Sumit

In the final post of the Adaptive RAG series, we explore how to treat selective retrieval as a core, learned skill, moving from passive observation to active, intelligent decision-making. blog.reachsumit.com/posts/2025/10/…

F2LLM-v2: Inclusive, Performant, and Efficient Embeddings for a Multilingual World Ant Group introduces a family of 8 multilingual embedding models (80M–14B) trained on 60M samples across 200+ languages. 📝 arxiv.org/abs/2603.19223 👨🏽‍💻 github.com/codefuse-ai/Co…

Negative Sampling Techniques in Information Retrieval: A Survey Presents a comprehensive survey and taxonomy of negative sampling techniques for dense retrieval, covering random, static/dynamic hard negative mining, false negative mitigation, and more. 📝arxiv.org/abs/2603.18005

Retrieval-Augmented LLM Agents: Learning to Learn from Experience Naver Labs presents a framework that combines experience retrieval with LoRA fine-tuning to improve LLM agent generalization to unseen tasks. 📝 arxiv.org/abs/2603.18272

Total Recall QA: A Verifiable Evaluation Suite for Deep Research Agents Introduces a benchmark for evaluating deep research agents that requires retrieving all relevant documents to answer aggregation-style questions. 📝 arxiv.org/abs/2603.18516 👨🏽‍💻 github.com/mahta-r/total-…

Hypothesis-Conditioned Query Rewriting for Decision-Useful Retrieval Proposes a training-free RAG framework that generates 3 retrieval queries (support, distinction, key features) conditioned on a working answer hypothesis. 📝 arxiv.org/abs/2603.19008 👨🏽‍💻 anonymous.4open.science/r/HCQR-1C2E

VLM2Rec: Resolving Modality Collapse in Vision-Language Model Embedders for Multimodal Sequential Recommendation Introduces a VLM embedder-based framework for sequential recommendation that resolves modality collapse. 📝 arxiv.org/abs/2603.17450

OPERA: Online Data Pruning for Efficient Retrieval Model Adaptation Amazon presents a data pruning framework for dense retriever finetuning that adaptively modulates sampling probabilities, improving both ranking and retrieval. 📝 arxiv.org/abs/2603.17205

CRE-T1 Preview Technical Report: Beyond Contrastive Learning for Reasoning-Intensive Retrieval Introduces a generative retrieval model that replaces static contrastive representation alignment with dynamic per-query reasoning trajectories. 📝 arxiv.org/abs/2603.17387

A Unified Language Model for Large Scale Search, Recommendation, and Reasoning @denadai2 et al. at Spotify introduce a tool-free framework that adapts a single LLM to jointly support search, recommendation, and reasoning over a 10M+ item catalog. 📝 arxiv.org/abs/2603.17533

Deploying Semantic ID-based Generative Retrieval for Large-Scale Podcast Discovery at Spotify Spotify presents a production-scale generative recommender that frames podcast discovery as an instruction-following task over Semantic IDs. 📝 arxiv.org/abs/2603.17540

Recursive Language Models Meet Uncertainty: The Surprising Effectiveness of Self-Reflective Program Search for Long Context @KeivanAlizadeh2 et al. at Apple use uncertainty-aware self-reflection to guide how language models interact with long contexts. 📝arxiv.org/abs/2603.15653

MiroThinker-1.7 & H1: Towards Heavy-Duty Research Agents via Verification Presents a research agent that improves long-horizon reasoning through agentic mid-training & verification-centric reasoning at local and global levels. 📝arxiv.org/abs/2603.15726 👨🏽‍💻github.com/MiroMindAI/Mir…

RecBundle: A Next-Generation Geometric Paradigm for Explainable Recommender Systems Introduces Fiber Bundle theory from differential geometry to decouple user collaboration topology from individual preference evolution in recommender systems. 📝 arxiv.org/abs/2603.16088

IndexRAG: Bridging Facts for Cross-Document Reasoning at Index Time Introduces a RAG approach that shifts cross-document reasoning from online inference to offline indexing by generating bridging facts through shared entities. 📝 arxiv.org/abs/2603.16415

OpenSeeker: Democratizing Frontier Search Agents by Fully Open-Sourcing Training Data Introduces a fully open-source search agent achieving frontier-level performance, using fact-grounded QA synthesis from web graphs. 📝 arxiv.org/abs/2603.15594 👨🏽‍💻 github.com/rui-ye/OpenSee…

APEX-Searcher: Augmenting LLMs' Search Capabilities through Agentic Planning and Execution Enhances multi-hop RAG by decoupling strategic task planning (via RL) from iterative retrieval execution (via SFT), enabling LLMs to decompose complex queries. 📝 arxiv.org/abs/2603.13853

Learning Retrieval Models with Sparse Autoencoders @thibault_formal et al. at Naver Labs present SPLARE, a learned sparse retrieval method that replaces vocabulary-based projections with sparse autoencoders, producing language-agnostic latent features. 📝arxiv.org/abs/2603.13277

MURE: Hierarchical Multi-Resolution Encoding via Vision-Language Models for Visual Document Retrieval Presents a multiresolution VLM encoding method for visual document retrieval that fuses coarse-to-fine features via Matryoshka representation learning 📝arxiv.org/abs/2603.13349

AMES: Approximate Multi-modal Enterprise Search via Late Interaction Retrieval Apple presents a multimodal late interaction retrieval method deployed in Apache Solr, combining parallel token-level ANN candidate generation with Exact MaxSim re-ranking. 📝 arxiv.org/abs/2603.13537

Bringing Model Editing to Generative Recommendation in Cold-Start Scenarios Proposes a training-free model editing framework that injects cold-start item knowledge into generative recommendation models . 📝 arxiv.org/abs/2603.14259