SIDInspector: A Mapping-First Diagnostic Resource for Semantic-ID Tokenizers Huawei presents a tool to validate and profile semantic-ID tokenizer mappings before generator training. 📝 arxiv.org/abs/2606.10375 👨🏽‍💻 github.com/jdding/sidin...

STORM: Stepwise Token Optimization with Reward-Guided Beam Search Trains LLM query rewriters with reward-guided beam search, turning retrieval rewards into a token-level signal so BM25-based lexical retrieval rivals dense retrievers at much lower cost. 📝 arxiv.org/abs/2606.10621

Effective Reinforcement Learning for Agentic Search by Recycling Zero-Variance Queries During Training Proposes recycling zero-variance rollout groups back into the RL training pool, letting a 1.7B search agent match or surpass 7B systems. 📝 arxiv.org/abs/2606.10709 👨🏽‍💻 github.com/cxcscmu/agen...

miniReranker: Efficient Multimodal Reranking through Visual Cache Reuse and Interaction Sparsity Speeds up multimodal LLM reranking via vision-first prompting, early exit, and visual token pruning, cutting reranking runtime by up to 99%. 📝 arxiv.org/abs/2606.10759

Conan-embedding-v3: Fusing Modality-Specific Models for Omni-Modal Embedding Tencent trains modality specialists separately, fuses them, and fixes "Projector Drift" to unify text, image, video, document, and audio retrieval in one backbone. 📝 arxiv.org/abs/2606.09331

SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research Ant Group introduces a harness and training recipe that teaches a model to decompose tasks and delegate subtasks to subagents. 📝 arxiv.org/abs/2606.09730

Multi-Turn Evaluation of Deep Research Agents Under Process-Level Feedback Studies whether deep research agents improve when given feedback. 📝 arxiv.org/abs/2606.09748 👨🏽‍💻 github.com/sabharwalris...

TrustMargin: Training-Free Arbitration between Parametric Memory and Retrieved Evidence in Large Language Models Introduces a training-free layer that decides whether to trust a direct answer or a retrieval-augmented answer. 📝 arxiv.org/abs/2606.08397 👨🏽‍💻 github.com/mojixu/Trust...

Beyond Item IDs: Scaling Short-Form-Video Recommendation via Semantic-Native Long Sequence Modeling Google presents a framework for ultra-long user sequences that combines content-native Semantic IDs with a Global-Aware Compression Transformer. 📝 arxiv.org/abs/2606.07546

MM-Matryoshka: Towards Budget-Elastic Visual Document Retrieval via a 2D Multimodal Matryoshka Training Framework Introduces a 2D Matryoshka training framework for visual document retrieval. 📝 arxiv.org/abs/2606.07654

DeRes: Decoupling Residual Stability and Adaptivity for Scalable CTR Prediction Introduces a dual-path inter-layer connector for CTR Transformers that pairs an identity residual with a block-level attention residual, replacing Softmax with SiLU. 📝 arxiv.org/abs/2606.07980

When Should Queries Be Decomposed? A Stage-Aware Study of Query Decomposition for Multi-Condition Retrieval Keeps queries monolithic during initial retrieval to preserve global context, while applying decomposition only during reranking. 📝 arxiv.org/abs/2606.08577 👨🏽‍💻 github.com/EIT-NLP/Quer...

Gryphon: A Unified Architecture for Semantic-ID Generation and Item-Level Scoring in Industrial Recommendations Yandex adds an item-level scoring module to re-score concrete items resolved from generated Semantic IDs, sidestepping SID collisions. 📝 arxiv.org/abs/2606.08604

Closing the Indexing-Decoding Gap in Multimodal Generative Retrieval via Prefix Retention Optimization Introduces a framework that improves target prefix survival during beam search. 📝 arxiv.org/abs/2606.09241 👨🏽‍💻 github.com/layingfish/M...

SlimSearcher: Training Efficiency-Aware Web Agents via Adaptive Reward Gating Introduces a training framework using Pareto-efficient trajectory filtering and Adaptive Reward Gating to make deep research agents more efficient. 📝 arxiv.org/abs/2606.07074 👨🏽‍💻 github.com/AQ-MedAI/Ant...

Towards Retrieving Interaction Spaces for Agentic Search Uses BM25 to build a bounded "interaction space" of documents an agent can explore with shell tools. 📝 arxiv.org/abs/2606.06880 👨🏽‍💻 github.com/texttron/RISE

DREAM: Dynamic Refinement of Early Assignment Mappings Kuaishou fixes the cold-start bottleneck in SID-based generative recommendation by replacing one-shot static item identifiers with prior-supported candidates. 📝 arxiv.org/abs/2606.06947

Beyond Matching: Category-Guided Latent Intent Reasoning for Generative Retrieval in E-Commerce Meituan learns latent intent states guided by product category hierarchies before SID decoding, bridging the gap between shopping queries & item identifiers. 📝 arxiv.org/abs/2606.07075

RISE: A Rust Library for Inverted Index Search Engines Presents a Rust library covering the full inverted-index stack, matching or surpassing state-of-the-art C++ libraries with speedups of up to 2x while keeping memory safety. 📝 arxiv.org/abs/2606.07187 👨🏽‍💻 github.com/AngeloSav/ri...

Gated Bidirectional Linear Attention for Generative Retrieval Yandex introduces a linear-time bidirectional attention layer for long-history generative retrieval that matches self-attention quality with 8x per-layer speedup. 📝 arxiv.org/abs/2606.07317 👨🏽‍💻 github.com/matfu-pixel/...

Bradley-Terry Rankings for Recommender Systems Across Dataset Taxonomies Introduces a method for fairly ranking recommendation algorithms, enabling dataset-aware baseline selection on unseen datasets without running the models. 📝 arxiv.org/abs/2606.07492 👨🏽‍💻 github.com/fallnlove/bt...

Your UnEmbedding Matrix is Secretly a Feature Lens for Text Embeddings Introduces a training-free linear transformation that refines LLM text embeddings by filtering out high-frequency token subspaces. 📝 arxiv.org/abs/2606.07502 👨🏽‍💻 github.com/CentreChen/E...

I published Vol. 159 of "Top Information Retrieval Papers of the Week" on Substack. 🔗 recsys.substack.com/p/bridging-m...

ReverseEOL: Improving Training-free Text Embeddings via Text Reversal in Decoder-only LLMs Augments forward text embeddings with a complementary embedding derived from reversed input text, addressing the causal attention bottleneck in frozen LLMs. 📝 arxiv.org/abs/2606.05858

Answer Presence Drives RAG Rewriting Gains Ant Group presents a controlled intervention audit showing that most F1 gains from LLM rewriters in RAG pipelines are causally driven by the gold answer string appearing in the rewritten context. 📝 arxiv.org/abs/2606.05633

QCFuse: Query-Aware Cache Fusion via Compressed View for Efficient RAG Serving Introduces a compressed-view selector for RAG KV cache fusion that enables query-aware token selection without stalling the serving pipeline. 📝 arxiv.org/abs/2606.05875 👨🏽‍💻 github.com/uYanJX/QCFuse

ColBERTSaR: Sparsified ColBERT Index via Product Quantization Introduces an embedding quantization method that turns a ColBERT index into a true inverted index, yielding an index 50-70% smaller than one-bit PLAID. 📝 arxiv.org/abs/2606.05568 👨🏽‍💻 github.com/hltcoe/ColBE...

Bridging the Semantic-Collaborative Gap: An Asymmetric Graph Architecture for Cold-Start Item Recommendation Introduces an asymmetric two-tower graph architecture that encodes cold-start content from intrinsic features alone. 📝 arxiv.org/abs/2606.06225

OneReason Technical Report Kuaishou introduces a generative recommendation model that activates reasoning by combining itemic token perception in pre-training with a cognition-enhanced chain-of-thought format. 📝 arxiv.org/abs/2606.06260

Argus-Retriever: Vision-LLM Late-Interaction Retrieval with Region-Aware Query-Conditioned MoE for Visual Document Retrieval Introduces a query-conditioned late-interaction visual document retriever. 📝 arxiv.org/abs/2606.04300 👨🏽‍💻 github.com/DataScienceU...

DSIRM: Learning Query-Bridged Discrete Semantic Identifiers for E-commerce Relevance Modeling Alibaba repositions discrete semantic identifiers as relevance features, using query-bridged contrastive quantization & LLM to predict item SIDs from queries. 📝 arxiv.org/abs/2606.04374

SAILRec: Steering LLM Attention to Dual-Side Semantically Aligned Collaborative Embeddings for Recommendation Proposes an LLM-based recommender that improves the use of injected collaborative embeddings through dual-side semantic alignment. 📝 arxiv.org/abs/2606.04514

ANN Search: Recall What Matters Argues that Recall@k overstates the cost of approximation in nearest neighbor search and proposes 1/Ratio@k, a judge-free, hyperparameter-free quality measure that tracks downstream task quality more faithfully. 📝 arxiv.org/abs/2606.04522

Cartridges at Scale: Training Modular KV Caches over Large Document Collections Amazon presents a training framework for scalable multi-cartridge learning that distills document collections into reusable KV caches. 📝 arxiv.org/abs/2606.04557

EviRank: Evidence-Based Confidence Estimation for LLM-Based Ranking Estimates position-level confidence for LLM-based ranking by aggregating semantic, attention, and output evidence, with position-aware calibration. 📝 arxiv.org/abs/2606.04727 👨🏽‍💻 anonymous.4open.science/r/EviRank-CDE0

ARBOR: Online Process Rewards via a Reusable Rubric Buffer for Search Agents Alibaba presents a reusable process-reward framework that maintains a shared rubric memory to supervise the search process. 📝 arxiv.org/abs/2606.03239

Attention Calibration for Position-Fair Dense Information Retrieval Introduces an inference-time attention calibration method with a tunable strength coefficient to reduce positional bias in dense retrieval. 📝 arxiv.org/abs/2606.02737 👨🏽‍💻 github.com/impresso/fai...

Do Neural Retrievers Prefer Certain Documents? Evidence of Learned Relevance Priors Shows that supervised dense retrievers implicitly learn a query-independent relevance prior from annotation biases, making relevant but niche docs harder to retrieve. 📝 arxiv.org/abs/2606.02814

LLM-Assisted Reranking to Operationalize Nuanced Objectives in Recommender Systems Investigates how LLM-based reranking of news recommendations can amplify exposure to extreme or conspiratorial political content. 📝 arxiv.org/abs/2606.02883

Slipstream: Locality-Aware Graph Index Construction for Streaming Approximate Nearest Neighbor Search Speeds up insertions in graph indexes for streaming nearest neighbor search by reusing candidates from previous insertions. 📝 arxiv.org/abs/2606.02992 👨🏽‍💻 github.com/suey3141/Sli...

VirtualMLE: A Virtual ML Engineer that Optimizes Sequential Recommenders Introduces an LLM-agent framework that tunes sequential recommenders through a closed loop of execution, reflection, and memory. 📝 arxiv.org/abs/2606.03221 👨🏽‍💻 github.com/cst20/Virtua...

Structures Facilitate Retrieve, Rerank, and Generate Integrates document structural information across retrieval, reranking, and generation for document-grounded dialogue systems in both Chinese and English. 📝 arxiv.org/abs/2606.03247

Can LLM Rerankers Predict Their Own Ranking Performance? Introduces reranker-internal query performance prediction, showing self-consistency is well-calibrated while verbalized confidence is overconfident, and proposes two supervised methods to fix it. 📝 arxiv.org/abs/2606.03535

Skill Is Not Document: A Query-Conditional Benchmark and Two-Stage Retriever for LLM Agent Skill Routing Tencent introduces a bilingual benchmark and two-stage retriever for agent skill routing. 📝 arxiv.org/abs/2606.03565

Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories Introduces a benchmark for span-level error localization in deep-research agent trajectories, and a claim-centric auditing framework that tracks agent claims. 📝 arxiv.org/abs/2606.02060

TVIR: Building Deep Research Agents Towards Text-Visual Interleaved Report Generation Introduces a hierarchical multi-agent framework for deep research reports that interleave text with semantically grounded charts and images. 📝 arxiv.org/abs/2606.02320 👨🏽‍💻 nju-link.github.io/TVIR/

When Is 0.1% Enough? Analyzing the Combined Effects of Dimensionality Reduction and Quantization on Text Embedding Compression Systematically studies combining dimensionality reduction and quantization for text embeddings. 📝 arxiv.org/abs/2606.01074

FineVerify: Scaling Test-Time Compute with Fine-Grained Self-Verification for Agentic Search Decomposes each question into checkable sub-questions, verifies sampled candidates against each, and selects the highest-scoring answer. 📝 arxiv.org/abs/2606.00660 👨🏽‍💻 github.com/XuZhao0/fine...

OCC-RAG: Optimal Cognitive Core for Faithful Question Answering Introduces a family of small language models for faithful context-grounded QA, mid-trained on a synthetic corpus of multi-hop examples. 📝 arxiv.org/abs/2606.00683 👨🏽‍💻 github.com/optimal-cogn... 🤗 huggingface.co/occ-ai

ExpWeaver: LLM Agents Learn from Experience via Latent RAG Lets LLM agents learn from experience via latent RAG, encoding past experiences in the model's own hidden states and integrating them through cross-attention and gated residuals. 📝 arxiv.org/abs/2606.01041 👨🏽‍💻 github.com/ulab-uiuc/Ex...