微软研究：不改训练，靠 harness 让 Llama3-8B 长上下文平均提升 64.18 分

技术路径是把长上下文任务拆解为若干能力原语（capability primitives），再由一个强 agent 在开发集上多轮自我改进，重写这些原语、它们追踪的状态以及控制逻辑。这套 harness 本身由 GPT-5.6-Sol 编写，实验显示如果让小模型自己生成 harness，成绩会明显下降。

这套为 Llama3-8B 构建的 harness 可以直接迁移到 Qwen2.5-3B。在 HELMET 和 RULER 两个基准上，它的表现超过依赖数据合成与训练的 NExtLong 和 PolicyLong，估算成本不到后者的一半。论文作者给出的建议是：给小模型做长上下文后训练之前，先尝试改进模型外围的 harness。

Learn to build a custom harness, folks. This Microsoft work builds a harness to improve long-context reasoning without synthetic data or post-training. (bookmark it) 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. 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. 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. The harness is written by GPT-5.6-Sol, and results drop sharply when the small models build their own. The takeaway: before post-training a small model for long context, try improving the harness around it. arxiv.org/abs/2611.10463 Chat with Paper: academy.dair.ai/papers/longcap…📊 21 ⚡

本条情报归属 AITOP「论文」核心栏目，重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中，该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：Llama3-8B、Qwen2.5、HELMET、RULER，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。

信源核验说明：本文原始线索采编自「elvis」（@omarsar0），经 AITOP 多维评分管线结构化提炼。