Knowledge, collaborative vectors, and semantic codes: how does the same content enter different recommender interfaces?¶
简体中文 · English
Updated 2026-10-27 · 60 topic entries / 4 representative comparisons
KAR generates external knowledge for an existing recommender. NoteLLM learns a searchable compressed state from co-reading. Dual-purpose SID uses compact codes for both identity and content reconstruction. FineSID transfers fine-grained tokens and a complete interaction backbone across stages. Their alignment signals, compressed objects, and inherited states differ even when they all use LLMs or SIDs.
Mechanisms · Trade-offs · Cross-paper questions · Updates
How to read this card
The table compares representative methods. It is not a shared-benchmark leaderboard. Teaching examples are not measured results.
Mechanism comparison¶
Axes: Knowledge from generated common sense, content, or interaction relations · Transfer objects: vectors, compact codes, or full predictor state · Joint constraints from content reconstruction and personalization · Offline caching, cold start, and full-reproduction limits
| Paper and source | Mechanism | Benefits and trade-offs |
|---|---|---|
| Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 本地复现 |
Generates user preference reasoning and item factual knowledge separately, compresses them with a hybrid-expert adapter, then sends them into a recommender alongside original ID representations. | Precomputing knowledge avoids online autoregressive generation, but descriptions are not actual preference labels. Caches need refreshes and facts need verification. The LLM does not do the entire online recommendation directly. |
| NoteLLM: A Retrievable Large Language Model for Note Recommendation · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 本地复现 |
A compression prompt puts content into a special-token state. GCL learns the retrieval space from co-reading; CSFT generates hashtags or categories from the same state. | Contrastive collaborative and generative supervision shape the representation together. Exposure bias in co-reading still matters. Content similarity and behavioral complementarity are not necessarily the same, and retrieval scores do not establish generation capability. |
| Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in Recommendation Systems · Original source Sections in the Chinese source: 背景与主要改动, 核心公式, 复现边界 |
Hierarchical SIDs encode collaborative identity, while SiDec reconstructs content embeddings from their discrete layers for retrieval, ranking, and low-I/O reconstruction. | Compact codes reduce content transfer but introduce quantization collisions and bottlenecks. Measure reconstruction MSE and recommendation separately. Local genre labels replace multimodal content; production codebooks and online reconstruction are not reproduced. |
| Progressive Learning from Content to Relevance to Personalization for CTR Prediction · Original source Sections in the Chinese source: 背景与主要改动, 三阶段结构, 成本判定：不执行复现 |
Retains multiscale visual and field-text tokens. Trains a full query–item relevance predictor, then keeps shared dense, embedding, and normalization state while switching to a CTR head. | Transfer is richer than one embedding but requires compatible backbone and state management across stages. It is not merely an SID lookup. The repository records the method and online evidence; full training was too expensive to implement. |
Trade-off analysis¶
Ask what actually gets transferred¶
KAR transfers features produced from generated knowledge via an adapter. NoteLLM transfers a collaboratively trained compressed vector. Dual-purpose SID transfers a discrete identity that reconstructs content. FineSID transfers shared predictor parameters and normalization state. Adding a random codebook to an existing CTR model reproduces neither KAR's knowledge source nor FineSID's stage inheritance. A token interface does not make them equivalent.
Sources: Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models · NoteLLM: A Retrievable Large Language Model for Note Recommendation · Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in Recommendation Systems · Progressive Learning from Content to Relevance to Personalization for CTR Prediction
Measure content fidelity and recommendation relevance separately¶
Dual-purpose SID reconstruction checks how much of an embedding the code preserves. NoteLLM's GCL uses co-reading to decide which content should be close, imposing a different geometry. FineSID trains one backbone first with soft relevance targets and then for personalized CTR. Lower reconstruction error need not improve click ranking, and more plausible LLM output need not match user preferences.
Sources: Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in Recommendation Systems · NoteLLM: A Retrievable Large Language Model for Note Recommendation · Progressive Learning from Content to Relevance to Personalization for CTR Prediction
Cross-paper questions¶
Try answering before expanding the explanation. Each question relates at least two papers.
Dual-purpose SID and FineSID both use discrete tokens. Why can't a single SID lookup experiment establish both results?¶
Show answer and example
Answer: Dual-purpose SID requires codes to support collaborative identity and content reconstruction. FineSID retains fine-grained multiscale positions and transfers the full shared relevance predictor to CTR. A lookup covers only part of the interface. Without reconstruction or stage inheritance respectively, it misses the defining mechanism.
Illustrative example: Using three codes to roughly reconstruct a shoe image is compression. Learning query–shoe relevance and then inheriting the entire interaction network to predict a user's click is a different training pipeline.
Sources: Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in Recommendation Systems · Progressive Learning from Content to Relevance to Personalization for CTR Prediction
Why can't KAR's generated preference descriptions replace NoteLLM's co-reading supervision?¶
Show answer and example
Answer: KAR offers candidate knowledge about common sense and preferences; the recommendation objective determines its usefulness. NoteLLM's co-reading pairs come from behavior and directly constrain retrieval geometry. Generated knowledge may fill semantic gaps or invent false explanations. Behavioral pairs are observed but affected by exposure and popularity. They complement one another and are not interchangeable ground truths.
Illustrative example: An LLM's suggestion that campers might like hiking boots is a knowledge hypothesis. Users actually reading about a shoe and waterproof socks together is a platform behavioral relation. The evidence differs.
Sources: Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models · NoteLLM: A Retrievable Large Language Model for Note Recommendation
Why can recommendation NDCG drop even when SID reconstruction is more accurate and text generation is more fluent?¶
Show answer and example
Answer: Reconstruction preserves content and CSFT preserves the generation interface, but ranking also needs collaborative distinctions among user–item pairs. A dual-purpose SID codebook may favor content at the expense of preferences; NoteLLM uses GCL for behavior geometry; FineSID continues personalized CTR training. Evaluate reconstruction, collisions, relevance, and isolated test ranking separately.
Illustrative example: Two visually similar coats may be reconstructed accurately, but one is for cycling and the other for commuting. Visual fidelity does not guarantee a correct click order if codes lose the differences between audiences.
Sources: Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in Recommendation Systems · NoteLLM: A Retrievable Large Language Model for Note Recommendation · Progressive Learning from Content to Relevance to Personalization for CTR Prediction
Update log¶
Updates record new approaches, changed trade-offs, and follow-up evidence.
2026-10-27 · Initial version: compared KAR, NoteLLM, Dual-purpose SID, and FineSID. Added analysis and cross-paper questions on knowledge sources and transfer of vectors, compact codes, or full predictor state.
Covers 60 topic entries; see the full list below.
All 60 topic entries
The full catalog follows. The main comparison covers the representative papers listed above.
- Beyond Static Collision Handling: Adaptive Semantic ID Learning for Multimodal Recommendation at Industrial Scale
- AgentWebRec: Compact Evidence Fusion over the Agent Web for Personalized Recommendation
- From Head to Tail: Asymmetric Knowledge Transfer in Long-tail Recommendation with Generative Semantic IDs
- Atomic Intent Reasoning: Bringing LLM Semantics to Industrial Cross-Domain Recommendations
- Breaking the Length Barrier: LLM-Enhanced CTR Prediction in Long Textual User Behaviors
- CAMIE: Co-Engagement-Aware Multimodal Item Embeddings for Snap Dynamic Product Ads Retrieval
- ConnectionMind: A General Social-Personalized Recommendation System with LLM Reasoning
- Zero-shot Cross-domain Knowledge Distillation: A Case study on YouTube Music
- Douyin Multimodal Embedding Model Technical Report
- DOS: Dual-Flow Orthogonal Semantic IDs for Recommendation in Meituan
- Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in Recommendation Systems
- From a Static Multi-Level Small Semantic Codebook to a Dynamic Single-Level Large Semantic Codebook for Generative Recommendation
- FilterLLM: Text-To-Distribution LLM for Billion-Scale Cold-Start Recommendation
- Progressive Learning from Content to Relevance to Personalization for CTR Prediction
- FLUID: From Ephemeral IDs to Multimodal Semantic Codes for Industrial-Scale Livestreaming Recommendation
- Structuring and Tokenizing Distributed User Interest Context for Generative Recommendation
- GALA: Generative Aligned Learning for Adaptive Multimodal Representation in the Taobao Shangou Recommender System
- GateSID: Adaptive Gating for Semantic-Collaborative Alignment in Cold-Start Recommendation
- Neither Black nor White: Balancing Semantic and Collaborative Signals with Graph-Informed Semantic IDs (GrIS)
- IDProxy: Cold-Start CTR Prediction for Ads and Recommendation at Xiaohongshu with Multimodal LLMs
- Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models
- Knowledge–Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation
- KuaFu: Compressing Long User Behavior into Understanding at Billion Scale
- L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation
- Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application
- A Large Language Model Enhanced Sequential Recommender for Joint Video and Comment Recommendation
- Unlocking Scaling Law in Industrial Recommendation Systems with a Three-step Paradigm based Large User Model
- LWGR: Lagrangian-Constrained Personalized World Knowledge for Generative Recommendation
- M6-Rec: Generative Pretrained Language Models are Open-Ended Recommender Systems
- Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations
- MIM: Multi-modal Content Interest Modeling Paradigm for User Behavior Modeling
- A General Framework for Multimodal LLM-Based Multimedia Understanding in Large-Scale Recommendation Systems
- Mosaic: A Fleet of User Embedding Specialists for Recommendation at Meta
- Balancing Efficiency and Effectiveness: An LLM-Infused Approach for Optimized CTR Prediction
- MuChator: Enabling Active Music Discovery via Conversational Music LLMs in Douyin Music
- Explainable Recommendations at Scale: LLM Rationales for YouTube Music Artist Discovery
- Multimedia Asset Personalization via Multimodal Embeddings at Netflix
- NoteLLM: A Retrievable Large Language Model for Note Recommendation
- Building a User Foundation Model for the Open Web
- Pin-SCALE: Semantic Cascading and Alignment Learning for Engagement-Aware IDs in Cold-Start Recommendations
- PinCLIP: Large-scale Foundational Multimodal Representation at Pinterest
- Fine-Tuned LLM as a Complementary Predictor Improving Ads System
- PIT: A Dynamic Personalized Item Tokenizer for End-to-End Generative Recommendation
- PRECISE: Pre-training Sequential Recommenders with Collaborative and Semantic Information
- Proximity Features: Privacy-Compliant Cold-Start Personalization at Airbnb
- Stop Treating Collisions Equally: Qualification-Aware Semantic ID Learning for Recommendation at Industrial Scale
- RECAP: Feedback-Driven Streaming Semantic User Profiles for Short-Video Recommendation
- RecGPT-Mobile: On-Device Large Language Models for User Intent Understanding in Taobao Feed Recommendation
- RecGPT-V2 Technical Report
- RecoReward: Recommender-Guided Multimodal Description Generation for Recommendation
- RQ-GMM: Residual Quantized Gaussian Mixture Model for Multimodal Semantic Discretization in CTR Prediction
- SARM: LLM-Augmented Semantic Anchor for End-to-End Live-Streaming Ranking
- SaviorRec: Semantic-Behavior Alignment for Cold-Start Recommendation
- Bursting Filter Bubble: Enhancing Serendipity Recommendations with Aligned Large Language Models
- SID-Coord: Coordinating Semantic IDs for ID-based Ranking in Short-Video Search
- Don't Rewrite the SID Map: Topology Preservation for Generative Retrieval in Dynamic E-Commerce Search
- SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control
- SSRLive: Live Streaming Recommendation with Dynamic Semantic ID
- TagLLM: A Fine-Grained Tag Generation Approach for Note Recommendation
- TokenMinds: Pretrained User Tokens and Embeddings for User Understanding in Large Recommender Systems