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/20…

Safe and Sufficient Caching: Per-Request Validated Evidence Reuse for Production Retrieval-Augmented Generation Amazon caches evidence chunks and checks access and source freshness before reuse, cutting latency & retrieval cost with zero invalid reuse. 📝amazon.science/publications/…

TransX Technical Report: Scaling Sequence Transduction for Recommendation LinkedIn presents a recommender that separates history encoding, behavior memory & candidate scoring so most work runs offline, letting models scale without raising serving cost. 📝arxiv.org/abs/2610.20913

Learning When to Think: Utility-Guided Adaptive Reasoning for Efficient Text Retrieval @yingtaoluo et al. train a retrieval model to stop reasoning once extra reasoning stops helping, cutting generation cost and latency. 📝 arxiv.org/abs/2610.21030 👨🏽‍💻 github.com/yingtaoluo/Adapti…

Beyond Retrieval Failure: Understanding Why LLMs Miss Evidence in Multi-Hop RAG Finds that missing evidence in multi-hop RAG often stems from search drift, not retriever limits. 📝 arxiv.org/abs/2610.21184

RecSoup: Efficient Model Merging Approach for Controllable Multi-Objective Recommendation Merges objective-specific recommenders after pruning conflicting parameters, so trade-offs can be tuned without retraining. 📝 arxiv.org/abs/2610.21220 👨🏽‍💻 github.com/35798642/RECSOUP.…

Aligning LLMs with Sparse Retrieval via Adaptive Sparse Supervision Aligns an LLM with SPLADE's vocabulary space using input-adaptive truncation, keeping inverted-index retrieval with much lower storage and latency than prior LLM sparse retrievers. 📝 arxiv.org/abs/2610.21877

CASI: Credit Assignment via Set Intervention for Agentic Retrieval Xiaohongshu proposes assigning credit to jointly generated retrieval actions under one shared reward by scoring subsets of cached retrievals with Shapley values. 📝 arxiv.org/abs/2610.20293

Revisiting Generative vs. Dense Retrieval: A Static Value Prior Reverses the Gap Shopify finds that generative retrieval's item-value lead in e-commerce comes largely from popularity. 📝 arxiv.org/abs/2610.19065 👨🏽‍💻 anonymous.4open.science/r/gr…

ListQA: A Benchmark for Evaluating List-Formatted Factual Knowledge Retrieval in Large Language Models @mayank_k91 et al. present a benchmark where LLMs recall facts from Wikipedia tables and return them as structured lists. 📝 arxiv.org/abs/2610.19165 👨🏽‍💻 listqa.github.io

RCER: Reconstructing Cross-Evidence Relations for Iterative Retrieval-Augmented Generation Introduces an iterative RAG method that reconstructs relations across retrieved passages to guide retrieval, select complementary passages, and compose answers. 📝 arxiv.org/abs/2610.19398

Retrieve What Helps, Only When It Helps: Agentic Reinforcement Learning for Selective Retrieval-Augmented Machine Translation Uses RL to teach an LLM when to retrieve for translation, and trains the retriever on translation quality, not similarity. 📝 arxiv.org/abs/2610.19528

Cascaded Retrieval with Cross-Modal Query Refinement for Efficient Visual Document Search Introduces a cascaded framework that cuts latency in visual document retrieval by reranking a small candidate pool and refining the query with offline summaries. 📝 arxiv.org/abs/2610.20367

Configuring RAG Systems: What Breaks, Where Originates, and How to Fix Presents an empirical study of configuration issues in open-source RAG systems, categorizing symptoms, causes and fix patterns across the pipeline stages. 📝arxiv.org/abs/2610.18567 👨🏽‍💻github.com/ideas-labo/rag-co…

Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders Introduces a system which rewards recommender lists against all responders' titles at once, making GRPO training cheaper. 📝arxiv.org/abs/2610.17771 👨🏽‍💻lixin.ai/ReMaRk

DRanker: A Transformer-Based Multi-Task Ranking Model for the Disney+ Homepage Disney describes a transformer ranker for the Disney+ homepage with a lightweight causal debiasing tower for popularity bias. 📝arxiv.org/abs/2610.15696

Personalized LLM Recommendation with Semantic IDs: Autoregressive Beam Search vs. Multi-Token Prediction at Industry Scale Netflix compares beam search and multi-token prediction for generating Semantic IDs in LLM recommenders. 📝 arxiv.org/abs/2610.15864

UniSGR-V2: Stepwise Target-Side Feature Anticipation and Crossing for Generative Recommendation Alibaba presents a system that predicts continuous target-item features at each semantic-ID decoding step & uses them for both code generation and ranking. 📝 arxiv.org/abs/2610.16132

Label-Centric Knowledge Retrieval for Zero-Shot Text Classification Presents a system where a large LLM builds label definitions and boundary rules offline, and a small LLM classifies zero-shot by retrieving that knowledge instead of labeled examples. 📝 arxiv.org/abs/2610.16182

From a Large-Scale Teacher to Efficient Multimodal Students for Search Relevance Prediction eBay distills a GPT-4.1 relevance judge into smaller multimodal Gemma-3 models for e-commerce search, keeping near-teacher quality at much lower cost & latency. 📝arxiv.org/abs/2610.16569

KNN Implementation Details Can Dramatically Change Performance: An Example from Cover Trees Shows how query mode, base parameter, and micro-optimizations affect Cover Tree speed. 📝 arxiv.org/abs/2610.16940 👨🏽‍💻 github.com/CrowdStrike/RustK…