Start low and escalate when a deterministic verifier fails, required evidence is absent, or the task belongs to a benchmarked hard class. The router should use observable features and never treat self-reported confidence as proof.
实施步骤
- Extract routing features such as tool count, constraint count, input length, and verifier availability.
- Choose effort tiers from measured task performance.
- Run the cheapest tier first when verification is inexpensive.
- Escalate with the original evidence and failure reason, not an ever-growing transcript.
验证清单
- Measure false escalations and missed hard tasks.
- Cap attempts and total spend.
- Log router version so regressions can be attributed.
常见错误
- Routing from prompt length alone.
- Trusting a confidence sentence generated by the same model.
- Retrying identical inputs without new information or changed settings.
实际决策
先交付能够通过上述检查的最小设计。记录测试中使用的模型、提示词、工具、数据集与运行时版本，让后续变化可以被比较而不是被猜测。依赖供应商的限制、模型名称、价格或预览功能之前，请重新检查链接文档；这些细节变化往往快于本文的工程原则。
