elvis(@omarsar0)

Learn to build custom harnesses, folks! Here is an impressive paper showing the impact of a good cu...

8.5内容质量
Learn to build custom harnesses, folks! Here is an impressive paper showing the impact of a good cu...

TL;DR · AI 摘要

定制化harness（LongHarness）可显著提升长上下文任务性能，GLM-5.3-Flash在OfficeQA Pro V2准确率从4.4%跃升至61.1%。

核心要点

  • LongHarness通过显式状态管理使GLM-5.3-Flash在OfficeQA Pro V2准确率提升13倍（4.4%→61.1%）
  • LongHarness在相同基准上超越Claude Code（12.1点）和Codex（18.9点）
  • LongHarness支持100M token以上语料库，适用于大规模长上下文场景

结构提纲

按章节快速跳转。

  1. §引言

    揭示定制harness对长上下文任务的关键影响，引用GLM-5.3-Flash的显著性能提升案例。

  2. ·LongHarness机制

    通过显式分离证据与结论状态，解决标准harness的检索失效问题。

  3. ›实验结果

    GLM-5.3-Flash在OfficeQA Pro V2准确率提升13倍，验证harness有效性。

  4. ·性能对比

    LongHarness在V1/V2基准分别超越Claude Code（12.1点）和Codex（18.9点）。

  5. ›技术特性

    支持100M+ token语料库，模型无关设计适配多种架构。

思维导图

用一张图看清主题之间的关系。

查看大纲文本（无障碍 / 无 JS 友好）
  • LongHarness技术解析
    • 核心机制
      • 显式状态分离（证据/结论）
      • 动态检索决策
    • 性能表现
      • OfficeQA V2 61.1%准确率
      • 超越Codex 18.9点
    • 技术特性
      • 100M+ token支持
      • 模型无关设计

金句 / Highlights

值得收藏与分享的关键句。

#LongHarness#长上下文处理#模型优化#OfficeQA#GLM
打开原文

elvis on X: "Learn to build custom harnesses, folks! Here is an impressive paper showing the impact of a good custom harness on long-context work. GLM-5.3-Flash goes from 4.4% to 61.1% exact match on OfficeQA Pro V2 with the same model. The authors report that the gain comes from the harness. T… / X

elvis

@omarsar0

Learn to build custom harnesses, folks! Here is an impressive paper showing the impact of a good custom harness on long-context work. GLM-5.3-Flash goes from 4.4% to 61.1% exact match on OfficeQA Pro V2 with the same model. The authors report that the gain comes from the harness. The model-agnostic harness name is LongHarness, built for corpora of 100M tokens and more. Why it works: 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. 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. On the same benchmark, it beats Claude Code by 12.1 points on V1 and Codex by 18.9 points on V2.

arxiv.org/abs/2610.13075

Chat with Paper:

academy.dair.ai/papers/longhar…

7:00 PM · Oct 13, 2026

·

18.5K

Views

83

53

376

450