微软论文：用规则蒸馏替代部分测试时推理以降本增效
What if you could pay the reasoning cost once, then reuse what the model learned…
Agent 工程必读，这篇论文给出了极具实操价值的“蒸馏+推理”混合架构思路，能显著降低 Token 成本并提升常规任务效率，建议收藏参考其技能提取流程。
What if you could pay the reasoning cost once, then reuse what the model learned across future tasks?
如果只需支付一次推理成本，就能让模型在未来任务中复用其所学内容，会怎样？
New Microsoft paper finds that some expensive test-time reasoning can be replaced with a small set of rules learned from previous agent runs.
微软最新论文发现，部分昂贵的测试时推理（test-time reasoning）可被一组从先前智能体运行中学习到的规则所替代。
The paper tests a cheaper alternative: collect 35–50 past trajectories, have a coding agent extract recurring failure patterns, then turn those patterns into a small markdown skill added to the non-reasoning model’s system prompt.
该论文测试了一种更便宜的替代方案：收集35–50条历史轨迹，由编码智能体提取反复出现的失败模式，然后将这些模式转化为小型markdown技能，加入非推理模型的系统提示中。
For GPT-5.4-mini, those skills recovered 55%–100%+ of the gap between non-reasoning and reasoning modes across 4 agent benchmarks, while using 2.9–4.5× fewer output tokens than reasoning.
在GPT-5.4-mini上，这些技能在4个智能体基准测试中恢复了非推理模式与推理模式之间55%–100%+的性能差距，同时输出token数仅为推理模式的2.9–4.5倍更少。
On ALFWorld and τ²-retail, the skilled non-reasoning model actually beat the reasoning mode.
在ALFWorld和τ²-retail上，具备技能的无推理模型实际上超越了推理模式。
The useful part is that the distiller did not need reasoning traces: skills built only from cheap non-reasoning rollouts were competitive across all 4 domains.
值得指出的是，蒸馏器无需推理轨迹：仅基于低成本的非推理 rollout 构建的技能在所有4个领域均具有竞争力。
The limit is equally useful: reasoning still won on telecom and SpreadsheetBench, where each task contains more instance-specific dependencies that a fixed skill cannot capture.
同样值得注意的是其局限性：在telecom和SpreadsheetBench上，推理模式仍占优势，因为每个任务包含更多固定技能无法捕捉的实例特定依赖关系。
So the practical split is: distill repeated procedures once, then reserve expensive test-time reasoning for the tasks that genuinely need fresh search.
因此，实际分工是：一次性提炼重复流程，然后将昂贵的测试时推理保留给真正需要全新搜索的任务。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力