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/…

Reinforced Preference Optimization for Reasoning-Augmented Recommendations Kuaishou integrates LLM chain-of-thought reasoning into recommendations via a dedicated retrieval head, using reinforcement learning to align reasoning quality. 📝 arxiv.org/abs/2605.21967

PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation TikTok presents a debiasing framework for recommendations that replaces raw engagement regression with contrastive percentile estimation. 📝 arxiv.org/abs/2605.21752

FLUID: From Ephemeral IDs to Multimodal Semantic Codes for Industrial-Scale Livestreaming Recommendation TikTok presents a livestreaming recommender that replaces item IDs with multimodal semantic codes generated by a cross-domain encoder. 📝 arxiv.org/abs/2605.21832

Bridging the Cold-Start Gap: LLM-Powered Synthetic Data Generation for Natural Language Search at Airbnb Airbnb combines contrastive listing pairs with seed queries to generate synthetic training data and relevance labels for natural language search. 📝 arxiv.org/abs/2605.21812

LLM Retrieval for Stable and Predictable Ad Recommendations Meta presents an LLM-based semantic candidate generation framework for ads retrieval that extracts hierarchical ad attributes and uses graph traversal to improve delivery performance. 📝 arxiv.org/abs/2605.21969

SGR-BENCH: Benchmarking Search Agents on State-Gated Retrieval Introduces a benchmark for evaluating search agents on specialized web retrieval tasks. 📝 arxiv.org/abs/2605.22219 👨🏽‍💻 huggingface.co/datasets/PKUAI…

Integrating Chain-of-Thought into Generative Retrieval: A Preliminary Study Interleaves chain-of-thought reasoning with document identifier generation in a single end-to-end pass for complex multi-hop retrieval. 📝 arxiv.org/abs/2605.22358

Search-E1: Self-Distillation Drives Self-Evolution in Search-Augmented Reasoning Proposes a self-evolution pipeline for search-augmented reasoning agents that alternates standard GRPO training with offline self-distillation. 📝 arxiv.org/abs/2605.22511

One prompt is not enough: Instruction Sensitivity Undermines Embedding Model Evaluation Shows that single-prompt evaluation of instruction-tuned embedding models is unreliable. 📝 arxiv.org/abs/2605.22544 👨🏽‍💻 github.com/centre-for-hum…

DeepWeb-Bench: A Deep Research Benchmark Demanding Massive Cross-Source Evidence and Long-Horizon Derivation Presents a deep research benchmark of 100 tasks requiring massive evidence collection, reconciliation, and derivation 📝arxiv.org/abs/2605.21482 👨🏽‍💻sixiongxie1001-dot.github.io/deep-research-…

DIVE: Embedding Compression via Self-Limiting Gradient Updates @realDonZhao introduces a compression adapter that combines a hinge-based triplet loss (zero gradient once margin is satisfied) with a head-wise NT-Xent contrastive loss. 📝 arxiv.org/abs/2605.20689

Layer-wise Token Compression for Efficient Document Reranking @ShengyaoZhuang et al. at Amazon propose applying token compression at intermediate transformer layers for cross-encoder rerankers, preserving early query-document interactions. 📝 arxiv.org/abs/2605.20683

Trust or Abstain? A Self-Aware RAG Approach Introduces a self-aware belief estimator that uses frozen LLM hidden states to independently assess parametric and contextual knowledge reliability. 📝 arxiv.org/abs/2605.18792 👨🏽‍💻 github.com/xizhu1022/SABER

ClusterRAG: Cluster-Based Collaborative Filtering for Personalized Retrieval-Augmented Generation Clusters users via HDBSCAN and retrieves documents from both the target user's profile and similar users' profiles. 📝 arxiv.org/abs/2605.18769 👨🏽‍💻 github.com/academicprojec…

LWGR: Lagrangian-Constrained Personalized World Knowledge for Generative Recommendation Alibaba transfers personalized world knowledge from LLMs into generative recommendation, using soft instructions and controllable knowledge fusion. 📝 arxiv.org/abs/2605.18771

Mask-to-Correct+: Leveraging Retriever Diversity for Masking-guided Faithful Fact Correction Introduces a training-free RAG framework that uses diversity-aware masking to identify erroneous spans in claims. 📝 arxiv.org/abs/2605.18776 👨🏽‍💻 github.com/payelsantra/Ma…

RecoAtlas: From Semantic Plausibility to Set-Level Utility in LLM Recommendation Agents @imad_aouali et al. at Criteo present a benchmark for evaluating LLM shopping agents on set-level objectives. 📝 arxiv.org/abs/2605.18805 👨🏽‍💻 github.com/imadaouali/rec… 🤗 huggingface.co/iaouali/reco-a…

The 99% Success Paradox: When Near-Perfect Retrieval Equals Random Selection Meta introduces a chance-corrected selectivity metric that reveals when high retrieval success rates actually perform no better than random selection. 📝 arxiv.org/abs/2605.18857

Understanding Wacky Weights: A Dissection of SPLADE's Learned Term Importance Formally defines "wacky weights" in SPLADE, finding that larger vocabularies increase their prevalence, stricter sparsity regularizers reduce it. 📝 arxiv.org/abs/2605.19628 👨🏽‍💻 github.com/polgrisha/unde…

Divergence Meets Consensus: A Multi-Source Negative Sampling Framework for Sequential Recommendation Introduces a "Teacher-Peer-Self" negative sampling framework inspired by Vygotsky's Zone of Proximal Development. 📝 arxiv.org/abs/2605.19651 👨🏽‍💻 github.com/Lyz103/SIGIR26…