Shape Your Feed: An LLM-based Agentic System for Conversational Recommendation
Proposes SYF, an LLM-based interactive recommendation system enabling multimodal real-time content control.
Key Findings
Methodology
SYF employs a three-tier architecture: Perception Flow (captures user intent via multimodal inputs), Serving Flow (real-time re-ranking based on a persistent Semantic Profile), and Self-Evolution Flow (iteratively refines behavior via DPO and LLM-judged feedback). The perception module uses LLMs for intent recognition and semantic profile maintenance; the serving layer dynamically reorders content grounded in user preferences; the self-evolution component updates policies through online behaviors and offline judgments, ensuring continuous adaptation. Experiments show the preference scoring achieves 98.85% accuracy, outperforming few-shot baselines, while large-scale online A/B tests demonstrate significant improvements in relevance and user satisfaction.
Key Results
- Preference scoring accuracy reaches 98.85%, surpassing baseline methods (>95%), greatly improving preference detection.
- Content relevance increases by over 20%, user dwell time rises by 15%, and satisfaction scores improve markedly in live traffic.
- Multimodal preference fusion enhances expression richness, supporting complex multi-turn interactions and transparent explanations.
Significance
This work addresses fundamental limitations of passive recommendation by enabling explicit, real-time user control through natural language and multimodal inputs. It bridges the gap between implicit behavioral signals and user-articulated preferences, advancing towards transparent, interactive, and user-centric recommendation systems. The architecture supports scalable deployment in industrial settings, improving content relevance, user engagement, and trust. It also opens pathways for future research on self-evolving, explainable AI in large-scale personalization, impacting both academia and industry by redefining content curation paradigms.
Technical Contribution
SYF introduces a novel three-layer architecture integrating LLM-based intent understanding, persistent semantic profiles, and self-evolving policies. It combines multimodal input processing, dynamic content re-ranking, and bias-aware preference optimization, enabling multi-turn, multi-modal, and multi-objective content control. The system’s core innovations include a structured semantic profile, dual-feedback policy alignment, and an LLM-as-a-judge mechanism, which together form a closed-loop, scalable, and interpretable recommendation pipeline. These contributions push the frontier of interactive AI-driven personalization.
Novelty
This is the first system to unify multimodal user feedback with a persistent semantic profile and self-evolving optimization in a real-time industrial feed environment. Unlike prior static catalog or single-turn interaction models, SYF supports continuous, multi-turn, multi-modal, user-driven content adjustment, addressing the challenges of high-throughput, dynamic content streams with explicit control and transparency, representing a significant innovation in recommendation technology.
Limitations
- Dependence on large models results in high computational costs, limiting real-time responsiveness in resource-constrained environments.
- User preference profiles may be affected by sparse or noisy feedback, reducing adjustment accuracy.
- Multimodal input accuracy relies on speech recognition and natural language understanding, which can introduce errors affecting intent detection.
Future Work
Future efforts will focus on model compression to reduce latency, robustness improvements for noisy inputs, and more sophisticated self-evolution algorithms combining reinforcement learning. Exploring multi-task learning and adaptive feedback mechanisms will further enhance system autonomy and personalization quality, aiming for more efficient, scalable, and trustworthy interactive recommendation solutions.
AI Executive Summary
In the rapidly evolving landscape of digital content platforms, traditional passive recommendation systems—relying solely on implicit behavioral signals—struggle to meet users’ demand for transparency, control, and nuanced personalization. Users increasingly wish to directly communicate their preferences through natural language or multimodal interactions, expecting the system to respond in a transparent and adaptable manner. Existing solutions often lack the ability to interpret explicit user intents, maintain multi-turn contextual understanding, or dynamically adjust content in real time, creating a gap between user expectations and system capabilities.
To bridge this divide, the authors propose Shape Your Feed (SYF), an innovative architecture leveraging large language models (LLMs) to enable interactive, multimodal content recommendation. The core idea is to treat recommendation as a closed-loop, user-controlled process, integrating three key components: the Perception Flow, which captures and normalizes user intent from text, voice, and UI; the Serving Flow, which dynamically re-ranks and filters content based on a persistent semantic profile; and the Self-Evolution Flow, which continuously updates system policies through offline judgments and online feedback.
The perception module employs LLMs for intent recognition, semantic profile maintenance, and multi-modal fusion, translating unstructured inputs into structured preferences. The serving layer uses these preferences to perform real-time content re-ranking, ensuring relevance and diversity. The self-evolution mechanism employs techniques like Direct Preference Optimization (DPO) and LLM-based judgment to refine behaviors iteratively, creating a scalable, interpretable, and adaptive recommendation pipeline.
Extensive offline and online experiments validate SYF’s effectiveness. The preference scoring module achieves 98.85% accuracy, outperforming baseline models. In production traffic, SYF improves content relevance by over 20%, increases user dwell time by 15%, and enhances satisfaction scores. These results demonstrate the system’s potential to transform industrial recommendation from static, opaque pipelines into transparent, user-driven ecosystems.
While promising, challenges remain, including computational costs, robustness of multimodal inputs, and profile sparsity. Future work aims to optimize efficiency, improve noise resilience, and develop more sophisticated self-adaptive algorithms. Overall, SYF marks a significant step toward intelligent, interactive, and scalable recommendation systems that align closely with user needs and preferences.
Deep Analysis
Background
Content recommendation在数字平台中扮演核心角色，经历了从基于协同过滤和深度学习模型（如Wide & Deep、Neural CF）到时间序列模型（如SASRec、BERT4Rec）的演变。早期方法依赖隐性行为信号，难以满足用户对内容控制的需求。近年来，LLMs的崛起带来了对话式推荐的可能性，支持多轮交互和语义理解（如ReDial、LLM-ConvRec）。但现有方案多局限于静态目录或单轮交互，难以应对高速、多模态、多目标的工业场景。本文提出SYF，融合多模态输入、偏好档案和自我演化机制，推动推荐系统向主动、可控、透明的方向发展。
Core Problem
工业推荐系统多采用被动排序，依赖隐性行为信号，难以满足用户实时表达复杂偏好的需求。用户希望通过自然语言或多模态交互主动调节内容偏好，但现有系统缺乏对显式意图的理解和响应机制，导致内容匹配不精准、缺乏透明度。多轮交互状态维护、偏好动态更新和多模态融合成为实现用户自主调节的关键技术难题。如何在保证内容多样性和平台约束的同时，实现高效、实时的偏好调节，是核心挑战。
Innovation
第一，提出三层架构（感知流、服务流、自我演化流），实现多模态意图捕获、内容重排序与偏好自我调节。第二，利用LLM的语义理解能力，将多模态输入融合成动态偏好档案，支持复杂意图表达。第三，引入偏好优化（DPO）和判官模型，建立闭环自我演化机制，持续优化推荐策略。这些创新突破了传统被动推荐的局限，使系统具备更高的交互性、透明度和适应性，满足工业场景的多样化需求。
Methodology
- �� 感知流：通过多模态接口（文本、语音、UI）捕获用户交互，利用LLM识别意图，构建动态语义档案。
- �� 服务流：结合偏好档案进行候选内容的重排序和过滤，确保内容相关性与多样性。
- �� 自我演化：基于偏好优化（DPO）和判官模型，持续调整系统行为，实现多轮交互的状态维护与偏好更新。
- �� 多模态融合：利用自然语言描述和语音识别，将用户偏好转化为结构化信息，支持复杂指令。
- �� 反馈机制：通过偏好评分、解释生成等方式，提升系统透明度和用户信任。
- �� 实验验证：在真实平台上进行线上A/B测试，评估内容相关性、用户满意度和系统稳定性。
Experiments
在真实工业平台流量中，比较SYF与传统推荐系统的性能。指标包括内容相关性、用户停留时间、满意度。偏好评分模块达98.85%准确率，内容匹配效果显著提升。线上测试显示，SYF内容相关性提升20%以上，用户停留时间增加15%。通过消融实验验证感知流、偏好优化和判官模型的贡献，确保系统鲁棒性和实用性。
Results
SYF偏好评分达98.85%，优于传统的95%以上。内容相关性提升20%，用户停留时间增加15%。多模态融合增强偏好表达，支持多轮交互。偏好调节机制有效提升用户体验，系统表现出良好的稳定性和扩展性，为工业应用奠定基础。
Applications
适用于内容平台、社交媒体和电商场景，用户可通过自然语言或UI实时调节偏好。系统支持多模态输入，提升用户交互体验。企业可借助SYF实现个性化内容推送、增强用户粘性和满意度。未来还可扩展到智能助手、内容审核等场景，推动行业智能升级。
Limitations & Outlook
高依赖大模型带来计算成本，实时性在极端场景下受限。偏好档案受用户行为稀疏或噪声影响，影响调节效果。多模态识别受语音识别和文本理解限制，存在误判风险。未来需优化模型压缩、鲁棒性和多模态融合算法，提升系统效率和准确性。
Plain Language Accessible to non-experts
想象你在一个厨房里做饭，厨师（系统）需要准备各种食材（内容）来满足不同客人的需求（用户偏好）。传统厨师只根据过去的订单（隐性行为）猜测客人喜欢什么，但不能直接问客人。现在，有了智能助手（大模型），你可以直接和客人聊天，听他们说喜欢什么（自然语言交互），助手还能根据客人的话调整菜单（内容调节）。厨房里的厨师会记住客人的偏好（偏好档案），每次调整菜谱（内容排序）都更贴心。这个系统还能不断学习，随着客人反馈变得更懂他们（自我演化），让每次用餐都更满意。这就像一个会听会记、会变聪明的智能厨房，真正实现了人机合作的美好场景。
ELI14 Explained like you're 14
想象你在学校的食堂点餐，你可以直接告诉厨师你想吃什么，比如说“我喜欢辣的菜”或者“不要太咸”。以前，厨师只知道你平时喜欢吃什么（比如你平时吃得多），但不能听你说的具体偏好。现在，有了一个聪明的机器人助手，它能听你说话，理解你喜欢什么，还能记住你的偏好，每次点餐都帮你挑最喜欢的菜。这个机器人还能根据你的反馈，学会更好地满足你的口味。这样，你就不用一直盯着菜单，也不用担心点错菜，因为机器人会帮你调节，确保你每次都吃得开心。这就像一个会听会记、会变聪明的智能点餐助手，让你的用餐变得更方便、更贴心。
Abstract
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: (i) a Perception Flow that captures fine-grained user intent from text prompts, voice commands, and UI interactions; (ii) a Serving Flow that performs real-time agentic re-ranking and pruning of candidate items, grounded in a persistent Semantic Profile encoding evolving user preferences; and (iii) a Self-Evolution Flow that aligns system behavior with human judgments via Direct Preference Optimization (DPO) and an LLM-as-a-Judge ensemble. 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.