🤖 AI

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

This paper introduces Shape Your Feed (SYF), an LLM-based agentic system that replaces passive ranking with real-time, conversational co-curation through a three-tier architecture, demonstrating significant improvements in feed relevance and user sentiment via both offline alignment accuracy and large-scale online A/B testing.

Original 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

Published 2026-08-10

📖 7 min read🧠 Deep dive

CC BY 4.0

Original 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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your phone is a magical, endless buffet of content—videos, posts, and stories—served up by a very busy, invisible chef. For years, this chef has been working in the dark, guessing what you like based only on what you've eaten before. If you lingered over a picture of a cat, the chef assumes you want a thousand more cats. If you scrolled past a recipe, they assume you hate cooking. This is called "passive ranking," and while it's good at guessing, it often misses the mark because it can't hear you speak. It doesn't know if you actually wanted to see that cat, or if you just stopped to pet your own.

Recently, a new kind of technology called Large Language Models (LLMs) has arrived. Think of these as super-smart, chatty assistants who can understand not just what you click, but what you say. They can listen to your voice, read your text messages, and understand the nuance of your mood. The big question for scientists building recommendation systems is: Can we teach this chatty assistant to take over the kitchen? Can we let you tell the chef, "Actually, I'm tired of cats today; bring me some space facts," and have the chef instantly change the menu? This is the challenge of "conversational recommendation"—moving from a system that guesses your taste to one that listens to your orders.

The "Shape Your Feed" System: A Chef Who Listens

In their paper, a team of researchers from Meta introduces a new system called Shape Your Feed (SYF). They built a digital kitchen where the chef doesn't just guess; they listen, learn, and adjust the menu in real-time based on your direct instructions. Instead of a passive system that waits for you to click "dislike," SYF is an "agentic" system, which is a fancy way of saying it acts like an intelligent assistant that can take action on your behalf.

The researchers found that by combining a chatty AI with a traditional recommendation engine, they could create a feed that feels much more like your feed. They showed that when users could talk to the system or use smart buttons to say exactly what they wanted, the system got much better at showing them the right stuff.

How the Magic Happens: The Three-Flow Kitchen

The SYF system works like a three-part team in a kitchen, each with a specific job:

1. The Perception Flow (The Ears and Brain)
This is where the system listens to you. You might type "I want more cooking videos but no spicy food," use your voice to say "Show me less politics," or tap a smart button that says "I'm bored of this topic." The Perception Flow takes these messy, human instructions and turns them into a clear, organized "Semantic Profile." Think of this as the chef writing a neat shopping list based on your rambling order. It remembers your history, so if you said "no politics" yesterday and "more tech" today, it updates the list accordingly. It even goes out and quietly grabs some new "ingredients" (videos or posts) that match your new list while you're still talking.

2. The Serving Flow (The Plating and Serving)
Once the list is ready, the Serving Flow takes over. It looks at the huge pile of content the system usually would show you (the "candidate pool") and starts sorting it. It uses the chef's new shopping list to:

Add: Pull in fresh items that match your new interests.
Prune: Toss out anything that violates your rules (like those political posts you banned).
Re-rank: Reorder the plate so the things you asked for are at the very top.
Crucially, this happens super fast. The system doesn't just guess; it uses a smart scoring method to decide exactly how much your new request should change the order. It blends your explicit orders with the system's usual knowledge of what you like, ensuring you don't lose all your favorite content just because you asked for a little change.

3. The Self-Evolution Flow (The Taste-Tester)
This is the part that makes the system get smarter over time. After the system serves you a meal, it watches what you do. Did you eat the new cooking video? Did you throw away the tech post? It also uses a team of "judge" AIs to check if the system's decisions were good. If the system made a mistake, it learns from it. The researchers used a technique called Direct Preference Optimization (DPO), which is like the chef practicing with a strict food critic until they get the taste perfect. This loop ensures the system doesn't just follow orders once, but gets better at following them every time.

What They Found: The Proof is in the Pudding

The researchers didn't just build this; they tested it to see if it actually worked.

The Offline Tests (The Practice Run)
First, they tested the system's "alignment scoring" module—the part that decides if a post matches your request. They compared their new system against older methods that just guessed based on a few examples.

The old "few-shot" method got about 83.84% of the decisions right.
Their new system, after training with the "judge" AIs and then refining with real user data (SFT + DPO), got 98.85% of the decisions right.
It also did this incredibly fast, taking only 323 milliseconds to score a list of items, which is fast enough to keep your phone from lagging.

The Online Tests (The Real Meal)
Then, they rolled out SYF to real users in the US and Canada for several months. They split the users into two groups: one group got the old system, and the other got the new "Shape Your Feed" system.

Users loved the new controls: When given a choice between old, static buttons and the new "Context-Aware Feedback Pills" (smart buttons that change based on what you're looking at), 76.02% of users chose the new, smarter buttons.
Less stuff people hated: The new system successfully reduced the number of posts users dismissed or disliked. The "Post Dismiss" rate dropped by 1.70%, and the "Post Dislike" rate fell by 2.74%. This means the system was better at filtering out the stuff users didn't want.
More discovery: Users started exploring new interests more often, with a 0.16% increase in "Interest Consumption." This suggests the system wasn't just hiding things; it was helping users find new things they actually liked.

What It Means

The paper shows that we can move away from systems that just guess what we want based on our past clicks. By adding a layer of AI that can understand our words and act on them, we can create feeds that feel more personal and less frustrating. The researchers found that this approach works well in the real world, making users happier and reducing the amount of content they have to skip over.

However, the authors are careful to note that this system relies on you speaking up. If you never tell the system what you want, it defaults to the old, passive way of guessing. They suggest that future versions might need to be even more proactive, perhaps asking you questions before you even realize you're bored, to help those who don't like to give feedback. But for now, Shape Your Feed proves that a recommendation system that listens is a system that works better.

Technical Summary: Shape Your Feed (SYF)

Problem Statement

Industrial recommendation systems currently rely on a passive ranking paradigm, inferring user preferences exclusively from implicit behavioral signals (e.g., clicks, dwell time). This approach creates a persistent discrepancy between a user's explicit interests and the content delivered by algorithms. Users lack the ability to express nuanced preferences, explain why content surfaces, or steer their feed in real time. Furthermore, existing Conversational Recommender Systems (CRS) often fail in high-throughput feed environments because they assume static catalogs, struggle with multi-turn consistency, and lack mechanisms to tightly bind generative intent to a production inventory without hallucination.

Methodology: The SYF Framework

The authors propose Shape Your Feed (SYF), an LLM-based agentic system designed to enable real-time, multimodal co-curation of content. SYF operates through a three-tier architecture that unifies natural language understanding with production-grade ranking pipelines:

1. Perception Flow

This module captures fine-grained user intent from unstructured multi-modal inputs (text prompts, voice commands, and UI interactions like "Context-Aware Feedback Pills").

Mechanism: An orchestrator parses user queries and routes them to a Recommendation Toolchain.
Key Components:
- Intent Detection: Disambiguates underlying behavioral motivations.
- Preference Synthesis: Updates a persistent Semantic Profile ( $S_t$ ) by performing de-duplication, strengthening, conflict resolution, and semantic expansion based on the user's history and current input.
- Candidate Sourcing: Asynchronously retrieves relevant inventory items based on new positive interests, decoupling retrieval latency from the real-time serving path.
Output: A structured Semantic Profile and an updated Intent Trajectory ( $T_t$ ).

2. Serving Flow

This flow executes real-time agentic re-ranking and pruning, grounding LLM reasoning in the platform's actual inventory.

Context Generation: Translates numerical ranking signals and the Semantic Profile into a natural language prompt ( $P_{align}$ ) interpretable by the LLM.
Agentic Refinement: The system performs a four-step refinement on the candidate pool:
1. Augmentation: Merges production candidates with items retrieved from the Candidate Cache.
2. Alignment Scoring: Uses an LLM to score candidates list-wise against the user's profile, assigning positive scores for matches and negative scores for conflicts.
3. Pruning: Hard-filters out candidates scoring below a threshold $\tau$ , instantly enforcing negative feedback.
4. Re-ranking: Blends the semantic alignment score with the base production engagement score ( $\hat{y}_{u,v}$ ) using a fusion weight $\alpha$ .
Justification: Generates natural language explanations for top items to enhance transparency.

3. Self-Evolution Flow

This module establishes a closed-loop optimization system to align agent behavior with human judgment.

Dual-Feedback Policy Alignment:
- Supervised Fine-Tuning (SFT): Trains the alignment scoring model using labels generated by an LLM-as-a-Judge ensemble (e.g., open-source models like Llama 4 Maverick or Qwen3-VL) to handle sparse explicit feedback.
- Direct Preference Optimization (DPO): Refines the model using real user behavioral feedback (e.g., "Show Less" clicks) to iteratively harden preference boundaries without separate reward models.
Agentic-to-Production Propagation: Distills high-order semantic insights from the agent into the base production ranker ( $f_\theta$ ) to improve retrieval and ranking features over time.

Key Contributions

System Architecture: A scalable, agentic framework that combines LLM-based intent understanding with structured, persistent memory and production-compatible ranking flows, enabling stable multi-turn behavior and catalog-grounded responses.
Multi-modal Preference Supervision: A unified approach to integrating text, voice, and UI controls into a consistent preference state, improving personalization and user agency.
Empirical Validation: Extensive offline and online evaluations demonstrating that SYF improves feed relevance and reduces negative feedback, offering a practical path for interactive recommendation in industrial settings.

Results

Offline Evaluation

Alignment Accuracy: The alignment scoring module achieved 98.85% accuracy (SFT + DPO), significantly outperforming few-shot baselines (83.84%) and conventional ranking signals ($P(CSL)$, $P(CSM)$), which showed negligible discrimination between matched and mismatched content.
Efficiency: The system maintains low latency (323ms for scoring), fitting within production constraints.
Model Selection: Llama 3 8B was selected as the optimal balance between reasoning capability and inference latency, outperforming smaller models (1B) and avoiding the prohibitive costs of larger models (70B).

Online A/B Testing

Conducted on live production traffic (US/Canada, 18+), the SYF system demonstrated:

Enhanced User Control: Users selected Context-Aware Feedback Pills 76.02% of the time, compared to 23.98% for static pills, indicating superior semantic alignment.
Reduction in Negative Feedback:
- Post Dismiss Rate decreased by 1.70%.
- Post Dislike Rate decreased by 2.74%.
Improved Interest Exploration: A statistically significant 0.16% uplift in new interest consumption, indicating the system successfully guides users toward novel domains.
Ablation Insights: The feedback interface (Pills) drove interaction volume, while the underlying agentic serving flow was required to convert those signals into actual feed quality improvements.

Significance and Claims

The paper claims that SYF represents a shift from passive personalization to interactive, user-algorithm co-curation. By bridging the gap between explicit user intent and implicit ranking signals, SYF offers a principled path toward more transparent and controllable feed ranking in large-scale social platforms.

The authors acknowledge a current limitation: the system relies on explicit user input to trigger the active personalization loop. In high-throughput environments where many users are passive, the system defaults to baseline ranking for those who do not interact with steering controls. Future work is proposed to address this sparsity through proactive preference elicitation and transfer learning to propagate inferred profiles to low-signal users.

Ultimately, the work validates that integrating agentic, LLM-driven reasoning into industrial recommendation stacks is technically feasible and beneficial, moving the field toward systems where users can actively steer their content consumption in real time.