从单体融合到 Agent 编排：大规模对话助手的动态响应架构实践
英文标题：From Monolithic Blending to Agentic Orchestration: Dynamic Response for Conversational Assistants at Scale
本文探讨了大型住宿平台在月均数百万次对话的真实生产场景下，将客服助手从全能单体模型向 Agent 编排架构迁移的实践。团队采用“动态响应”（Dynamic Response）系统，以基于类型化工具的有界 ReAct 编排器配合小型上下文生成器，替代了原有的单一 Qwen3-235B-A22B 混合模型。架构重构使实体选择准确率从 8.3% 暴增至 89.1%，彻底消除了结构化操作幻觉（2.14% 降至 0%），并显著降低了转人工升级率。同时，优化后的自托管方案将编排器 P90 延迟压缩至 2.24 秒并节省三分之一 GPU 资源，预估年服务成本降低超过一个数量级，为工业级智能助手的大规模工程落地提供了标杆范式。
2.1 KB
612 B
380 B
9.1 KB
32.7 KB
22.8 KB
24.2 KB
34.7 KB
中文
[于 2026 年 9 月 4 日提交 (v1)，最后修改于 2026 年 9 月 9 日 (当前版本，v2)]
标题：从单体融合到 Agent 编排：大规模对话助手的动态响应架构 (From Monolithic Blending to Agentic Orchestration: Dynamic Response for Conversational Assistants at Scale)
查看 Cen Mia Zhao 等 9 位作者撰写的论文《From Monolithic Blending to Agentic Orchestration: Dynamic Response for Conversational Assistants at Scale》PDF 全文
摘要：对话助手既可以将检索、动作选择、问题升级（转人工）与语言表达融合成单一的模型路径，也可以将这些职责明确解耦拆分。我们报告了一家大型住宿预订平台（每月数百万次对话、覆盖 11 种语言、P90 响应时间 10 秒）客服助手的生产环境架构迁移。动态响应（Dynamic Response, 简称 DR）方案使用运行在类型化工具上的有界 ReAct 编排器，配合一个根据后端验证上下文契约进行文本生成的小型生成器，替换了原先单一的 Qwen3-235B-A22B 混合应答器。由于本次生产迁移同时涉及了 Prompt、模型对齐以及推理服务的变更，我们对每种效果进行了归因分析，并仅将在完全相同的回放轮次中测量到的改进归结为架构效益：类型化实体选择将预订选择器切换至“精准率优先”的操作区间（精准率从 8.3% 跃升至 89.1%，召回率由 75.2% 微调至 67.3%），而带有成员资格校验的类型化动作 ID 则彻底消除了已观测到的结构化动作幻觉（从 2.14% 降至 0.0%）。低比例流量灰度 A/B 测试复现了重放评估中的转人工比例下降：强升级（hard-escalation）回复比例从 5.60% 下降至 3.08%，软升级（soft-escalation）回复从 9.56% 下降至 2.49%，同时生产环境的人工交接总量保持基本稳定；自助解决率呈现正向改善趋势（+5.1 个百分点，95% 置信区间 [-2, +12]）。在服务优化方面，编排器 P90 延迟在 GPU 资源占用减少约三分之一的前提下，从 3.87 秒缩减至 2.24 秒，且自托管方案将预估的年度模型推理成本降低了一个数量级以上。
提交历史
来自：Cen Mia Zhao [查看邮箱]
[v1] 2026 年 9 月 4 日，星期五 22:38:10 UTC (43 KB)
[v2] 2026 年 9 月 9 日，星期三 01:43:26 UTC (43 KB)
English / 原文
[Submitted on 4 Sep 2026 (
), last revised 9 Sep 2026 (this version, v2)]
Title:From Monolithic Blending to Agentic Orchestration: Dynamic Response for Conversational Assistants at Scale
View a PDF of the paper titled From Monolithic Blending to Agentic Orchestration: Dynamic Response for Conversational Assistants at Scale, by Cen Mia Zhao and 8 other authors
Abstract:Conversational assistants can blend retrieval, action selection, escalation, and wording in a single model path, or separate those roles. We report a production migration of a customer-support assistant at a large accommodation marketplace (millions of conversations per month, 11 languages, 10-second P90). Dynamic Response (DR) replaces a single Qwen3-235B-A22B blended responder with a bounded ReAct orchestrator over typed tools plus a smaller generator that writes from a backend-validated context contract. Because the migration also changed prompts, alignment, and serving, we attribute each effect to its cause and claim as architecture effects only those measured on identical replayed turns: typed entity selection moves the reservation selector to a precision-first operating point (precision 8.3% to 89.1%, recall 75.2% to 67.3%), and typed action IDs with a membership check remove observed structured-action hallucination (2.14% to 0.0%). A low-ramp A/B test reproduces the replay escalation reductions: hard-escalation responses fall from 5.60% to 3.08% and soft-escalation responses from 9.56% to 2.49%, while production handoff volume holds roughly steady; self-solve is directional (+5.1 points, 95% CI [-2, +12]). Serving optimizations cut orchestrator P90 latency from 3.87s to 2.24s on a GPU footprint reduced by roughly one-third, and self-hosting reduces estimated annual model-serving cost by more than an order of magnitude.
Submission history
From: Cen Mia Zhao [
]
Fri, 4 Sep 2026 22:38:10 UTC (43 KB)
[v2]
Wed, 9 Sep 2026 01:43:26 UTC (43 KB)