Step-DeepResearch Technical Report Presents a cost-effective 32B parameter deep research agent achieving expert-level performance through atomic capability training and progressive optimization from mid-training to RL. 📝 arxiv.org/abs/2512.20491 👨🏽‍💻 github.com/stepfun-ai/S...

M³KG-RAG: Multi-hop Multimodal Knowledge Graph-enhanced Retrieval-Augmented Generation Proposes a framework with multi-hop multimodal knowledge graphs and a pruning mechanism to improve audio-visual reasoning in multimodal large language models. 📝 arxiv.org/abs/2512.20136

Retrieval-augmented Prompt Learning for Pre-trained Foundation Models Presents a retrieval-augmented framework that enhances prompt learning by decoupling knowledge from memorization through a knowledge-store of training data. 📝 arxiv.org/abs/2512.20145 👨🏽‍💻 github.com/zjunlp/Promp...

Collaborative Group-Aware Hashing for Fast Recommender Systems Proposes a hashing method that integrates inherent group information to learn hash codes, improving both accuracy in sparse settings and efficiency for large-scale recommendations. 📝 arxiv.org/abs/2512.20172

Mem³R: Memory Retrieval via Reflective Reasoning for LLM Agents Introduces an autonomous memory-retrieval controller that transforms retrieve-then-answer pipelines into closed-loop processes using evidence-gap tracking. 📝 arxiv.org/abs/2512.20237 👨🏽‍💻 github.com/Leagein/memr3

Representation-Enhanced Cascading Multi-Level Interest Learning for Multi-Behavior Recommendation Presents a parallel learning framework that captures multi-level positive and negative feedback signals from behavioral sequences. 📝 dl.acm.org/doi/10.1145/... 👨🏽‍💻 github.com/lhybq/PPN-ARE

Laser: Governing Long-Horizon Agentic Search via Structured Protocol and Context Register Introduces a framework that stabilizes agentic search through symbolic action protocols and compact context management. 📝 arxiv.org/abs/2512.20458 👨🏽‍💻 github.com/ShootingWong...

Making Large Language Models Efficient Dense Retrievers Finds that MLP layers in LLM-based retrievers are highly redundant while attention layers remain critical, enabling substantial compression through coarse-to-fine MLP pruning. 📝 arxiv.org/abs/2512.20612 👨🏽‍💻 github.com/Yibin-Lei/Ef...

MegaRAG: Multimodal Knowledge Graph-Based Retrieval Augmented Generation Presents a multimodal graph-based RAG method that automatically constructs knowledge graphs from visual documents, enabling cross-modal reasoning for better content understanding. 📝 arxiv.org/abs/2512.20626

How important is Recall for Measuring Retrieval Quality? Introduces a simple measure independent of total relevant documents and evaluates retrieval quality metrics against LLM-based response quality judgments across multiple datasets. 📝 arxiv.org/abs/2512.20854

Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning NVIDIA releases a 30B parameter MoE model with competitive accuracy and 3.3x higher throughput while supporting 1M token contexts. 📝 arxiv.org/abs/2512.20848 🤗 huggingface.co/nvidia/NVIDI...

NVIDIA Nemotron 3: Efficient and Open Intelligence NVIDIA introduces a family of models (Nano, Super, Ultra) using hybrid Mamba-Transformer MoE architecture with up to 1M token context and state-of-the-art reasoning performance. 📝 arxiv.org/abs/2512.20856

Accurate and Diverse Recommendations via Propensity-Weighted Linear Autoencoders Introduces a propensity scoring method using sigmoid functions on logarithmic item frequency to improve recommendation diversity while maintaining accuracy. 📝 arxiv.org/abs/2512.20896 👨🏽‍💻 github.com/cars1015/IPS...

ReaSeq: Unleashing World Knowledge via Reasoning for Sequential Modeling Alibaba introduces a reasoning-enhanced framework that leverages LLMs to address knowledge poverty and systemic blindness in recommender systems. 📝 arxiv.org/abs/2512.21257

C2LLM Technical Report: A New Frontier in Code Retrieval via Adaptive Cross-Attention Pooling Ant Group introduces a family of code embedding models using Pooling by Multihead Attention to break information bottlenecks in code retrieval. 📝 arxiv.org/abs/2512.21332 👨🏽‍💻 github.com/codefuse-ai/...

Faster Distributed Inference-Only Recommender Systems via Bounded Lag Synchronous Collectives Huawei proposes a bounded lag synchronous operation for distributed recommender systems that improves both latency and throughput in inference-only DLRM runs. 📝 arxiv.org/abs/2512.19342

LIR³AG: A Lightweight Rerank Reasoning Strategy Framework for Retrieval-Augmented Generation Enables non-reasoning models to transfer reasoning strategies by restructuring retrieved evidence into coherent reasoning chains. 📝 arxiv.org/abs/2512.18329

Factorized Transport Alignment for Multimodal and Multiview E-commerce Representation Learning Etsy unifies multimodal and multi-view learning for e-commerce search, using factorized transport to efficiently align primary and non-primary images with text views. 📝 arxiv.org/abs/2512.18117

Efficient Optimization of Hierarchical Identifiers for Generative Recommendation Proposes greedy and hybrid tree construction methods that reduce identifier building time to 2-8% while maintaining or improving retrieval quality. 📝 arxiv.org/abs/2512.18434 👨🏽‍💻 github.com/joshrosie/re...

Generative Giants, Retrieval Weaklings: Why do Multimodal Large Language Models Fail at Multimodal Retrieval? Analyzes why MLLMs underperform in zero-shot multimodal retrieval, revealing text-dominated spaces and misaligned feature components. 📝 arxiv.org/abs/2512.19115

QuCo-RAG: Quantifying Uncertainty from the Pre-training Corpus for Dynamic Retrieval-Augmented Generation Presents a dynamic RAG approach that uses pre-training corpus statistics to determine when to retrieve. 📝 arxiv.org/abs/2512.19134 👨🏽‍💻 github.com/ZhishanQ/QuC...

Scalable Distributed Vector Search via Accuracy Preserving Index Construction Presents a hierarchical vector index system that achieves up to 9.64x higher throughput by using balanced partition granularity and accuracy-preserving recursive construction. 📝 arxiv.org/abs/2512.17264

DEER: A Comprehensive and Reliable Benchmark for Deep-Research Expert Reports Introduces a benchmark for evaluating AI-generated research reports with 50 tasks across 13 domains, combining expert-grounded rubrics and document-level fact-checking. 📝 arxiv.org/abs/2512.17776

Exploiting ID-Text Complementarity via Ensembling for Sequential Recommendation Snap introduces a method that trains ID-based and text-based sequential recommendation models independently, then combines them through ensembling to leverage their complementary strengths. 📝 arxiv.org/abs/2512.17820

A Reproducible and Fair Evaluation of Partition-aware Collaborative Filtering Presents a benchmark of partition-aware collaborative filtering methods, revealing that FPSR models remain competitive but don't consistently outperform block-aware baselines. 📝 arxiv.org/abs/2512.17015

TCDE: Topic-Centric Dual Expansion of Queries and Documents with Large Language Models for Information Retrieval Introduces a training-free framework that expands both queries and documents using LLMs to achieve topic-centric alignment, improving retrieval performance. 📝 arxiv.org/abs/2512.17164

MMRAG-RFT: Two-stage Reinforcement Fine-tuning for Explainable Multi-modal Retrieval-augmented Generation Introduces a two-stage reinforcement fine-tuning framework that enables multi-modal LLMs to provide explainable reasoning in RAG. 📝 arxiv.org/abs/2512.17194

Mindscape-Aware Retrieval Augmented Generation for Improved Long Context Understanding Equips RAG systems with global context awareness through hierarchical summarization, enabling better retrieval and reasoning over long documents across languages. 📝 arxiv.org/abs/2512.17220

Warmer for Less: A Cost-Efficient Strategy for Cold-Start Recommendations at Pinterest Pinterest presents a cost-efficient solution to cold-start recommendations through residual connections, score regularization, and manifold mixup, adding only 5% to model parameters. 📝 arxiv.org/abs/2512.17277

I published Vol. 135 of "Top Information Retrieval Papers of the Week" on Substack. 🔗 recsys.substack.com/p/understand...

How Do Graph Signals Affect Recommendation: Unveiling the Mystery of Low and High-Frequency Graph Signals Proves that low-frequency and high-frequency graph signals are equivalent in recommendation tasks. 📝 arxiv.org/abs/2512.15744 👨🏽‍💻 github.com/mojosey/SimGCF

Leveraging Spreading Activation for Improved Document Retrieval in Knowledge-Graph-Based RAG Systems Introduces a RAG framework using spreading activation algorithm to retrieve information from automatically constructed knowledge graphs. 📝 arxiv.org/abs/2512.15922

On Recommending Category: A Cascading Approach eBay presents a cascading category recommender with variational autoencoder to encode item-level information for category-level recommendations. 📝 arxiv.org/abs/2512.16033

Needle in the Web: A Benchmark for Retrieving Targeted Web Pages in the Wild Introduces a benchmark with 663 queries across seven domains to evaluate search agents on fuzzy exploratory tasks. 📝 arxiv.org/abs/2512.16553 👨🏽‍💻 github.com/TangoWhiskym...

AdaSearch: Balancing Parametric Knowledge and Search in Large Language Models via Reinforcement Learning Presents an RL framework that helps LLMs decide when to search externally versus relying on internal knowledge. 📝 arxiv.org/abs/2512.16883 👨🏽‍💻 github.com/hank0316/Ada...

LinkedOut: Linking World Knowledge Representation Out of Video LLM for Next-Generation Video Recommendation LinkedIn extracts world knowledge from video LLMs to enable personalized video recommendation while operating directly on raw frames without handcrafted labels. 📝 arxiv.org/abs/2512.16891

Inference Time Feature Injection: A Lightweight Approach for Real-Time Recommendation Freshness Tubi proposes injecting recent watch history at inference time without model retraining, achieving 0.47% engagement lift in long-form video streaming. 📝 arxiv.org/abs/2512.14734

Where to Explore: A Reach and Cost-Aware Approach for Unbiased Data Collection in Recommender Systems Tubi presents a cost-aware exploration strategy that places randomized content at optimal scroll depths to collect unbiased interaction data while preserving engagement. 📝 arxiv.org/abs/2512.14733

Behavior Tokens Speak Louder: Disentangled Explainable Recommendation with Behavior Vocabulary Tokenizes user behaviors into discrete sequences to enable LLMs to generate explainable recommendations without fine-tuning. 📝 arxiv.org/abs/2512.15614 👨🏽‍💻 github.com/fxsxjtu/BEAT

DTRec: Learning Dynamic Reasoning Trajectories for Sequential Recommendation Adaptively adjusts reasoning direction and depth for sequential recommendation through hierarchical process supervision and adaptive reasoning halting. 📝 arxiv.org/abs/2512.14036

From Feature Interaction to Feature Generation: A Generative Paradigm of CTR Prediction Models Tencent presents a generative framework for CTR prediction that shifts from feature interaction to feature generation, reducing embedding collapse. 📝 arxiv.org/abs/2512.14041 👨🏽‍💻 github.com/USTC-StarTea...

AsarRec: Adaptive Sequential Augmentation for Robust Self-supervised Sequential Recommendation Learns adaptive data augmentation strategies for sequential recommendation through constrained matrix transformations. 📝 arxiv.org/abs/2512.14047 👨🏽‍💻 github.com/Kaike-Zhang/...

Dynamic Context Selection for Retrieval-Augmented Generation: Mitigating Distractors and Positional Bias Introduces a classifier that dynamically predicts the optimal number of documents to retrieve for each query. 📝 arxiv.org/abs/2512.14313

RecGPT-V2 Technical Report Alibaba presents an enhanced LLM-powered recommendation system that uses hierarchical multi-agent reasoning and reduces GPU consumption by 60% while improving recommendation metrics. 📝 arxiv.org/abs/2512.14503

Extending the Context of Pretrained LLMs by Dropping Their Positional Embeddings Introduces a method that removes positional embeddings after pretraining to enable zero-shot context extension without expensive long-context finetuning. 📝 arxiv.org/abs/2512.12167 👨🏽‍💻 github.com/SakanaAI/DroPE

Memory in the Age of AI Agents Surveys agent memory through forms (token-level, parametric, latent), functions (factual, experiential, working), and dynamics, distinguishing it from LLM memory, RAG, and context engineering. 📝 arxiv.org/abs/2512.13564 👨🏽‍💻 github.com/Shichun-Liu/...

Breaking the Curse of Dimensionality: On the Stability of Modern Vector Retrieval Introduces stability theory for multi-vector search showing Chamfer distance preserves stability while average pooling may destroy it. 📝 arxiv.org/abs/2512.12458 👨🏽‍💻 github.com/vihan-lakshm...

FuXi-γ: Efficient Sequential Recommendation with Exponential-Power Temporal Encoder and Diagonal-Sparse Positional Mechanism Huawei accelerates sequential recommendation training through temporal encoding and sparse attention. 📝 arxiv.org/abs/2512.12740 👨🏽‍💻 github.com/Yeedzhi/FuXi...

Hindsight Is 20/20: Building Agent Memory That Retains, Recalls, and Reflects Presents a memory architecture organizing agent knowledge into four networks (world, experience, opinion, observation) with retain, recall, and reflect operations. 📝 arxiv.org/abs/2512.12818 👨🏽‍💻 github.com/vectorize-io...

Do Reviews Matter for Recommendations in the Era of Large Language Models? Introduces a framework to evaluate review-aware recommender systems, showing that removing or distorting reviews doesn't necessarily hurt recommendation accuracy. 📝 arxiv.org/abs/2512.12978 👨🏽‍💻 github.com/zhytk/RAREva...