MTA-Agent: An Open Recipe for Multimodal Deep Search Agents Salesforce introduces a tool-augmented agent that synthesizes verified multi-hop vision-language QA trajectories from existing VQA datasets. 📝 arxiv.org/abs/2604.06376 👨🏽‍💻 github.com/SalesforceAI...

Feedback Adaptation for Retrieval-Augmented Generation Qualcomm introduces an inference-time method that incorporates user feedback into RAG systems without retraining, enabling immediate correction and strong generalization. 📝 arxiv.org/abs/2604.06647

CASE: Cadence-Aware Set Encoding for Large-Scale Next Basket Repurchase Recommendation Walmart presents a next basket repurchase model that encodes item histories as calendar-time signals using multi-scale temporal CNNs. 📝 arxiv.org/abs/2604.06718 👨🏽‍💻 github.com/ycao21/CASE_...

Improving Semantic Proximity in Information Retrieval through Cross-Lingual Alignment Introduces a training strategy combining Jensen-Shannon Divergence and InfoNCE losses to enhance cross-lingual alignment in multilingual retrievers. 📝 arxiv.org/abs/2604.05684

A Framework for Prompt Optimization and Translation Across Foundation Models Amazon presents a governed hierarchical framework for cross-model prompt adaptation that decomposes prompts into semantic, structural, and interface components. 📝 www.amazon.science/publications...

RAG or Learning? Understanding the Limits of LLM Adaptation under Continuous Knowledge Drift in the Real World Presents a benchmark of real-world dynamic events and a time-aware retrieval baseline that organizes evidence into an Event Evolution Graph. 📝 arxiv.org/abs/2604.05096

Learning to Retrieve from Agent Trajectories Trains retrievers directly from deep research agent trajectories, using browsing actions as positives, unbrowsed results as negatives. 📝 arxiv.org/abs/2604.04949 👨🏽‍💻 yuqi-zhou.github.io/LRAT-homepage/

Tencent Advertising Algorithm Challenge 2025: All-Modality Generative Recommendation Tencent presents large-scale de-identified ad datasets with collaborative IDs and multi-modal embeddings, and findings from their 2025 challenge. 📝 arxiv.org/abs/2604.04976 👨🏽‍💻 algo.qq.com/2025

Document Optimization for Black-Box Retrieval via Reinforcement Learning Recasts document expansion as an optimization problem, fine-tuning an LM/VLM with GRPO to rewrite documents into retriever-aligned surrogates. 📝 arxiv.org/abs/2604.05087 👨🏽‍💻 github.com/omriuz/docum...

CRAB: Codebook Rebalancing for Bias Mitigation in Generative Recommendation Walmart introduces a post-hoc debiasing method for generative recommenders that identifies and splits over-popular semantic tokens in the codebook. 📝 arxiv.org/abs/2604.05113

Spike Hijacking in Late-Interaction Retrieval Adobe presents a mechanistic study showing that MaxSim pooling in late-interaction retrievers concentrates gradients on few patches, making models brittle and vulnerable compared to softer pooling alternatives. 📝 arxiv.org/abs/2604.05253

Pay Attention to Sequence Split: Uncovering the Impacts of Sub-Sequence Splitting on Sequential Recommendation Models Reveals that sub-sequence splitting significantly inflates reported performance in recent SR models. 📝 arxiv.org/abs/2604.05309 👨🏽‍💻 github.com/KingGugu/SSS...

Next-Scale Generative Reranking: A Tree-based Generative Rerank Method at Meituan Meituan introduces a tree-based generative reranking framework that progressively expands recommendation lists in a coarse-to-fine manner. 📝 arxiv.org/abs/2604.05314

Semantic Trimming and Auxiliary Multi-step Prediction for Generative Recommendation Alibaba combines adaptive token pruning with multi-step auxiliary prediction to accelerate SID-based generative recommendation training while preserving performance. 📝 arxiv.org/abs/2604.05329

Graph of Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills Presents an inference-time structural retrieval layer that builds a dependency-aware skill graph and uses reverse-weighted Personalized PageRank. 📝 arxiv.org/abs/2604.05333 👨🏽‍💻 github.com/davidliuk/gr...

Retrieve-then-Adapt: Retrieval-Augmented Test-Time Adaptation for Sequential Recommendation Handles preference shift at inference by retrieving collaboratively similar items from a memory database and fusing them with the base model's prediction. 📝 arxiv.org/abs/2604.05379

Generative Retrieval Overcomes Limitations of Dense Retrieval but Struggles with Identifier Ambiguity Shows that generative retrieval bypasses the vector bottleneck of dense retrievers on the LIMIT benchmark. 📝 arxiv.org/abs/2604.05764 👨🏽‍💻 anonymous.4open.science/r/GR-on-LIMIT/

The LLM Effect on IR Benchmarks: A Meta-Analysis of Effectiveness, Baselines, and Contamination Presents a meta-analysis of 143 papers showing LLM-based retrieval systems achieve notable gains, but also reveals measurable data contamination. 📝 arxiv.org/abs/2604.05766

The LLM Effect on IR Benchmarks: A Meta-Analysis of Effectiveness, Baselines, and Contamination Presents a meta-analysis of 143 papers showing LLM-based retrieval systems achieve notable gains, but also reveals measurable data contamination. 📝 arxiv.org/abs/2604.05766 👨🏽‍💻 github.com/MoritzStaudi...

CLEAR: Cross-Lingual Enhancement in Alignment via Reverse-training Introduces a cross-lingual retrieval loss that uses English passages as a bridge and a reverse training scheme to strengthen alignment with target languages. 📝 arxiv.org/abs/2604.05821 👨🏽‍💻 github.com/dltmddbs100/...

Masking or Mitigating? Deconstructing the Impact of Query Rewriting on Retriever Biases in RAG Systematically explores how query enhancement techniques affect dense retrieval biases. 📝 arxiv.org/abs/2604.06097

Data, Not Model: Explaining Bias Toward LLM Texts in Neural Retrievers Shows that retrievers' preference for LLM-generated text stems from artifact imbalances (fluency, lexical specificity) between positives and negatives in training data. 📝 arxiv.org/abs/2604.06163

PRAISE: Prefix-Based Rollout Reuse in Agentic Search Training Reuses prefix states from multi-turn search rollouts to generate intermediate answers, deriving step-level rewards from adjacent prefix gains for more efficient agentic search training. 📝 arxiv.org/abs/2604.03675

Search, Do not Guess: Teaching Small Language Models to Be Effective Search Agents Introduces a distillation method that trains SLMs to always retrieve evidence rather than rely on parametric knowledge. 📝 arxiv.org/abs/2604.04651 👨🏽‍💻 github.com/yizhou0409/A...

Align then Train: Efficient Retrieval Adapter Learning Introduces a label-efficient framework that trains retrieval adapters via self-supervised alignment, then supervised adaptation. 📝 arxiv.org/abs/2604.03403 👨🏽‍💻 github.com/megagonlabs/...

LLM-based Listwise Reranking under the Effect of Positional Bias Mitigates positional bias in LLM listwise rerankers via inverse propensity scoring and position-aware data augmentation during fine-tuning. 📝 arxiv.org/abs/2604.03642

Are LLM-Based Retrievers Worth Their Cost? An Empirical Study of Efficiency, Robustness, and Reasoning Overhead Presents an empirical study on BRIGHT across 14 retrievers. 📝 arxiv.org/abs/2604.03676 👨🏽‍💻 github.com/DataScienceU...

LightThinker++: From Reasoning Compression to Memory Management Lets LLMs dynamically compress intermediate thoughts and explicitly manage memory via commit/expand/fold primitives, cutting peak token usage by ~70%. 📝 arxiv.org/abs/2604.03679 👨🏽‍💻 github.com/zjunlp/Light...

Fusion and Alignment Enhancement with Large Language Models for Tail-item Sequential Recommendation Tackles the tail-item problem in sequential recommendation by adaptively fusing ID and LLM embeddings and aligning them. 📝 arxiv.org/abs/2604.03688 👨🏽‍💻 github.com/ZhifuWei/FAE...

Semantic IDs for Recommender Systems at Snapchat: Use Cases, Technical Challenges, and Design Choices Snap shares practical lessons on deploying Semantic IDs in production, introducing RQ-VAE improvements with straight-through estimation and multi-modal fusion. 📝 arxiv.org/abs/2604.03949

Formalized Information Needs Improve Large-Language-Model Relevance Judgments Shows that synthetically formalizing queries into TREC-style topics (title, description, narrative) improves LLM relevance judgments. 📝 arxiv.org/abs/2604.04140 👨🏽‍💻 github.com/irgroup/SIGI...

Beyond Fluency: Toward Reliable Trajectories in Agentic IR Google presents a position paper on failure modes in agentic IR systems, proposing verification gates and calibrated abstention to ensure trajectory integrity over plausible but unverified completions. 📝 arxiv.org/abs/2604.04269

Retrieval Augmented Conversational Recommendation with Reinforcement Learning Presents a 2-stage retrieval-augmented conversational recommender that uses LLM feedback as reward signals to align retrieval and generation. 📝 arxiv.org/abs/2604.04457 👨🏽‍💻 github.com/Yueeeeeeee/RAR

Beyond Hard Negatives: The Importance of Score Distribution in Knowledge Distillation for Dense Retrieval Proposes a sampling strategy that uniformly covers the entire teacher score spectrum, outperforming top-K and random sampling for distilling dense retrievers. 📝 arxiv.org/abs/2604.04734

SkillX: Automatically Constructing Skill Knowledge Bases for Agents Presents an automated framework for building plug-and-play skill knowledge bases for LLM agents, organizing experience into hierarchical planning, functional, and atomic skills 📝 arxiv.org/abs/2604.04804 👨🏽‍💻 github.com/zjunlp/SkillX

BidirLM: From Text to Omnimodal Bidirectional Encoders by Adapting and Composing Causal LLMs Proposes a recipe to transform causal LLMs into bidirectional encoders via masking, contrastive training, and weight merging. 📝 arxiv.org/abs/2604.02045 🤗 huggingface.co/BidirLM

Distance Comparison Operations Are Not Silver Bullets in Vector Similarity Search: A Benchmark Study on Their Merits and Limits Benchmarks 8 DCO algorithms across 10 datasets and diverse hardware. 📝 arxiv.org/abs/2604.02801 👨🏽‍💻 github.com/zzlin237/dco...

Unified and Efficient Approach for Multi-Vector Similarity Search Introduces the first native hierarchical graph index for multi-vector data, using an edge-weight function that reduces search latency by up to 14x. 📝 arxiv.org/abs/2604.02815 👨🏽‍💻 github.com/oldherd/MV-H...

Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents Finds 3–13% of citation URLs from commercial LLMs and deep research agents are hallucinated, and releases an open-source tool that reduces non-resolving citations. 📝 arxiv.org/abs/2604.03173

Overcoming the "Impracticality" of RAG: Proposing a Real-World Benchmark and Multi-Dimensional Diagnostic Framework Fujitsu introduces an enterprise RAG benchmark with a 4-axis difficulty taxonomy to diagnose system weaknesses beyond end-to-end accuracy. 📝 arxiv.org/abs/2604.02640

MBGR: Multi-Business Prediction for Generative Recommendation at Meituan Meituan presents a generative recommendation framework tailored for multi-business scenarios, using business-aware semantic IDs. 📝 arxiv.org/abs/2604.02684

AnnoRetrieve: Efficient Structured Retrieval for Unstructured Document Analysis Proposes an annotation-driven retrieval paradigm that replaces vector search with lightweight structured queries over automatically induced schemas. 📝 arxiv.org/abs/2604.02690 👨🏽‍💻 anonymous.4open.science/r/AnnoRetrie...

Bilateral Intent-Enhanced Sequential Recommendation with Embedding Perturbation-Based Contrastive Learning Injects shared intent prototypes into both item and sequence representations and uses bounded embedding-level perturbations. 📝 arxiv.org/abs/2604.02833 👨🏽‍💻 anonymous.4open.science/r/BIPCL-8E78/

Self-Optimizing Multi-Agent Systems for Deep Research Zeta Alpha explores automatic prompt optimization (TextGrad and GEPA) for multi-agent Deep Research systems, showing that self-play with a task-specific meta-prompt can match or outperform expert-crafted prompts. 📝 arxiv.org/abs/2604.02988

I published Vol. 150 of "Top Information Retrieval Papers of the Week" on Substack. 🔗 recsys.substack.com/p/on-strengt...

ContextBudget: Budget-Aware Context Management for Long-Horizon Search Agents Alibaba formulates context compression as a sequential decision problem, using curriculum-based RL to help LLM agents adaptively manage context under strict window limits. 📝 arxiv.org/abs/2604.01664

Grounded Token Initialization for New Vocabulary in LMs for Generative Recommendation Learns semantically meaningful embeddings for new vocabulary tokens before fine-tuning, addressing the collapse caused by standard mean initialization. 📝 arxiv.org/abs/2604.02324

Optimizing RAG Rerankers with LLM Feedback via Reinforcement Learning Proposes an RL framework that aligns rerankers with LLM generation quality by formulating reranking as sequential decision-making. 📝 arxiv.org/abs/2604.02091

Neuro-RIT: Neuron-Guided Instruction Tuning for Robust Retrieval-Augmented Language Model Presents a neuron-level instruction tuning framework that disentangles neurons responsible for processing relevant vs. irrelevant contexts to improve RAG robustness. 📝 arxiv.org/abs/2604.02194

Multi-Agent Video Recommenders: Evolution, Patterns, and Open Challenges Google surveys the evolution of multi-agent video recommendation systems, presenting a taxonomy of collaborative patterns from early MARL systems to LLM-powered architectures. 📝 arxiv.org/abs/2604.02211