跳转至

MACS：用于可靠对话式电商推荐的混合多智能体框架

文章背景与核心概要

在大语言模型（LLM）驱动的对话式电商推荐系统中，许多实际应用场景往往有着极为严格的运营限制：推荐的产品必须完全来自商家的固定目录，不能依赖网络搜索，且不得产生任何无依据的产品虚假宣传。为了解决这一痛点，本文作者推出了 MACS（Multi-Agent Commerce System，多智能体商业系统），这是一个专为固定目录场景设计的混合多智能体框架，旨在实现可靠的对话式推荐。

MACS 的核心设计理念是将面向语言的任务（如用户请求理解、偏好引导和回复生成）交由 LLM 处理，而将对正确性要求极高的操作（如产品检索、硬约束过滤、品牌排除以及渐进式放宽）通过专用的商家智能体进行确定性执行。此外，系统引入了会话持久化偏好层来跨轮次追踪约束。评估结果表明，MACS 在单轮和多轮基准测试中均展现出极高的通过率、完美的品牌合规性以及零约束漂移，尤其在处理复杂的约束累积和排除反转任务时优势显著。

摘要

Conversational e-commerce recommendations mediated by Large Language Models (LLMs) often struggle with strict operational requirements—such as recommending products exclusively from a merchant's fixed catalog without relying on web search or making unsupported product claims.

由大语言模型（LLM）调解的对话式电商推荐系统，往往难以满足严格的运营要求——例如，必须纯粹从商家的固定目录中推荐产品，而不能依赖网络搜索或做出无依据的产品声明。

To address this, the authors introduce MACS (Multi-Agent Commerce System), a hybrid multi-agent framework designed for reliable conversational recommendations in fixed-catalog settings. MACS delegates language-facing tasks (like user request interpretation, preference elicitation, and response generation) to LLMs, while executing correctness-critical operations (such as product retrieval, hard-constraint filtering, brand exclusion, and progressive relaxation) deterministically through a dedicated merchant agent.

为了解决这一问题，作者推出了 MACS（Multi-Agent Commerce System，多智能体商业系统），这是一个专为固定目录场景下的可靠对话式推荐而设计的混合多智能体框架。MACS 将面向语言的任务（如用户请求解释、偏好引导和回复生成）委派给 LLM，同时通过专用的商家智能体确定性地执行对正确性至关重要的操作（如产品检索、硬约束过滤、品牌排除和渐进式放宽）。

Key Findings & Benchmark Performance

  • Single-Turn Benchmark (140 queries): MACS achieves the highest pass rate (87.1%) and perfect brand compliance (1.000).
  • Multi-Turn Benchmark (10 scenarios): MACS secures the strongest macro Pass@5 (72% vs. 56% for GPT+Catalog and 52% for Gemini+Catalog) with zero constraint drift.
  • Specific Advantages: The system performs exceptionally well on complex conversational challenges, notably exclusion reversal (100% vs. 20% / 0%) and constraint accumulation (100% vs. 60% / 40%), while maintaining comparable judged response quality.

关键发现与基准性能

  • 单轮基准测试（140个查询）： MACS 取得了最高的通过率（87.1%）以及完美的品牌合规性（1.000）。
  • 多轮基准测试（10个场景）： MACS 获得了最强的宏平均 Pass@5（72%，相比之下 GPT+Catalog 为 56%，Gemini+Catalog 为 52%），且实现了零约束漂移。
  • 特定优势： 该系统在复杂的对话挑战中表现优异，特别是在排除反转（100% 对比 20% / 0%）和约束累积（100% 对比 60% / 40%）方面，同时保持了相当的评判回复质量。

Metadata & Publication Details

  • arXiv Identifier: arXiv:2608.14068 [cs.IR]
  • Subjects: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI)
  • Submission Date: August 14, 2026
  • Authors: Juli Huang, Hannah Clay, Sajjad Beygi, Thomas Sarda, Negin Golrezaei, Amin Saberi
  • Conference Presentation: Accepted/Presented at the Stanford Trust & Safety Conference and the Stanford Market AI Conference.

元数据与出版详情

  • arXiv 标识符： arXiv:2608.14068 [cs.IR]
  • 学科分类： 信息检索 (cs.IR)；人工智能 (cs.AI)
  • 提交日期： 2026年8月14日
  • 作者： Juli Huang, Hannah Clay, Sajjad Beygi, Thomas Sarda, Negin Golrezaei, Amin Saberi
  • 会议发表： 被斯坦福信任与安全会议（Stanford Trust & Safety Conference）及斯坦福市场人工智能会议（Stanford Market AI Conference）接受并发表。

Abstract

Conversational recommendation for e-commerce is increasingly mediated by large language models (LLMs), yet many real-world deployments operate under a stricter requirement: recommendations must be drawn only from a merchant's fixed catalog, without web search or unsupported product claims. In this setting, the main challenge is reliability under hard constraints: the system must satisfy user requirements, remain grounded in available inventory, and preserve preferences across multiple conversational turns. We present MACS (Multi-Agent Commerce System), a hybrid multi-agent framework for reliable conversational recommendation in fixed-catalog settings. MACS uses LLMs for language-facing tasks such as interpreting user requests, eliciting preferences, and generating responses, while correctness-critical operations, including product retrieval, hard-constraint filtering, brand exclusion, and progressive relaxation, are executed deterministically by the merchant agent. A session-persistent preference layer tracks constraints across turns, enabling consistent handling of budget overwrites and exclusion reversals. On a 140-query single-turn benchmark, MACS achieves the highest pass rate (87.1%) and perfect brand compliance (1.000). On a 10-scenario multi-turn benchmark, MACS achieves the strongest macro Pass@5 (72% vs. 56% GPT+Catalog / 52% Gemini+Catalog) with zero constraint drift. The advantage is sharpest on exclusion reversal (100% vs. 20% / 0%) and constraint accumulation (100% vs. 60% / 40%). Mean judged response quality is similar across systems (0.751 vs. 0.736). These results suggest that hybrid architectures combining deterministic constraint enforcement with session-persistent preference tracking provide stronger reliability-oriented performance than catalog-bound prompt-only baselines in the fixed-catalog merchant setting.

电商领域的对话式推荐正越来越多地由大语言模型（LLM）进行调解，然而许多实际部署运行在更严格的要求之下：推荐必须仅从商家的固定目录中提取，且不能进行网络搜索或产生不受支持的产品声明。在此设定下，主要的挑战在于硬约束下的可靠性：系统必须满足用户需求、立足于现有库存，并在多个对话轮次中保持偏好。我们提出了 MACS（Multi-Agent Commerce System），这是一个用于固定目录设定下可靠对话式推荐的混合多智能体框架。MACS 使用 LLM 来处理面向语言的任务，例如解释用户请求、引导偏好和生成回复，而对正确性至关重要的操作（包括产品检索、硬约束过滤、品牌排除和渐进式放宽）则由商家智能体确定性地执行。会话持久化的偏好层可跨轮次追踪约束，从而实现对预算覆盖和排除反转的一致处理。在一个包含 140 个查询的单轮基准测试中，MACS 实现了最高的通过率（87.1%）和完美的品牌合规性（1.000）。在一个包含 10 个场景的多轮基准测试中，MACS 实现了最强的宏平均 Pass@5（72% 对比 GPT+Catalog 的 56% / Gemini+Catalog 的 52%），且零约束漂移。这种优势在排除反转（100% 对比 20% / 0%）和约束累积（100% 对比 60% / 40%）上最为明显。各系统的平均评判回复质量相近（0.751 对比 0.736）。这些结果表明，在固定目录商家设定中，将确定性约束执行与会话持久化偏好追踪相结合的混合架构，能够提供比仅依赖提示词的目录基线更强的面向可靠性的性能。