跳转至

文章背景与核心概要

快照支持的沙箱技术使得计算分叉（Computational Branching）的成本变得非常低廉，但它们往往忽视了证据依赖（Evidence Dependence）的问题。当并行分支重复使用共享元素（如模型、提示词、代码库、测试、观测数据或执行祖先）时，天真地汇总输出会导致有偏见的共识，从而将单个重复的错误放大为极高的置信度。

为了解决这一问题，该论文引入了一种证据感知归约契约（evidence-aware reduction contract）。该框架要求每个工作节点（worker）报告估计值、估计信息、证据标识符、分叉血缘以及执行元数据。对于估计公共参数的独立工作节点，系统会以高斯/瓦尔德（Gaussian/Wald）形式应用标准的逆信息聚合。生成的固定维度数值摘要可以按任何树状顺序无缝合并，而证据 ID 和血缘则遵循专用的追踪规则。

面向可分叉计算的证据感知 MapReduce

作者： Yossi Eliaz
学科： 人工智能 (cs.AI)；概率论 (math.PR)；统计学理论 (math.ST)
arXiv ID： arXiv:2607.09689 [cs.AI]
许可证： 知识共享署名 4.0 国际许可协议

📌 摘要

Snapshot-backed sandboxes make computational branching inexpensive, but they often leave evidence dependence unaddressed. When parallel branches reuse shared elements—such as models, prompts, repositories, tests, observations, or execution ancestors—naively counting outputs can lead to biased consensus by amplifying a single repeated error into high confidence.

To solve this, the paper introduces an evidence-aware reduction contract. This framework requires each worker to report: * An estimate * Estimated information * Evidence identifiers * Fork lineage * Execution metadata

For independent workers estimating a common parameter, the system applies standard inverse-information pooling in its Gaussian/Wald form. The resulting fixed-dimensional numeric summary merges seamlessly in any tree order, while evidence IDs and lineage adhere to dedicated tracking rules.

方法的核心组件：

Key Components of the Approach:

  • 分歧度量（\(\Delta\)）： 测量工作节点的散度，在标量逆方差情况下简化为科克伦 \(Q\) 统计量（Cochran’s \(Q\) statistic），并自然地显现在乘积积分（product integral）中。
  • 参考实现： 验证序列化记录，拒绝重复的非空证据标识符，通过树归约传播证据和血缘，并依赖基于乔列斯基分解（Cholesky-based）的数值线性代数。

  • 测试与验证： 综合单元测试和带种子的合成检查验证了底层代数、不等信息缩放以及伪造精度处理。包含四个工作节点的命名快照追踪展示了端到端执行路径。

  • Disagreement Metric (\(\Delta\)): Measures worker divergence, reducing to Cochran’s \(Q\) statistic in the scalar inverse-variance case and naturally surfacing in the product integral.

  • Reference Implementation: Validates serialized records, rejects duplicate non-empty evidence identifiers, propagates evidence and lineage through tree reductions, and relies on Cholesky-based numerical linear algebra.
  • Testing & Verification: Comprehensive unit tests and seeded synthetic checks validate the underlying algebra, unequal information scaling, and forged precision handling. A trace involving a four-worker named snapshot demonstrates the end-to-end execution path.

Ultimately, the work highlights a critical open challenge for modern distributed systems: formalizing evidence identity and fork lineage into a robust dependence model capable of handling correlated and adaptively selected AI execution branches.

🔗 链接与资源

📅 提交历史

📅 Submission History

  • [v1] 2026年6月17日 星期三 16:26:18 UTC (538 KB)
  • [v2] 2026年7月14日 星期二 20:12:20 UTC (57 KB)
  • [v3] 2026年7月17日 星期五 06:58:06 UTC (13 KB)
  • [v4] 2026年8月22日 星期六 02:29:54 UTC (13 KB) — 当前版本
  • [v1] Wed, 17 Jun 2026 16:26:18 UTC (538 KB)
  • [v2] Tue, 14 Jul 2026 20:12:20 UTC (57 KB)
  • [v3] Fri, 17 Jul 2026 06:58:06 UTC (13 KB)
  • [v4] Sat, 22 Aug 2026 02:29:54 UTC (13 KB) — Current Version