Tagged articles
Re-ranking
31 articles · Page 1 of 1
Himalaya uses LLMs as user preference proxies to automatically optimize recommendation parameters across recall, ranking, and re-ranking stages, leveraging user context and long-term memory to generate personalized parameters via a parameter center, achieving significant metric gains for low-frequency users.
HimalayaLLMPBox
0 likes · 26 min read

This article systematically dissects the re-ranking stage in recommender systems, explaining how hard constraints (filtering, quotas, safety) and soft optimizations (diversity, blending) are orchestrated in a prioritized pipeline to transform ranked candidates into a final display list that balances user experience, platform health, and operational needs.
MMRRe-rankingdiversity
0 likes · 40 min read

When a long document is split into many highly similar chunks, vector‑based top‑k retrieval tends to return multiple pieces from the same source, causing document dominance; applying a per‑document chunk limit together with Maximal Marginal Relevance (MMR) re‑ranking introduces diversity while preserving relevance, offering a low‑cost way to improve RAG answer quality.
DPPMMRRAG
0 likes · 17 min read

This article examines practical ways to improve Retrieval‑Augmented Generation (RAG) retrieval quality—covering vector database choices, data chunking, embedding models, query expansion, and re‑ranking—while weighing performance gains against operational costs through multiple real‑world case studies.
LLMRAGRe-ranking
0 likes · 16 min read

This article examines why Retrieval‑Augmented Generation systems that work in demos often fail in production, detailing common pitfalls—from chunking and vector‑database selection to hybrid retrieval and re‑ranking—and offers concrete strategies, configuration tips, and a decision tree to build reliable enterprise‑grade RAG solutions.
Hybrid RetrievalRAGRe-ranking
0 likes · 12 min read

The article analyzes a recent paper that introduces a generative reasoning re‑ranker for LLM‑driven recommendation, detailing its SFT and RL training pipeline, semantic‑ID embedding, target vs. reject sampling strategies, and experimental gains of 2.4% Recall@5 and 1.3% NDCG@5 over the OneRec‑Think baseline.
Generative ReasoningLLMRe-ranking
0 likes · 9 min read

This article dissects why naive Retrieval‑Augmented Generation (RAG) often yields only 60% accuracy, then presents eleven concrete ingestion, query, and hybrid techniques—complete with code samples, performance trade‑offs, and real‑world case studies—that together can raise RAG accuracy to 94% while outlining practical implementation roadmaps and common pitfalls.
EmbeddingLLMRAG
0 likes · 31 min read

This article examines the evolution of re‑ranking systems from traditional pointwise models to a two‑stage generation‑evaluation framework, compares autoregressive and non‑autoregressive generative approaches, details inference speed optimizations with GPU and model‑server upgrades, and outlines a future end‑to‑end sequence generation architecture enhanced by reinforcement learning and contrastive learning.
AIGenerative ModelsInference Optimization
0 likes · 14 min read

This article walks developers through three essential upgrades for Retrieval‑Augmented Generation systems—hybrid search combining vector and keyword retrieval, query rewriting to clarify conversational inputs, and re‑ranking with a cross‑encoder—providing step‑by‑step code examples using LangChain to dramatically improve answer quality.
AILangChainQuery Rewriting
0 likes · 9 min read

This article recounts a JD engineer’s journey from theory to practice, detailing the development of the PODM‑MI re‑ranking framework, its three‑layer distribution‑based design, the discovery of a novel SID bottleneck, and the resulting multi‑million‑order impact validated at SIGIR 2024.
Distribution ModelingRe-rankingSIGIR
0 likes · 8 min read

This article dives into the practical implementation of Retrieval‑Augmented Generation (RAG), covering document chunking, semantic and reverse HyDE indexing, embedding, hybrid search, and re‑ranking techniques, and provides concrete code examples and optimization tips for building high‑performance AI applications.
EmbeddingRAGRe-ranking
0 likes · 18 min read


JD TechJul 11, 2025 · Artificial Intelligence
JD’s algorithm engineer recounts how his team transformed e‑commerce search by developing the PODM‑MI re‑ranking framework, uncovering a novel “hourglass” bottleneck in generative retrieval, and implementing lightweight solutions that boosted diversity, relevance, and order volume, culminating in a SIGIR publication.
Gaussian modelingRe-rankinge-commerce
0 likes · 8 min read

JD’s algorithm engineer describes how a three‑layer PODM‑MI re‑ranking framework, combining Gaussian preference modeling, mutual‑information optimization, and utility‑matrix fusion, overcame the hourglass bottleneck in generative retrieval, dramatically improving search diversity, user experience, and generating over ten million additional orders.
AIRe-rankinge-commerce
0 likes · 9 min read

This article presents LAST, a novel Learning-at-Serving-Time approach that enables real‑time online learning for re‑ranking in industrial recommendation pipelines, eliminating feedback latency, detailing its architecture, challenges, experimental validation, and practical advantages over traditional online learning methods.
LAST algorithmRe-rankingonline serving
0 likes · 12 min read


JavaEdgeOct 2, 2024 · Artificial Intelligence
This article presents a contextual retrieval technique that combines contextual embeddings and contextual BM25 to reduce RAG miss rates by up to 67%, explains the underlying methods, implementation steps, cost considerations, experimental results, and practical deployment guidance.
AIBM25Contextual Retrieval
0 likes · 17 min read


NewBeeNLPSep 9, 2024 · Artificial Intelligence
This article introduces LAST, a novel online learning approach that updates recommendation models instantly at serving time, addressing real‑time learning challenges, re‑ranking complexities, and demonstrating superior offline and online performance in industrial e‑commerce scenarios.
AILASTRe-ranking
0 likes · 12 min read

This paper, accepted at SIGIR 2024, introduces PODM‑MI, a preference‑oriented diversity re‑ranking model for e‑commerce search that jointly optimizes accuracy and diversity by modeling user intent with multivariate Gaussian distributions and maximizing mutual information between user preferences and candidate items.
E-Commerce SearchRe-rankingUser Preference Modeling
0 likes · 11 min read


JD TechAug 26, 2024 · Artificial Intelligence
This article presents the SIGIR 2024 accepted PODM‑MI model, which uses variational inference and mutual‑information maximization to jointly optimize relevance and diversity in JD e‑commerce search re‑ranking, demonstrating significant gains in both user conversion and result diversity through extensive online experiments.
E-Commerce SearchRe-rankingdiversity
0 likes · 11 min read

The paper proposes PODM‑MI, a mutual‑information‑driven, preference‑oriented diversity model that jointly optimizes accuracy and diversity in e‑commerce search re‑ranking by modeling user preferences with multivariate Gaussian distributions and adapting rankings via a learnable utility matrix, showing significant gains in JD's main search experiments.
AIE-Commerce SearchRe-ranking
0 likes · 12 min read

This paper introduces PODM‑MI, a preference‑oriented diversity model that jointly optimizes relevance and diversity in e‑commerce search re‑ranking by leveraging variational inference and mutual‑information, demonstrating significant gains in both user conversion and result variety on JD.com.
E-Commerce SearchRe-rankingdiversity
0 likes · 10 min read

This article introduces PODM‑MI, a preference‑oriented diversity model that uses mutual information and variational Gaussian representations to jointly optimize accuracy and diversity in e‑commerce search re‑ranking, and reports significant online A/B test improvements on JD.com.
Re-rankingdiversitye-commerce
0 likes · 10 min read

The paper introduces a real‑time controllable, multi‑objective re‑ranking framework for Taobao’s feed recommendation that combines actor‑critic reinforcement learning with hypernetworks to instantly adjust objective weights, handling diverse media and cold‑start constraints while delivering higher click‑through, diversity, and cold‑start ratios with only 20‑25 ms latency.
AlibabaHypernetworksMulti-Objective Optimization
0 likes · 34 min read

This article presents a comprehensive study of a controllable multi‑objective re‑ranking model for Taobao's information‑flow recommendation, detailing the challenges of complex feed scenarios, three modeling paradigms (V1‑V3), an actor‑critic reinforcement learning framework with hypernet‑generated weights, and extensive online evaluation results.
HypernetworksMulti-Objective OptimizationRe-ranking
0 likes · 31 min read

Meituan/Dianping’s edge‑AI re‑ranking system moves large‑scale deep‑learning models onto users’ devices, using dense networks and cloud‑served embeddings, advanced feedback‑sequence and multi‑view attention models, and aggressive compression to deliver real‑time, privacy‑preserving search personalization that boosts click‑through rates by up to 0.43 %.
Re-rankingedge AImobile search
0 likes · 25 min read

The article surveys the re‑ranking stage of modern recommendation pipelines, detailing its architecture after recall and precise ranking, and examining how shuffling and diversity improve user experience, while multi‑task fusion, context‑aware learning‑to‑rank, real‑time online learning, and traffic‑control strategies balance accuracy, efficiency, and business responsiveness.
Re-rankingReal-timediversity
0 likes · 15 min read

This article details the technical evolution of Kuaishou's short‑video recommendation pipeline, focusing on sequence re‑ranking, multi‑content mixing, and on‑device re‑ranking, and explains how transformer‑based models, generator‑evaluator frameworks, and reinforcement‑learning strategies are employed to maximize overall sequence value, user engagement, and revenue.
KuaishouOn-device InferenceRe-ranking
0 likes · 15 min read

This article reviews the Sliding Spectrum Decomposition (SSD) model presented by Xiaohongshu at KDD 2021, explaining how it incorporates sliding‑window diversity into the re‑ranking stage, combines content‑based and collaborative‑filtering embeddings via the CB2CF framework, and demonstrates its effectiveness through offline and online A/B experiments.
EmbeddingRe-rankingSSD
0 likes · 14 min read

This article introduces a personalized re‑ranking model for recommendation systems, explaining the limitations of traditional point‑wise ranking, describing the PRM architecture with input, encoding, and output layers using multi‑head attention and pre‑trained personalization features, and presenting experimental results and future extensions.
AttentionCTRPersonalization
0 likes · 7 min read

The article presents a comprehensive study of re‑ranking in large‑scale industrial recommendation pipelines, identifies four key challenges—context awareness, permutation specificity, computational complexity, and business constraints—and proposes a two‑stage List Retrieval System that combines fast sequence search and a generative re‑ranking network with a deep context‑wise model, achieving significant online gains across multiple Taobao feed scenarios.
Re-rankingcontext-awaredeep learning
0 likes · 28 min read

Meituan’s search ranking system integrates Transformer‑based models across feature engineering, behavior sequence modeling, and re‑ranking, adapting AutoInt‑style embeddings and multi‑stage attention mechanisms to boost QV_CTR and NDCG, while outlining future enhancements with BERT, graph neural networks, and reinforcement learning.
Behavior ModelingMeituanRe-ranking
0 likes · 16 min read

Watchdog team won TOP1 in iQIYI & ACMMM2019 multimodal video person recognition challenge using pre‑extracted multimodal features, a 2048‑dim classifier with BCE loss, re‑ranking, DALI‑accelerated re‑detection, fine‑tuned InsightFace, and multi‑model ensembling achieving ~91% test accuracy.
Feature FusionRe-rankingmodel ensemble
0 likes · 12 min read
