LongHarness:通过显式状态管理实现百万token级长上下文推理
Learn to build custom harnesses, folks!
Agent开发者的硬核干货,直接给出了百万token场景下解决长上下文失效的具体架构方案,参数和逻辑清晰可落地,建议做RAG或长文本Agent的同学重点参考。
Learn to build custom harnesses, folks!
学会构建自定义 harness,各位!
Here is an impressive paper showing the impact of a good custom harness on long-context work.
这里有一篇令人印象深刻的论文,展示了良好的自定义 harness 对长上下文工作的影响。
GLM-5.3-Flash goes from 4.4% to 61.1% exact match on OfficeQA Pro V2 with the same model.
使用相同的模型,GLM-5.3-Flash 在 OfficeQA Pro V2 上的精确匹配率从 4.4% 提升至 61.1%。
The authors report that the gain comes from the harness.
作者指出,这种提升来自于 harness。
The model-agnostic harness name is LongHarness, built for corpora of 100M tokens and more.
该与模型无关的 harness 名为 LongHarness,专为包含 1 亿 token 及以上的语料库构建。
Why it works:
其工作原理:
Standard agent harnesses keep evidence and conclusions mixed into one growing chat history. Retrieval keeps expanding without closing the actual gap, and corrections stop reaching the answer.
标准的 agent harness 将证据和结论混合在一个不断增长的聊天历史中。检索过程不断扩展,但并未真正缩小差距,且修正信息无法有效传达至答案。
LongHarness stores both as explicit state. Open conclusions decide what to retrieve next, and each new piece of evidence confirms or revises the conclusions that asked for it.
LongHarness 将两者作为显式状态存储。开放的结论决定下一步检索什么,而每一条新证据都会确认或修订提出该请求时的结论。
On the same benchmark, it beats Claude Code by 12.1 points on V1 and Codex by 18.9 points on V2.
在同一基准测试中,它在 V1 上比 Claude Code 高出 12.1 分,在 V2 上比 Codex 高出 18.9 分。
https://arxiv.org/abs/2610.13075
Chat with Paper: https://academy.dair.ai/papers/longharness-advancing-agentic-reasoning-over-100m-token-corpora-via-stateful-evi-2610.13075
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