Rescaling MLM-Head for Neural Sparse Retrieval
Finds that pretrained encoders with large MLM-head scales face degradation in sparse retrieval, and introduces a zero-cost rescaling correction to stabilize training.
š arxiv.org/abs/2606.18811
On the Memorization Behavior of LLMs in Generative Recommendation: Observations, Implications, and Training Strategies
Snap shows that LLM recommenders rely heavily on one-hop memorization and teaches them richer relations.
š arxiv.org/abs/2606.17276
šØš½āš» github.com/snap-researc...
Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search
Diversifies the first-turn query across parallel search rollouts to avoid redundant retrieval, improving multi-hop QA at matched compute.
š arxiv.org/abs/2606.17209
šØš½āš» github.com/cxcscmu/dive...
MCompassRAG: Topic Metadata as a Semantic Compass for Paragraph-Level Retrieval
Introduces a framework enriching coarse chunks with topic metadata to improve paragraph-level retrieval efficiency and precision.
š arxiv.org/abs/2606.18508
šØš½āš» github.com/AmirAbaskohi...
Do Generative Recommenders Deepen the Information Cocoon? A Closed-Loop Simulation with LLM-powered User Simulators
Introduces a framework with LLM-powered user agents to study information cocoons in generative recommenders.
š arxiv.org/abs/2606.17707
šØš½āš» github.com/Dregen-Yor/R...
SproutRAG: Attention-Guided Tree Search with Progressive Embeddings for Long-Document RAG
Introduces a framework organizing sentence chunks into a binary tree for multi-granularity retrieval without external LLMs.
š arxiv.org/abs/2606.18381
šØš½āš» github.com/AmirAbaskohi...
RSRank: Learning Relevance from Representational Shifts
Adobe introduces a lightweight reranking framework that learns relevance from the representational shift a document induces in a query's internal state, filtering irrelevant content at a natural zero threshold.
š arxiv.org/abs/2606.17468
Lost in a Single Vector: Improving Long-Document Retrieval with Chunk Evidence Aggregation
Presents a training-free method to improve long-document dense retrieval by aggregating chunk embeddings into a single vector.
š arxiv.org/abs/2606.18781
šØš½āš» github.com/PunchlineAAA...
Querit-Reranker: Training Compact Multilingual Rerankers via Efficient Label-Free Distribution Adaptation
Presents a multilingual cross-encoder reranker family trained via a data-centric pipeline for label-efficient adaptation.
š arxiv.org/abs/2606.19037
š¤ huggingface.co/Querit/Querit
Learned sparse retrieval (LSR) models such as SPLADE have traditionally used BERT-style masked language models as backbone encoders. A natural expectation is that replacing BERT with stronger pretrain...
Generative recommendation (GR) has emerged as a promising direction for recommender systems. Recently, large language models (LLMs) have been increasingly adopted for GR, as their rich pretrained know...
arxiv.org
Test-time scaling for agentic search typically increases depth (i.e., more turns and tokens per trajectory) or breadth (i.e., more parallel rollouts). Here we focus on breadth scaling, showing that st...
Retrieval-augmented generation (RAG) systems depend critically on how documents are chunked and searched. Fine-grained chunks can improve retrieval precision but expand the search space, increasing la...
arxiv.org
Recommender systems alleviate information overload, yet repeated feedback between recommendations and user interactions can reinforce existing preferences and narrow users' exposure, forming informati...
arxiv.org
Retrieval-augmented generation (RAG) systems must balance retrieval granularity with contextual coherence, a challenge that existing methods address through LLM-guided chunking, single-level context e...
As enterprises deploy RAG-based systems to provide grounded responses to user queries, reranking has become a critical component for the final filtering step that separates relevant from distracting o...
Deployable multilingual rerankers must generalize across languages, domains, and target ranking tasks while remaining efficient enough for second-stage reranking. However, adapting them to new target ...
arxiv.org
Dense retrieval ranks one query vector against one document vector. On long documents, this interface can fail when a short but decisive span is weakened during document encoding before ranking. We st...
