增值税判定中的大模型智能体任务拆解规模优化：一项试点对照实验

文章背景与核心概要

在大语言模型（LLM）智能体架构的演进过程中，一个根本性的设计冲突逐渐浮现：我们应该将任务拆解为众多窄视野、高度专一的智能体，还是交给单个功能强大、具备工具调用能力的通用智能体来处理？本文在跨境增值税（VAT）判定（包含反向征收机制）这一边界清晰的领域内，对这一架构权衡展开了深入研究。在该领域中，每个案例都具备客观的标准答案（Oracle 标签）以及可独立评分的中间决策。

研究在保持整体作业面恒定的前提下——包括子任务、工具、输入输出（I/O）架构、校验机制、编排器、基础模型和合并策略——系统性地调整了子任务在不同工作节点间的分配方式。研究通过总计 4,400 次运行的严格对照实验，探讨了性能表现、帕累托最优性、Token 预算效应以及系统鲁棒性，为大模型智能体系统的任务拆解规模提供了具有参考价值的试点启发式方法与开源研究资产。

执行摘要 (Executive Summary)

Recent developments in Large Language Model (LLM) agent architectures have highlighted a fundamental design conflict: should tasks be decomposed across many narrow, specialized agents, or handled by a single, powerful tool-using agent?

近期大语言模型（LLM）智能体架构的发展凸显了一个根本性的设计冲突：任务究竟应该被拆解并分发给众多狭窄且专一的智能体，还是应该由一个功能强大的、具备工具使用能力的单一智能体来处理？

This paper investigates this architectural trade-off within the bounded domain of cross-border Value Added Tax (VAT) determination (incorporating reverse charges), where every case features an objective oracle label and independently scorable intermediate decisions. The study holds the overall activity surface constant—including subtasks, tools, I/O schemas, validation checks, orchestrators, base models, and merge policies—while systematically varying the distribution of subtasks among workers.

本文在跨境增值税（VAT）判定（包含反向征收规则）这一受限领域内研究了这种架构权衡。在该领域中，每个案例都具有客观的标准标签（Oracle label）以及可独立评分的中间决策。本研究在保持整体作业面（包括子任务、工具、I/O 架构、验证检查、编排器、基础模型和合并策略）不变的同时，系统性地调整了子任务在工作节点之间的分配方式。

Key Findings

Performance: Intermediate configurations (balancing narrow and wide scopes) achieved the highest accuracy at 0.830, compared to endpoints of 0.720 and 0.770. However, because this did not cross the pre-stated statistical threshold against the fine-grained endpoint, the intermediate-optimum hypothesis remains formally unsupported at pilot scale.
Pareto Optimality: A single agent does not universally dominate the orchestrated configurations.
Token Budget Effects: A budget-matched single agent scored 6.5 points below the leading configuration, though the confidence interval includes zero, meaning the advantage is consistent with a prompt-budget explanation.
Robustness (Failure Injection):
Availability faults were successfully absorbed across all granularities, with wide-scope restart over-recovering baseline performance by +0.160.
Conversely, a single schema-conforming hallucinated record degraded every configuration and inverted the performance hierarchy, disproportionately impacting highly fragmented setups.

核心发现

性能表现： 中间配置（平衡了狭窄与宽泛的视野范围）实现了最高准确率 0.830，而两端配置的准确率分别为 0.720 和 0.770。然而，由于该结果未能在与细粒度端点的对比中突破预先设定的统计学显著性阈值，因此在试点规模下，“中间最优假设”在形式上暂未得到充分支持。

帕累托最优性： 单一智能体并未在所有方面普遍优于多智能体编排配置。

Token 预算效应： 预算匹配的单一智能体得分比领先配置低 6.5 个百分点，但由于置信区间包含零，这一优势与提示词预算（prompt-budget）的解释相吻合。

鲁棒性（故障注入）：

可用性故障在所有粒度级别下都被成功吸收，其中宽视野重启甚至超额恢复了基线性能，提升了 +0.160。

相反，单个符合 Schema 规范的幻觉记录会降低所有配置的性能并颠覆性能层级，对高度碎片化的架构造成了不成比例的严重冲击。

Methodology & Experimental Design

Task Domain: Bounded cross-border VAT determination with reverse-charge rules.
Oracle: A deterministic rule engine serving as the ground-truth baseline.
Configurations Tested:
Four orchestrated configurations ranging from 1 wide worker to 5 narrow workers.
\(S_0\): A tuned, no-orchestrator single agent.
Scale of Experiment: Spanned 4,400 total runs, comprising a 40-case, five-repeat main sweep, matched-token arms to isolate prompt-budget vs. agent-count effects, and three dedicated failure-injection arms evaluated against pre-registered falsification criteria.

方法论与实验设计

任务领域： 带有反向征收规则的受限跨境增值税判定。

基准真值（Oracle）： 作为基准真值基线的确定性规则引擎。

测试配置：

四种编排配置，工作节点范围从 1 个宽视野工作节点到 5 个窄视野工作节点不等。

\(S_0\)：经过调优的、无编排器的单一智能体。

实验规模： 共计进行了 4,400 次运行，包括 40 个案例、重复 5 次的主扫描；用于分离提示词预算与智能体数量效应的匹配 Token 实验组；以及针对预注册证伪标准进行评估的三个专用故障注入实验组。

Contributions

The primary contribution of this work is a bounded, pre-registered pilot heuristic for right-sizing task decomposition—suggesting the placement of a partition boundary at the midpoint of the dependency layer.

The complete research artifact has been publicly released, including the rule-engine oracle, evaluation datasets, execution harnesses, raw execution traces, and the analysis pipeline.

研究贡献

本工作的主要贡献是一项用于确定任务拆解规模的、受限且预注册的试点启发式方法——建议将分区边界放置在依赖层的中点位置。

完整的研究产物已向公众开源，包括规则引擎基准、评估数据集、执行框架（execution harnesses）、原始执行轨迹（traces）以及分析流水线。