💬 NLP

RGAlign-Rec: Ranking-Guided Alignment for Latent Query Reasoning in Recommendation Systems

The paper proposes RGAlign-Rec, a closed-loop framework that integrates an LLM-based semantic reasoner with a ranking model through a novel Ranking-Guided Alignment training paradigm to bridge the semantic gap and align reasoning with ranking objectives, thereby significantly improving proactive intent prediction and recommendation performance in industrial e-commerce systems.

Original authors: Junhua Liu, Yang Jihao, Cheng Chang, Kunrong LI, Bin Fu, Kwan Hui Lim

Published 2026-02-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Junhua Liu, Yang Jihao, Cheng Chang, Kunrong LI, Bin Fu, Kwan Hui Lim

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 you walk into a massive, chaotic department store (like a huge e-commerce site). You look a little confused, maybe you're holding a package that hasn't arrived, or you're worried about a payment that failed. You haven't said a word yet, but a helpful shop assistant (the chatbot) immediately walks up to you and says, "I see you're waiting for a delivery. Would you like to check the status or expedite the shipping?"

That is the goal of this paper: Proactive Intent Prediction. The system wants to guess what you need before you even ask.

However, the authors found that existing systems were like a shop assistant who speaks a different language than the customers, or one who is great at chatting but terrible at selling. Here is how they fixed it using RGAlign-Rec.

The Two Big Problems

1. The "Translation Gap" (Semantic Gap)
Imagine the shop assistant has a notebook of customer problems written in complex code (User IDs, Order Numbers like "8492"). But the customer speaks natural language ("My package is late!").

The Issue: Old systems tried to match the code directly to the code. They missed the meaning. They couldn't understand that "Order 8492" + "7 days old" = "I'm angry about a late delivery."
The Fix: They introduced an LLM Reasoner (a smart AI translator). This AI reads the code, understands the story behind it, and translates it into a natural question: "The status isn't updating; can we speed this up?"

2. The "Wrong Goal" (Objective Misalignment)
Imagine the shop assistant is trained to be a great poet. They write beautiful, polite sentences. But the store manager only cares about one thing: Did the customer click the link?

The Issue: The AI was writing beautiful poetry (good for humans) but not the specific "sales pitch" that gets the customer to click (good for the business). The AI was aligned with human feelings, not ranking metrics.
The Fix: They needed to retrain the AI to care about the "click," not just the poetry.

The Solution: RGAlign-Rec (The "Closed-Loop" Team)

The authors created a three-step training camp to fix these issues. Think of it as a Coach, a Player, and a Referee working together.

Step 1: The Coach Sets Up the Game (QE-Rec)

First, they built a Query-Enhanced Ranking Model.

Analogy: Imagine a three-legged stool.
- Leg 1 (User): Who are you? (Your history, profile).
- Leg 2 (Intent): What are the possible problems? (The list of intents).
- Leg 3 (Query): What is the AI thinking you need? (The new "Translator" AI).
They trained this stool to be very stable. This model acts as the Referee. It knows exactly which "guess" leads to a click.

Step 2: The Training Camp (Ranking-Guided Alignment)

Now, they take the "Translator" AI (the LLM) and put it in a boot camp.

The Drill: The AI generates 4 different guesses for what the user needs.
The Referee's Call: The "Coach" (the Ranking Model) scores these guesses. "Guess A got a click! Guess B was ignored."
The Lesson: The AI learns: "Oh, I shouldn't just write a polite sentence. I need to write the specific sentence that the Coach knows will get a click."
The Magic Trick (Best-of-N): They don't just pick one. They generate many options, pick the winners, and teach the AI to mimic the winners. They also use a technique called Contrastive Learning, which is like showing the AI two pictures: one of a "good guess" and one of a "bad guess," and forcing the AI to understand the difference in the deep math behind them.

Step 3: The Closed Loop (Calibration)

Once the AI gets better at guessing, they feed those better guesses back into the Coach (the Ranking Model).

Analogy: The Coach gets better at judging because the Player is now giving better inputs. The Player gets better because the Coach is judging more accurately. They keep doing this until they are perfectly in sync.

The Results: Why It Matters

When they tested this in the real world (on Shopee, a massive shopping app):

Better Guessing: The system got significantly better at predicting what users wanted. The "Error Rate" (guessing wrong) dropped by 3.5%. That sounds small, but in a system with millions of users, that's thousands of happy customers.
More Clicks: The Click-Through Rate (CTR) went up. People were clicking on the suggestions more often.
Real-World Speed: It works fast enough to be used in a live chatbot without making the user wait.

The Big Takeaway

This paper is about teaching an AI to stop being a "general conversationalist" and start being a "specialized sales assistant."

Old Way: "Here is a list of things you might like based on your ID number." (Boring, often wrong).
New Way (RGAlign-Rec): "I see you're waiting for a package. Here is the fastest way to track it." (Smart, helpful, and gets the click).

They did this by creating a feedback loop where the AI's "brain" (the LLM) and the system's "goal" (the Ranking Model) constantly talk to each other until they are perfectly aligned.

1. Problem Statement

The paper addresses the challenge of Proactive Intent Prediction in e-commerce chatbots. Unlike traditional recommendation systems that rely on explicit user queries or long-term interaction history, proactive systems must anticipate user needs (e.g., delivery delays, payment failures) upon entering a chatbot interface without any explicit input ("Zero-Query" setting).

The authors identify two fundamental challenges in existing industrial systems:

Semantic Gap & Modality Mismatch: Traditional Deep Learning Recommendation Models (DLRMs) rely on discrete identifiers (User IDs, Intent IDs). However, chatbot intents are text-rich and semantically fine-grained. Bridging the gap between sparse behavioral features and continuous semantic representations in a zero-query context is difficult.
Objective Misalignment: Large Language Models (LLMs) are typically aligned for human preferences (fluency, helpfulness) via techniques like SFT or DPO. These objectives do not directly correlate with downstream ranking metrics (e.g., Click-Through Rate or CTR). Consequently, LLM-generated semantic representations may be linguistically sound but sub-optimal for ranking tasks.

2. Methodology: RGAlign-Rec Framework

The authors propose RGAlign-Rec, a closed-loop framework that integrates an LLM-based semantic reasoner with a Query-Enhanced (QE) ranking model. The framework operates in three iterative stages:

A. Core Architecture Components

LLM-Based User Query Reasoner ( $\pi$ ):
- Feature Verbalization: Discrete user features (profile, order status, behavior sequences) are mapped to natural language strings to create a "Chain-of-Thought" context prompt.
- Latent Query Reasoning: An LLM (Qwen3-4B) processes this context to synthesize a latent user query.
- Representation Extraction: Instead of mean pooling or external sentence encoders, the system uses Last Token Pooling (LTP) on the LLM's output. The hidden state of the final token serves as the high-fidelity semantic embedding ( $e_{LTP}$ ) representing the user's latent intent.
Query-Enhanced Ranking Model (QE-Rec):
- A three-tower architecture: User Tower (hierarchical encoding of static/sequential features), Intent Tower (hybrid ID + semantic embeddings), and Query Tower (projects the LLM's $e_{LTP}$ into the ranking space).
- The scoring function combines collaborative signals (User-Intent) and semantic matching (Query-Intent).

B. The Ranking-Guided Alignment (RGA) Process

To align the LLM's reasoning with ranking objectives, the authors introduce a multi-stage training paradigm:

Stage 1: Initial QE-Rec Training:
- A frozen LLM reasoner generates embeddings.
- The QE-Rec model is trained to maximize ranking metrics. This model then serves as a Reward Model (RM).
Stage 2: Ranking-Guided Alignment (RGA):
- Multi-LLM Distillation: Queries are generated by multiple teacher models (GPT-5, Gemini, CompassMax) and the student model (Qwen3-4B).
- Best-of-N Sampling: The QE-Rec Reward Model scores all candidate queries. The top-performing queries (based on NDCG) are selected as positive samples ( $Q_w$ ), and others serve as negatives ( $Q_l$ ).
- RG-SFT (Supervised Fine-Tuning): The LLM is fine-tuned to generate the "best" queries identified by the reward model.
- RG-CL (Contrastive Learning): To address the limitations of token-level optimization (like in DPO), the authors employ InfoNCE-based contrastive learning. This aligns the LLM's internal hidden states (last token) directly with the semantic space of the optimal queries, ensuring geometric consistency with ranking preferences.
- Note: The paper finds that representation-level alignment (CL) is more robust and scalable than preference-based alignment (DPO) in this context.
Stage 3: Closed-Loop Calibration:
- The refined LLM is used to retrain the QE-Rec model. This creates a feedback loop where the ranker and reasoner mutually improve, synchronizing semantic reasoning with ranking objectives.

3. Key Contributions

RGAlign-Rec Architecture: A novel integration of an LLM semantic reasoner into a query-enhanced three-tower recommendation framework, specifically designed to bridge the semantic gap in Zero-Query intent prediction.
Closed-Loop Ranking-Guided Alignment (RGA): A training paradigm that uses the downstream ranking model as a Reward Model. It introduces Ranking-Guided Contrastive Learning (RG-CL) to align latent semantic representations directly with ranking utilities, overcoming the misalignment between human preference and ranking metrics.
Industrial Validation: The framework is validated on a massive-scale industrial dataset from Shopee, demonstrating that ranking-aware alignment significantly improves both prediction accuracy and service quality in production.

4. Experimental Results

The system was evaluated on a dataset of ~7.53M samples from Shopee's chatbot (Southeast Asia and Latin America).

Offline Performance:

GAUC: Achieved a +0.12% absolute gain over the production baseline (a 3.52% relative error reduction).
Recall@3: Improved by +0.56%.
Comparison: The full RGAlign-Rec (SFT + CL + Calibration) outperformed variants using only SFT, DPO, or raw LLM integration. Representation-level alignment (CL) proved more robust than preference-based (DPO) methods.

Online A/B Testing:

CTR (Click-Through Rate): The QE-Rec model alone yielded a +0.98% lift. The subsequent Ranking-Guided Alignment stage added a further +0.13% gain.
Intent Hit Rate (IHR): Improved by +1.43% (QE-Rec) and an additional +0.30% (RGAlign-Rec).
CSAT (Customer Satisfaction): Showed a marginal decline (-0.21%), suggesting that while intent identification improved, the quality of the response content to those intents remains a separate challenge.

Ablation Studies:

Last Token Pooling (LTP): Outperformed traditional BGE sentence embeddings, proving that leveraging the LLM's internal hidden states is superior for end-to-end optimization.
Sampling Strategy: A "Non-Qwen3 Filtering" strategy (excluding the student model's own generated queries during training) prevented self-fitting bias and yielded the best final results.

5. Significance

This paper provides a critical solution to the "semantic gap" in industrial recommendation systems. By moving beyond treating LLMs as mere feature extractors and instead creating a closed-loop system where ranking signals guide semantic reasoning, the authors demonstrate a practical path to deploying generative AI in high-stakes, real-time recommendation scenarios. The work highlights that representation-level alignment is often more effective than token-level preference optimization for tasks where the ultimate goal is a specific utility metric (like CTR) rather than human conversational quality.