跳转至

塑造你的信息流：基于大语言模型的对话式推荐智能体系统

文章背景与核心概要

传统的推荐系统主要依赖点击率和停留时间等被动行为信号，这导致自动化算法与用户在实时中真正想要表达的需求之间存在鸿沟。为了弥合这一差距，本文推出了 Shape Your Feed (SYF)，这是一个基于大语言模型（LLM）的智能体框架，专为实时、多模态的内容共创而设计。

SYF 采用三层架构——包括感知流（Perception Flow）、服务流（Serving Flow）和自我演进流（Self-Evolution Flow）——允许用户通过文本、语音和 UI 交互主动引导其推荐信息流。该研究已被 RecSys 2026 工业赛道 录用，离线评估和大规模线上 A/B 测试表明，SYF 在生产环境中显着提升了信息流相关性、对齐准确率以及用户情感体验。

Shape Your Feed: An LLM-based Agentic System for Conversational Recommendation

Summary

Traditional recommendation systems primarily rely on passive behavioral signals like clicks and dwell time, creating a gap between automated algorithms and what users actually want to express in real time. To bridge this gap, this paper introduces Shape Your Feed (SYF), an LLM-based agentic framework designed for real-time, multimodal content co-curation. SYF features a three-tier architecture—comprising a Perception Flow, a Serving Flow, and a Self-Evolution Flow—that allows users to actively steer their recommendation feeds using text, voice, and UI interactions. Accepted into the RecSys 2026 Industrial Track, offline evaluations and large-scale online A/B tests demonstrate that SYF significantly improves feed relevance, alignment accuracy, and user sentiment in production environments.

Metadata

Authors

  • Ziyun Xu
  • Bosen Ding
  • Yue Zhang
  • Ji Qi
  • Qingyuan Song
  • Jizhou Huang
  • Liwei Wang
  • Jefferey Santelli
  • Yue Weng
  • Qichao Que
  • Zhenheng Yang
  • Junfeng Pan
  • Linhong Zhu

Abstract

工业推荐系统主要采用被动排序范式，通过隐式行为信号（如点击、停留时间）来推断用户偏好，而不是显式的自然语言输入。因此，用户会发现其显式兴趣与被动行为算法所提供的内容之间存在持续的差异，这限制了他们实时表达细微偏好或引导信息流的能力。

为了解决推荐优化方式与用户表达兴趣愿望之间日益扩大的鸿沟，我们提出了 Shape Your Feed (SYF)，这是一个基于 LLM 的智能体推荐框架，支持实时、多模态的内容共创。SYF 采用了三层架构： 1. 感知流（Perception Flow）： 从文本提示、语音指令和 UI 交互中捕捉细粒度的用户意图。 2. 服务流（Serving Flow）： 基于编码演进用户偏好的持久化语义画像（Semantic Profile），执行实时的智能体重新排序和候选物品剪枝。 3. 自我演进流（Self-Evolution Flow）： 通过直接偏好优化（DPO）和 LLM-as-a-Judge 集成，将系统行为与人类判断相对齐。

Industrial recommendation systems predominantly adopt a passive ranking paradigm that infers user preferences from implicit behavioral signals (e.g., clicks, dwell time) rather than explicit, natural language inputs. As a result, users experience a persistent discrepancy between their explicit interests and what passive behavioral algorithms deliver, limiting their ability to express nuanced preferences or steer their feed in real time.

To address this growing gap between how recommendations are optimized and how users wish to articulate their interests, we present Shape Your Feed (SYF), an LLM-based agentic recommendation framework that enables real-time, multimodal co-curation of content. SYF employs a three-tier architecture: 1. Perception Flow: Captures fine-grained user intent from text prompts, voice commands, and UI interactions. 2. Serving Flow: Performs real-time agentic re-ranking and pruning of candidate items, grounded in a persistent Semantic Profile encoding evolving user preferences. 3. Self-Evolution Flow: Aligns system behavior with human judgments via Direct Preference Optimization (DPO) and an LLM-as-a-Judge ensemble.

离线评估表明，SYF 的对齐评分模块达到了 98.85% 的准确率，大幅优于强少样本基线。对生产流量的大规模线上 A/B 实验进一步证明，SYF 提高了信息流的相关性和用户情感体验，为工业场景中实现交互式、可由用户引导的推荐指明了一条实用且可扩展的道路。

Offline evaluations show that SYF's alignment scoring module achieves 98.85% accuracy, substantially improving over strong few-shot baselines. Large-scale online A/B experiments on production traffic further demonstrate that SYF improves feed relevance and user sentiment, indicating a practical and scalable path toward interactive, user-steerable recommendation in industrial settings.

External References & Tools