微软新论文：用蒸馏技能替代昂贵测试时推理
What if you could pay the reasoning cost once, then reuse what the model learned…
What if you could pay the reasoning cost once, then reuse what the model learned across future tasks?
如果你能只支付一次推理成本，然后在未来的任务中重用模型学到的内容，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.
微软的新论文发现，一些昂贵的测试时推理可以用从先前代理运行中学习到的一小组规则来替代。
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条过去的轨迹，让编码代理提取重复出现的失败模式，然后将这些模式转化为添加到非推理模型系统提示中的小型markdown技能。
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.
对于GPT-5.4-mini，这些技能在4个代理基准上恢复了非推理和推理模式之间差距的55%–100%以上，同时使用的输出令牌比推理少2.9–4.5倍。
On ALFWorld and τ²-retail, the skilled non-reasoning model actually beat the reasoning mode.
在ALFWorld和τ²-retail上，具备技能的非推理模型实际上击败了推理模式。
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.
有用之处在于，蒸馏器不需要推理轨迹：仅从廉价的非推理回滚中构建的技能在所有4个领域都具有竞争力。
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.
局限性同样有用：在电信和SpreadsheetBench上，推理仍然胜出，因为每个任务包含更多实例特定的依赖关系，固定技能无法捕捉。
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，实际的分工是：一次性蒸馏重复的程序，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。
– arxiv. org/abs/2608.07885
– arxiv.org/abs/2608.07885
Title: "Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills"
标题：“宽推理而非深推理：将推理溢价摊销为蒸馏技能”
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