A RAG system shouldn't need a retriever to answer "What is the capital of France?". But how do we systematically identify these cases? blog.reachsumit.com/posts/2025/0...

Relevance to Utility: Process-Supervised Rewrite for RAG Directly optimizes document rewriting in RAG systems by maximizing answer generation probability through process supervision from large language models. 📝 arxiv.org/abs/2509.15577 👨🏽‍💻 anonymous.4open.science/r/R2U-D547/

SERVAL: Surprisingly Effective Zero-Shot Visual Document Retrieval Powered by Large Vision and Language Models Introduces a zero-shot approach that uses VLMs to generate text descriptions of document images, then embeds them with standard text encoders for retrieval. 📝 arxiv.org/abs/2509.15432

Chunk Knowledge Generation Model for Enhanced Information Retrieval: A Multi-task Learning Approach Proposes a T5-based multi-task framework that simultaneously generates titles, candidate questions, and keywords from document chunks. 📝 arxiv.org/abs/2509.15658

Understanding Embedding Scaling in Collaborative Filtering Discovers two phenomena when scaling collaborative filtering models, revealing that interaction noise significantly affects embedding scalability. 📝 arxiv.org/abs/2509.15709

CodeRAG: Finding Relevant and Necessary Knowledge for Retrieval-Augmented Repository-Level Code Completion Introduces a framework for repository-level code completion. 📝 arxiv.org/abs/2509.16112 👨🏽‍💻 github.com/KDEGroup/Cod...

I published Vol. 122 of "Top Information Retrieval Papers of the Week" on Substack. 🔗 recsys.substack.com/p/a-unified-...

Evaluating Large Language Models for Cross-Lingual Retrieval Systematically evaluates LLMs for cross-lingual information retrieval, comparing dense multilingual bi-encoders with translation-based approaches. 📝 arxiv.org/abs/2509.14749 👨🏽‍💻 github.com/mainlp/llm-c...

Enhancing Retrieval Augmentation via Adversarial Collaboration Introduces a framework with two heterogeneous agents that engage in adversarial collaboration to address retrieval hallucinations in RAG systems. 📝 arxiv.org/abs/2509.14750 👨🏽‍💻 anonymous.4open.science/r/AC-RAG/

What Matters in LLM-Based Feature Extractor for Recommender? A Systematic Analysis of Prompts, Models, and Adaptation Introduces a modular framework that systematically analyzes LLM-as-feature-extractor pipelines for sequential recommendation. 📝 arxiv.org/abs/2509.14979

MA-DPR: Manifold-aware Distance Metrics for Dense Passage Retrieval Introduces manifold-aware distance metrics that capture non-linear embedding structures through graph-based representations, improving dense passage retrieval performance by up to 26%. 📝 arxiv.org/abs/2509.13562

Modernizing Facebook Scoped Search: Keyword and Embedding Hybrid Retrieval with LLM Evaluation Meta introduces a blended retrieval system combining traditional keyword search with embedding-based retrieval for Facebook Group search. 📝 arxiv.org/abs/2509.13603

Sequential Data Augmentation for Generative Recommendation Snap presents a framework that models data augmentation as a three-step sampling process to improve generative recommendation systems through better training distribution control. 📝 arxiv.org/abs/2509.13648

Improving Context Fidelity via Native Retrieval-Augmented Reasoning Teaches LLMs to integrate in-context evidence within their reasoning process, improving context fidelity and reducing hallucinations. 📝 arxiv.org/abs/2509.13683 👨🏽‍💻 foundationagents.github.io/CARE/

Enhancing Time Awareness in Generative Recommendation Introduces a generative recommendation framework that captures temporal dynamics through time-aware prompting and trend-aware inference. 📝 arxiv.org/abs/2509.13957 👨🏽‍💻 github.com/skleee/GRUT

GEM-Bench: A Benchmark for Ad-Injected Response Generation within Generative Engine Marketing Introduces a benchmark for evaluating ad-injected response generation in generative engine marketing. 📝 arxiv.org/abs/2509.14221

Large Language Models in Document Intelligence: A Comprehensive Survey, Recent Advances, Challenges and Future Trends Surveys 300+ papers on LLMs in document intelligence, covering RAG, long context processing, and fine-tuning approaches. 📝 dl.acm.org/doi/10.1145/...

Conan-Embedding-v2: Training an LLM from Scratch for Text Embeddings Tencent introduces a 1.4B-parameter LLM trained from scratch for text embeddings, achieving SOTA performance on MTEB benchmarks through soft-masking and cross-lingual retrieval. 📝 arxiv.org/abs/2509.12892

WebResearcher: Unleashing unbounded reasoning capability in Long-Horizon Agents Alibaba presents an iterative deep-research paradigm that reformulates research as a Markov Decision Process, enabling high-quality reasoning. 📝 arxiv.org/abs/2509.13309 👨🏽‍💻 github.com/Alibaba-NLP/...

zELO: ELO-inspired Training Method for Rerankers and Embedding Models Introduces a multi-stage training pipeline that uses ELO scoring systems to train state-of-the-art open-weight reranker models. 📝 arxiv.org/abs/2509.12541 👨🏽‍💻 huggingface.co/zeroentropy/...

What News Recommendation Research Did (But Mostly Didn't) Teach Us About Building A News Recommender Reflects on building a live news recommendation platform, highlighting gaps between research literature and real-world systems. 📝 arxiv.org/abs/2509.12361

Topic Coverage-based Demonstration Retrieval for In-Context Learning Selects demonstrations for in-context learning by identifying required topics and retrieving examples that cover previously uncovered knowledge areas. 📝 arxiv.org/abs/2509.12451 👨🏽‍💻 github.com/WonbinKweon/...

LEAF: Knowledge Distillation of Text Embedding Models with Teacher-Aligned Representations MongoDB presents a knowledge distillation framework that creates smaller embedding models aligned to their teachers, enabling flexible asymmetric architectures. 📝 arxiv.org/abs/2509.12539

InfoGain-RAG: Boosting Retrieval-Augmented Generation via Document Information Gain-based Reranking and Filtering Measures how much retrieved documents actually improve LLM generation confidence, enabling better document filtering in RAG systems. 📝 arxiv.org/abs/2509.12765

A Learnable Fully Interacted Two-Tower Model for Pre-Ranking System Ant Group introduces a two-tower architecture that enables rich information interactions while maintaining inference efficiency for large-scale recommendation systems. 📝 arxiv.org/abs/2509.12948

Learning Decomposed Contextual Token Representations from Pretrained and Collaborative Signals for Generative Recommendation Preserves pretrained semantics while enhancing token adaptability through contextualized composition and embedding fusion. 📝 arxiv.org/abs/2509.10468

A Survey on Retrieval And Structuring Augmented Generation with Large Language Models Presents a survey of Retrieval-And-Structuring augmented generation that integrates dynamic retrieval with structured knowledge representations. 📝 arxiv.org/abs/2509.10697

Automated MCQA Benchmarking at Scale: Evaluating Reasoning Traces as Retrieval Sources for Domain Adaptation of Small Language Models Introduces a framework for generating scientific MCQAs from literature. 📝 arxiv.org/abs/2509.10744 👨🏽‍💻 github.com/ramanathanla...

CultureSynth: A Hierarchical Taxonomy-Guided and Retrieval-Augmented Framework for Cultural Question-Answer Synthesis Alibaba combines hierarchical cultural taxonomy with RAG-based synthesis to generate culturally relevant QA pairs. 📝 arxiv.org/abs/2509.10886 👨🏽‍💻 github.com/Eyr3/Culture...

Do Large Language Models Favor Recent Content? A Study on Recency Bias in LLM-Based Reranking Finds a bias in LLM-based rerankers across 7 models, including GPT-4o and Qwen2.5, where "fresh" passages are promoted over older content regardless of relevance. 📝 arxiv.org/abs/2509.11353

Decoding in Latent Spaces for Efficient Inference in LLM-based Recommendation Bypasses autoregressive decoding, achieving over 10× faster inference than language-space decoding while maintaining performance. 📝 arxiv.org/abs/2509.11524

HiChunk: Evaluating and Enhancing Retrieval-Augmented Generation with Hierarchical Chunking Tencent introduces a benchmark and framework for multi-level document structuring, achieving better chunking quality within a reasonable time. 📝 arxiv.org/abs/2509.11552 👨🏽‍💻 github.com/TencentCloud...

SAQ: Pushing the Limits of Vector Quantization through Code Adjustment and Dimension Segmentation Introduces a vector quantization method that achieves up to 80% reduction in quantization error and accelerates encoding speed by over 80x. 📝 arxiv.org/abs/2509.12086

DeepDive: Advancing Deep Search Agents with Knowledge Graphs and Multi-Turn RL Presents automated synthesis of complex questions from knowledge graphs and multi-turn RL to enhance LLMs' long-horizon reasoning with deep search capabilities. 📝 arxiv.org/abs/2509.10446 👨🏽‍💻 github.com/THUDM/DeepDive

Faster and Memory-Efficient Training of Sequential Recommendation Models for Large Catalogs Introduces a method for GPU-efficient cross-entropy loss with negative sampling, accelerating training by up to 2x while reducing memory consumption. 📝 arxiv.org/abs/2509.09682 👨🏽‍💻 github.com/On-Point-RND...

Powering Job Search at Scale: LLM-Enhanced Query Understanding in Job Matching Systems LinkedIn presents a unified LLM framework that replaces multiple NER models for job search query understanding, improving relevance while reducing system complexity. 📝 arxiv.org/abs/2509.09690

Personas within Parameters: Fine-Tuning Small Language Models with Low-Rank Adapters to Mimic User Behaviors JP Morgan shows how small language models can be fine-tuned with low-rank adapters to create scalable user behavior agents for recommender systems. 📝 arxiv.org/abs/2509.09689

HANRAG: Heuristic Accurate Noise-resistant Retrieval-Augmented Generation for Multi-hop Question Answering ANT Group efficiently handles multi-hop queries by routing questions to appropriate processing chains and filtering noise from retrieved documents. 📝 arxiv.org/abs/2509.09713

RecoWorld: Building Simulated Environments for Agentic Recommender Systems Meta introduces a blueprint for creating simulated environments where virtual users interact with AI-powered recommender systems through multi-turn conversations. 📝 arxiv.org/abs/2509.10397

I published Vol. 121 of "Top Information Retrieval Papers of the Week" on Substack. 🔗 recsys.substack.com/p/scaling-mu...

Generative Engine Optimization: How to Dominate AI Search Introduces a framework for optimizing content visibility in AI-powered search engines like ChatGPT and Perplexity, revealing systematic biases toward earned media over brand-owned content. 📝 arxiv.org/abs/2509.08919

We're Still Doing It (All) Wrong: Recommender Systems, Fifteen Years Later Argues that recommender systems research continues to prioritize narrow performance gains over human-centered evaluation, calling for fundamental shifts. 📝 arxiv.org/abs/2509.09414

Boosting Data Utilization for Multilingual Dense Retrieval Enhances multilingual dense retrieval by constructing high-quality hard negative samples and effective mini-batches. 📝 arxiv.org/abs/2509.09459 👨🏽‍💻 github.com/miaomiao1205...

Beyond Sequential Reranking: Reranker-Guided Search Improves Reasoning Intensive Retrieval Proposes a graph-based search method that strategically selects documents for reranking. 📝 arxiv.org/abs/2509.07163 👨🏽‍💻 github.com/xuhaike/Rera...

Benchmarking Information Retrieval Models on Complex Retrieval Tasks Presents a benchmark of eight complex retrieval tasks showing top models struggle with multi-aspect queries, scoring only 0.346 nDCG@10. 📝 arxiv.org/abs/2509.07253 👨🏽‍💻 github.com/jfkback/crumb

Multi-view-guided Passage Reranking with Large Language Models Introduces a non-generative LLM reranking method that captures diverse relevance signals through multi-view encoding. 📝 arxiv.org/abs/2509.07485 👨🏽‍💻 github.com/bulbna/MVP

MoLoRAG: Bootstrapping Document Understanding via Multi-modal Logic-aware Retrieval Presents a graph-based retrieval method that combines semantic and logical relevance for multi-page document understanding. 📝 arxiv.org/abs/2509.07666 👨🏽‍💻 github.com/WxxShirley/M...

A Survey of Long-Document Retrieval in the PLM and LLM Era Provides a comprehensive survey of long-document retrieval methods, documenting the evolution from classical approaches to modern PLM and LLM-based techniques. 📝 arxiv.org/abs/2509.07759

Query Expansion in the Age of Pre-trained and Large Language Models: A Comprehensive Survey Surveys query expansion techniques in the PLM/LLM era through paradigms, model taxonomy, and practice-oriented guidance across domains. 📝 arxiv.org/abs/2509.07794

Revealing the Numeracy Gap: An Empirical Investigation of Text Embedding Models Exposes significant limitations in how current text embedding models preserve numerical information, with accuracy barely above chance on financial numeric reasoning tasks. 📝 arxiv.org/abs/2509.05691