微软提出长上下文推理自定义Harness无需训练提升性能
Learn to build a custom harness, folks.
给做长上下文优化的同学一个全新思路：不训模型也能大幅提效。这套无训练Harness框架直接给出可复现的拆解与自改进流程，性价比极高，值得尝试。
Learn to build a custom harness, folks.
学会构建自定义的 harness（辅助框架），各位。
This Microsoft work builds a harness to improve long-context reasoning without synthetic data or post-training.
这项来自微软的研究构建了一个 harness，旨在无需合成数据或后训练的情况下提升长上下文推理能力。
(bookmark it)
（收藏它）
They show that Llama3-8B gains 64.18 points on average across long-context benchmarks with no training at all. The gain comes from the harness.
他们表明，Llama3-8B 在长上下文基准测试中平均提升了 64.18 分，且完全未经过训练。这一提升归功于该 harness。
It breaks long-context tasks into capability primitives. A strong agent then rewrites those primitives, the state they track, and the control logic over many rounds of self-improvement on a development set.
它将长上下文任务拆解为能力基元（capability primitives）。随后，一个强大的智能体在开发集上经过多轮自我改进，重写这些基元、它们所跟踪的状态以及控制逻辑。
The harness built for Llama3-8B transfers to Qwen2.5-3B. On HELMET and RULER, it beats NExtLong and PolicyLong, which rely on data synthesis and training, at less than half their estimated cost.
为 Llama3-8B 构建的 harness 可迁移至 Qwen2.5-3B。在 HELMET 和 RULER 基准上，它以不到一半的预估成本超越了依赖数据合成和训练的 NExtLong 与 PolicyLong。
The harness is written by GPT-5.6-Sol, and results drop sharply when the small models build their own.
该 harness 由 GPT-5.6-Sol 编写，当小模型自行构建时，结果会急剧下降。
The takeaway: before post-training a small model for long context, try improving the harness around it.
核心启示：在对小模型进行长上下文后训练之前，先尝试优化其周围的 harness。
https://arxiv.org/abs/2611.10463
Chat with Paper: https://academy.dair.ai/papers/longcapharness-capability-guided-auto-harnessing-for-long-context-reasoning-2611.10463
论文对话：https://academy.dair.ai/papers/longcapharness-capability-guided-auto-harnessing-for-long-context-reasoning-2611.10463
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