🤖 AI

Atomic Intent Reasoning: Bringing LLM Semantics to Industrial Cross-Domain Recommendations

This paper presents AIR (Atomic Intent Reasoning), an industrial-grade cross-domain recommendation framework that bridges the semantic gap between content and e-commerce by leveraging offline LLM inference and efficient online retrieval to achieve significant inference acceleration and substantial GMV improvements in Kuaishou's real-world deployment.

Original authors: Zhuohang Jiang, Yuxin Chen, Shijie Wang, Haohao Qu, Zhou Jindong, Wenqi Fan, Li Qing, Dongxu Liang, Jun Wang

Published 2026-06-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhuohang Jiang, Yuxin Chen, Shijie Wang, Haohao Qu, Zhou Jindong, Wenqi Fan, Li Qing, Dongxu Liang, Jun Wang

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 are a very smart, but incredibly slow and expensive, personal shopper named "LLM." This shopper is brilliant at understanding human desires. If you tell them, "I watched a video of a cute cat and then bought cat food," they can instantly deduce, "Ah, this person loves cats and might want a cat tree next."

However, there's a problem: This shopper takes 8 seconds to give you an answer. In the world of online shopping (like on Kuaishou, a massive Chinese app), you need an answer in milliseconds. If the shopper is too slow, the customer has already left the store. Also, hiring this shopper for every single user interaction costs a fortune.

The paper introduces a new system called AIR (Atomic Intent Reasoning) to solve this. Think of AIR as a brilliant architect and librarian who works before the customer arrives, so the actual shopping happens instantly.

Here is how AIR works, broken down into simple steps:

1. The "Pre-Game" Prep (Offline Phase)

Instead of asking the slow "LLM shopper" to think in real-time, AIR does all the heavy thinking offline (when no one is watching).

The Atomic Breakdown: The system takes a user's messy history (watching videos, clicking ads, buying things) and asks the LLM to break it down into tiny, clear "intent atoms."
- Analogy: Instead of saying, "I watched a cat video, then a car video, then bought cat food," the system translates this into specific, labeled cards: [Action: Watch] + [Object: Cat Video] = Intent: Cat Lover.
The Library: These "intent cards" are organized into a massive, highly organized library (an Intent Tree) and stored away. This is like pre-cooking a meal and freezing it, so you don't have to cook from scratch when a customer orders.

2. The "Flash" Service (Online Phase)

When a user actually visits the app, the system doesn't call the slow LLM. Instead, it acts like a super-fast librarian.

The Search: The system looks at what the user is doing right now (e.g., they are looking at a "Badminton" racket).
The Retrieval: It instantly runs to the library and pulls out only the "intent cards" that match Badminton. It ignores the irrelevant stuff (like the cat videos the user watched last month, unless they are relevant).
The Assembly: It quickly stacks these relevant cards together to form a clear picture of what the user wants right now.

3. The "Noise Filter"

User behavior data is often messy and huge (like a library with millions of books, most of which are irrelevant to the current customer).

The Metaphor: Imagine trying to find a specific needle in a haystack. The old way was to look at the whole haystack. AIR uses a magnet (Target-Aware Retrieval) that only pulls out the needles (relevant intents) and ignores the hay (noise).
This allows the system to handle massive amounts of data without getting confused or slow.

4. The Result: Speed and Smarts

By doing the "thinking" beforehand and the "retrieving" instantly, AIR achieves two amazing things:

Speed: It is 400 times faster than calling the LLM directly. It goes from taking 8 seconds to just 20 milliseconds.
Accuracy: Because it uses the LLM's "brain" to understand the meaning behind the actions (not just the numbers), it predicts what users want to buy much better than older systems.

Real-World Proof

The authors tested this on Kuaishou, a giant platform with hundreds of millions of users.

The Test: They ran a real-world experiment (A/B test) where half the users got the old system and half got AIR.
The Win: The AIR group made 3.4% more money (GMV) for the company. In the world of big business, that is a massive victory. It also helped users who didn't have much history (the "cold start" problem) by guessing their interests based on the "intent cards" created by the LLM.

In Summary:
AIR is a clever trick that lets companies use the "super-smart" brain of a Large Language Model without paying the "slow and expensive" price. It pre-digests user data into easy-to-read cards, so when a user clicks, the system can instantly say, "I know exactly what you want," in the blink of an eye.

Technical Summary: Atomic Intent Reasoning (AIR)

1. Problem Statement

The paper addresses the critical challenges of Cross-Domain Recommendation (CDR) in industrial content-to-merchant platforms (e.g., Kuaishou), where the goal is to infer user purchasing intent from content-side behaviors (e.g., watching videos) to drive e-commerce conversions.

Existing approaches face two primary bottlenecks in real-world deployment:

Semantic Gaps: Traditional methods relying on ID-level collaborative filtering (shared embeddings, GNNs) capture co-occurrence correlations but lack the semantic grounding to understand fine-grained intent transitions between heterogeneous domains (e.g., from "watching a cat video" to "buying cat food").
Scalability and Latency of LLMs: While Large Language Models (LLMs) offer superior semantic understanding and reasoning, their direct application in online serving is prohibitive due to:
- Latency: Real-time LLM inference takes seconds, whereas industrial systems require millisecond-level responses.
- Noise and Scale: User behavior sequences are massive, heterogeneous, and noisy. Feeding raw, long histories into models amplifies noise and computational costs.
- Timeliness: Periodic offline updates of user profiles fail to capture rapidly evolving user interests.

2. Methodology: Atomic Intent Reasoning (AIR)

The authors propose AIR, an LLM-driven framework that decouples heavy semantic reasoning from online serving. It employs an offline-to-online pipeline to achieve real-time semantic awareness with millisecond latency. The framework consists of four stages:

Stage 1: Atomic Intent Pair Generation (Offline)

Process: Instead of processing full sequences online, AIR uses LLMs offline to decompose complex user interaction sequences into atomic behavior-intent units.
Mechanism: Each user event $e = (action, object, user\_profile)$ is fed into an LLM prompt combining context, behavioral semantics, and object descriptions. The LLM outputs hierarchical intent paths (e.g., $Cat \rightarrow Cat\ Food$ ).
Output: These atomic intents are cached in an Intent Knowledge Base, organized by action types (click, order, view) to allow differentiated weighting in downstream modeling.

Stage 2: Real-time Unified User Intent Tree Construction (Online)

Process: During online serving, the system retrieves cached atomic intents from the user's recent behavior history.
Structure: These intents are aggregated into a Unified User Intent Tree.
- Nodes: Represent intent concepts at varying levels of abstraction.
- Weights: Each node is assigned a preference weight $w_u$ calculated by combining multi-action counts (clicks, orders, views) with a temporal decay factor to emphasize recent behaviors.
Benefit: This transforms massive, heterogeneous sequences into a structured, semantically expressive representation without invoking the LLM online.

Stage 3: Target-aware Intent Retrieval

Process: To address noise and scale, the system performs retrieval conditioned on the target item (the candidate product).
Mechanism:
1. Semantic Similarity: Computes similarity between the target item embedding and intent node embeddings.
2. Intent Chain Finding: Selects top-ranked intent chains based on target relevance and user preference strength.
3. Sequence Searching: Retrieves compact, high-evidence behavioral instances (e.g., specific clicks or orders) supporting the selected chains, applying denoising (filtering stale or irrelevant signals).
Output: A concise User Intent Chain tailored specifically to the target item.

Stage 4: MHA-based Intent Extraction

Process: The retrieved intent chain is fused with the target item embedding using a Multi-Head Attention (MHA) module.
Function: Different attention heads capture complementary facets of user intent (e.g., coarse-grained category alignment vs. fine-grained preference patterns).
Integration: The resulting representation is integrated with standard user/item features for downstream Click-Through Rate (CTR) prediction and ranking.

3. Key Contributions

AIR Framework: A novel cross-domain recommendation framework that achieves effectiveness comparable to real-time LLM calls while maintaining millisecond-level serving latency.
Atomic Intent Caching: A mechanism that distills user events into atomic intent units offline, yielding a 400× throughput gain over real-time LLM invocation.
Target-Aware Retrieval: A method to extract compact, high-evidence intents from long, noisy histories, combined with MHA for fine-grained preference modeling.
Industrial Validation: Successful deployment and A/B testing in Kuaishou's real-world e-commerce environment.

4. Experimental Results

The paper evaluates AIR on public benchmarks and a large-scale industrial dataset.

Public Benchmarks (Amazon Movie-Book & Food-Kitchen):
- AIR outperforms state-of-the-art baselines (including LLMCDSR, SASRec, and traditional CDR models) in NDCG@10 and HR@10.
- Notable improvements: In the Movie-Book scenario, HR@10 improved by 41.31% over the second-best method (LLMCDSR).
Ablation Studies:
- Removing the Intent Tree or Intent Retrieval components causes significant performance drops (HR@10 drops of 17–27%), confirming their critical, coupled role in handling data sparsity and organizing knowledge.
Latency Analysis:
- Direct LLM inference (Qwen3-4b) takes ~8 seconds.
- AIR reduces this to ~20.1 ms, achieving a 400× throughput gain.
Industrial Online A/B Test (Kuaishou):
- Deployed on 5.08% of platform traffic.
- GMV (Gross Merchandise Value): +3.446% increase.
- GPM (Gross Merchandise per Mille): +3.662% increase.
- Paid Order Count: +1.043% increase.
- OPM (Orders per Mille): +1.254% increase.
- User Activity: Consistent gains across low, mid, high, and ultra-activity user groups, with the highest lift (+8.04%) observed in high-activity users, validating the effectiveness of noise filtering in long sequences.

5. Significance and Claims

The paper claims that AIR successfully bridges the gap between the semantic reasoning capabilities of LLMs and the strict latency/throughput constraints of industrial recommendation systems.

Practical Value: By migrating LLM inference to the offline phase and using efficient retrieval/composition online, AIR validates that LLM-level semantics can be operationalized in real-time industrial scenarios without prohibitive costs.
Robustness: The framework effectively handles the "noisy, massive, and heterogeneous" nature of real-world user behavior data, outperforming methods that rely on overlapping users or shallow co-occurrence patterns.
Business Impact: The significant uplift in core business metrics (GMV, GPM) in a live production environment demonstrates the framework's direct contribution to commercial value, proving its viability for large-scale content-to-commerce loops.