Sumit

recsys.substack.comSearchJev: A Small Calibrated Model for the Routine Decisions in Search Agents, Training-Free Semantic IDs for Generative Recommendation, and More!Vol.177 for Oct 05 - Oct 11, 2026

I published Vol. 177 of "Top Information Retrieval Papers of the Week" on Substack. 🔗 recsys.substack.com/p/searchjev-...

arxiv.orgEmbeddingGemma 2: Unified and Efficient Multimodal RepresentationsWe present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space. Our approach introduces improvements ...

EmbeddingGemma 2: Unified and Efficient Multimodal Representations Google presents an open embedding model that maps text, images, audio, and video into one shared space, with modular encoders for smaller deployments. 📝 arxiv.org/abs/2610.37364 👨🏽‍💻 ai.google.dev/gemma/docs/e...

arxiv.orgJevRec: Linguistic Encoding with Jev-like Decision Modeling for LLM-based Generative RecommendationGenerative recommendation (GR) represents items as semantic identifiers (SIDs) and autoregressively generates the next item's SID from interaction history. However, most methods assign the same SID le...

10m

JevRec: Linguistic Encoding with Jev-like Decision Modeling for LLM-based Generative Recommendation Gives popular items shorter semantic IDs and long-tail items longer ones, and uses separate heads for predicting ID length and generating item codes. 📝 arxiv.org/abs/2610.36937

arxiv.orgMAREval: A Multi-Agent Framework for Evaluating Natural Language Recommendation ExplanationsEvaluating natural language explanations in recommender systems is essential for fostering user trust, transparency, and engagement. However, existing evaluation approaches like human evaluations, whi...

11m

MAREval: A Multi-Agent Framework for Evaluating Natural Language Recommendation Explanations Walmart proposes a multi-agent LLM framework with planner, moderator and arbitrator agents for evaluating recommendation explanations in line with human judgments. 📝 arxiv.org/abs/2610.36389

arxiv.orgQESChunker: A Single Objective Unifies Overlapping and Non-Overlapping Chunking for RAGChunking is a fundamental preprocessing step in retrieval-augmented generation (RAG), yet most existing chunkers rely primarily on heuristics. We introduce the Question--Evidence-Supervised Chunker (Q...

12m

QESChunker: A Single Objective Unifies Overlapping and Non-Overlapping Chunking for RAG Alibaba introduces a RAG chunker that sets boundaries from question-evidence pairs, using exact dynamic programming without overlap and greedy selection with overlap. 📝 arxiv.org/abs/2610.36557

arxiv.orgThe Vocabulary Is Already Here: Adapting Pretrained Interpretable Features for Generative RecommendationGenerative recommendation (GR) predicts the next item through its semantic ID (SID). Existing methods commonly fit SID vocabularies to pretrained language model (PLM) representations of item descripti...

12m

The Vocabulary Is Already Here: Adapting Pretrained Interpretable Features for Generative Recommendation Uses features from a pretrained sparse autoencoder as the item vocabulary for generative recommendation. 📝 arxiv.org/abs/2610.36763 👨🏽‍💻 anonymous.4open.science/r/API-C1F9/

arxiv.orgNatural Hard-Negative Mining from Positives for Instruction-Following RetrievalInstruction-following information retrieval (IFIR) requires retrievers to rank documents by both a query and its instance-specific instructions, yet instruction-conditioned contrastive supervision rem...

13m

Natural Hard-Negative Mining from Positives for Instruction-Following Retrieval Builds instruction-following training data by promoting a retrieved negative via a new instruction, with the original positive as a hard negative. 📝 arxiv.org/abs/2610.36900

arxiv.orgDo We Need an LLM to Know Why Sources Conflict? Decision Models for Knowledge-Conflict Typing in RAGWhen the sources retrieved for a query disagree, the right response depends on why they disagree: a question with several valid answers, a contested topic, stale information and a false source each ca...

13m

Do We Need an LLM to Know Why Sources Conflict? Decision Models for Knowledge-Conflict Typing in RAG Tests fast decision models for classifying why retrieved sources conflict. Most match Gemini 2.5 Flash at lower cost. 📝 arxiv.org/abs/2610.36915 👨🏽‍💻 github.com/Parv17k/deci...

arxiv.orgPailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal SearchGenerative retrieval models have been increasingly adopted in search and recommendation systems. However, industrial generative retrieval models are commonly specialized to a fixed online scenario: th...

14m

PailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search Alibaba trains one vision-language model to map an image and a free-form request to product IDs, with adjustable reasoning budgets. 📝 arxiv.org/abs/2610.36958

arxiv.orgKL-MOPD: Discrepancy-Aware Multi-Objective On-Policy Distillation for Generative RecommendationCoordinating multiple recommendation objectives (e.g., click and conversion) is critical for real-world generative recommendation (GR). However, since a positive item in GR can be retrieved only when ...

15m

KL-MOPD: Discrepancy-Aware Multi-Objective On-Policy Distillation for Generative Recommendation Weights each objective's teacher per token by its KL divergence from the student when distilling a generative recommender. 📝 arxiv.org/abs/2610.36973 👨🏽‍💻 github.com/Yandorn/KL-M...

arxiv.orgCompote: Collapse-Resistant Multiple Hypothesis Learning For Sequential RecommendationNext-item prediction is inherently ambiguous: a single interaction history often admits multiple plausible continuations across distinct catalogue regions. Standard sequential recommenders compress th...

15m

Compote: Collapse-Resistant Multiple Hypothesis Learning For Sequential Recommendation Trains a multi-head sequential recommender with an annealed winner-takes-all objective, so each head covers a different part of the catalogue instead of collapsing. 📝 arxiv.org/abs/2610.37005

arxiv.orgTessera: Structure-Completed Evidence Units for Web Retrieval-Augmented GenerationWeb retrieval-augmented generation (RAG) splits a page into fixed-size text chunks or, in recent visual pipelines, screenshot tiles, and the retrieved item serves at once as the unit of indexing, rank...

16m

Tessera: Structure-Completed Evidence Units for Web Retrieval-Augmented Generation Introduces a training-free layer that rebuilds retrieved pages into small evidence units that keep their titles and table headers. 📝 arxiv.org/abs/2610.37080 👨🏽‍💻 github.com/sarendis56/T...

arxiv.orgRethinking Positional Encoding and Interaction Granularity for CTR PredictionClick-through rate (CTR) prediction relies on historical behaviors and contextual features to estimate clicks on a target item. Existing Transformer-based CTR models often adopt index-based or tempora...

16m

Rethinking Positional Encoding and Interaction Granularity for CTR Prediction Proposes a CTR model that encodes each behavior's sequence rank and its time relative to the target, and lets behaviors attend to each other within temporal groups. 📝 arxiv.org/abs/2610.37140

arxiv.orgCoverage Under Contradiction: A Benchmark for Negation-Aware RetrievalCoverage-evaluation benchmarks quantify whether retrieved evidence covers the different pieces of information needed to satisfy an information need. Existing benchmarks mostly consider factual claims ...

17m

Coverage Under Contradiction: A Benchmark for Negation-Aware Retrieval Introduces a benchmark that edits documents to support negated claims, testing how retrievers handle evidence that conflicts with prior knowledge. 📝 arxiv.org/abs/2610.37162 🤗 huggingface.co/datasets/pas...

arxiv.orgLossless Compression of Proximity Graph Indexes for Efficient Approximate Nearest Neighbors SearchGraph-based Approximate Nearest Neighbor (ANN) indexes, such as HNSW, are the backbone of modern vector search. However, their massive memory footprint creates a severe infrastructure bottleneck. Whil...

18m

Lossless Compression of Proximity Graph Indexes for Efficient Approximate Nearest Neighbors Search Compresses HNSW edge lists by reordering node IDs offline and using fast SIMD integer codecs, cutting memory with negligible effect on query latency. 📝 arxiv.org/abs/2610.37176

arxiv.orgAutoRSE: Evaluating LLM Agents as Recommendation Engineers for Self-Evolving Recommender SystemsRecommender Systems Engineer Agents (RSE-Agents), which use LLM-based agents to autonomously improve recommendation algorithms, are emerging as a promising paradigm. However, their capabilities and li...

19m

AutoRSE: Evaluating LLM Agents as Recommendation Engineers for Self-Evolving Recommender Systems Benchmarks LLM agents on six recommendation tasks. Their gains come mostly from tuning and ensembling, not new algorithms. 📝 arxiv.org/abs/2610.37228 👨🏽‍💻 github.com/Ray-Tiann/Au...

arxiv.orgLESSRec: Level-Specific History Modeling for Semantic ID-Based RecommendationSemantic ID-based recommendation represents each item as a sequence of semantic tokens obtained through residual quantization. This paradigm generates the next item's tokens autoregressively from coar...

19m

LESSRec: Level-Specific History Modeling for Semantic ID-Based Recommendation Finds that history tokens at the target token's semantic level are more informative. Proposes encoding each level separately & limiting attention to the matching level. 📝 arxiv.org/abs/2610.37286

arxiv.orgREGER: Modeling inter-item Relations with Semantic IDs for Generative RecommendationPersonalized recommendation depends on learning semantic and collaborative relations among items from online content and user interactions. Generative recommendation (GR) offers a promising paradigm b...

21m

REGER: Modeling inter-item Relations with Semantic IDs for Generative Recommendation Adds explicit item relations to generative recommendation through semantic pretraining, adjacent-item attention, and distillation from a discriminative teacher. 📝 arxiv.org/abs/2610.37584

doi.orgEffBench: A Holistic Efficiency Benchmark for CTR Models | ACM Transactions on Information SystemsClick-Through Rate (CTR) prediction is a critical task in recommender systems and online advertising, so CTR models design has been widely researched in both academia and industry. However, since the model parameters increase larger and their computations ...

EffBench: A Holistic Efficiency Benchmark for CTR Models Introduces a benchmark that compares CTR models on efficiency as well as accuracy, with latency and throughput metrics and a profiling tool that finds bottlenecks. 📝 doi.org/10.1145/3857... 👨🏽‍💻 github.com/Yeedzhi/EffB...

doi.orgER-RAG: Enhance RAG with ER-Based Unified Modeling of Heterogeneous Data Source | ACM Transactions on Knowledge Discovery from DataRetrieval-Augmented Generation (RAG) enhances the reliability of large language models (LLMs) by integrating external evidence for question-answering tasks. However, when supporting evidence is scattered across heterogeneous data sources (e.g., web pages, ...

ER-RAG: Enhance RAG with ER-Based Unified Modeling of Heterogeneous Data Source Introduces a RAG framework that retrieves from web pages, databases and knowledge graphs through one shared API, with a trained module that picks which sources to query. 📝 doi.org/10.1145/3860...

github.comGitHub - EthanNing/SearchHorizonBench: A benchmark for search value in AI agents. Paired execution, gain prediction, and cost-aware decisions across 45,632 executions.A benchmark for search value in AI agents. Paired execution, gain prediction, and cost-aware decisions across 45,632 executions. - EthanNing/SearchHorizonBench

SearchHorizonBench: A Benchmark for Search Value in AI Agents Introduces a benchmark for judging when an extra web search is worth its cost for an AI agent, using paired runs that either answer directly or search further from the same point. 📝 arxiv.org/abs/2610.34714 👨🏽‍💻 github.com/EthanNing/Se...

arxiv.orgMuSID-KD: Multigranularity Knowledge Distillation for Generative Recommendation with Semantic IDsGenerative recommendation with semantic IDs (SIDs) represents each item as a sequence of discrete tokens and autoregressively predicts the next item's SID. Despite strong recommendation performance, r...

MuSID-KD: Multigranularity Knowledge Distillation for Generative Recommendation with Semantic IDs Improves distillation of semantic ID recommenders by aligning candidate group distributions at each ID level. 📝 arxiv.org/abs/2610.35218 👨🏽‍💻 anonymous.4open.science/r/m7k2-revie...

arxiv.orgCan System-1-Style Decision Models Replace Dedicated Rerankers in RAG? A Systems EvaluationRetrieval-augmented generation (RAG) systems often use a dedicated reranker to choose which retrieved candidates reach the language model. Reranking improves relevance but adds latency, so teams look ...

Can System-1-Style Decision Models Replace Dedicated Rerankers in RAG? A Systems Evaluation Tests decision models as reranker replacements. Truncation makes some competitive, but a tuned dedicated reranker is still faster and uses far less GPU memory. 📝 arxiv.org/abs/2610.34816

arxiv.orgMemory Tax: Dissecting the Net Value of Agent MemoryMemory is increasingly used in long-horizon agents, yet the net value of building memory over direct retrieval from raw history remains largely unclear. Through a large-scale evaluation across multipl...

Memory Tax: Dissecting the Net Value of Agent Memory Finds that many agent memory systems score below naive RAG over raw history. Splits memory's net value into benefit and tax. 📝 arxiv.org/abs/2610.34073 👨🏽‍💻 hexuan21.github.io/MemoryTax 🤗 huggingface.co/datasets/hex...

arxiv.orgPinGER: Token-Efficient Generative Retrieval with Semantic IDsGenerative retrieval replaces nearest-neighbor search with token-by-token decoding of an item's semantic ID (SID). But existing SID-based methods primarily use L levels of tokens sequentially to repre...

PinGER: Token-Efficient Generative Retrieval with Semantic IDs Pinterest presents its generative retriever that replaces nearest-neighbor search. An asymmetric attention mask trains with semantic ID supervision but serves only one token per item. 📝 arxiv.org/abs/2610.33435

arxiv.orgRethinking Retrieval in the Long-Context Era: Training-Free Semantic Trees for Adaptive Context ConstructionLong-context language models can increasingly process entire documents, making full-context prompting a practical alternative to passage retrieval. However, greater capacity determines how much inform...

Rethinking Retrieval in the Long-Context Era: Training-Free Semantic Trees for Adaptive Context Construction Builds a training-free semantic tree of a long document and gives the LLM only relevant sections. 📝 arxiv.org/abs/2610.34146 👨🏽‍💻 anonymous.4open.science/r/treefuse-E...

arxiv.orgLoopGR: Parameter-Efficient Scaling for Generative Recommendation via Recurrent DepthGenerative recommendation formulates item recommendation as autoregressive generation rather than a discriminative scoring problem. Although empirical studies exhibit favorable scaling behavior in rec...

LoopGR: Parameter-Efficient Scaling for Generative Recommendation via Recurrent Depth Reuses a shared decoder and a light context updater in recurrent loops to add depth without adding parameters, then distills it into a cheaper low-loop student. 📝 arxiv.org/abs/2610.34157

arxiv.orgTheoretically Sound Item Similarity Measures for Recommender SystemsQuantifying the similarity between items is fundamental to neighborhood-based collaborative filtering, yet many commonly used similarity measures are heuristics that lack theoretical justification. Th...

Theoretically Sound Item Similarity Measures for Recommender Systems Derives item similarity measures from Bernoulli and multinomial Naive Bayes, and shows cosine similarity solves a specific optimization problem. 📝 arxiv.org/abs/2610.34402 👨🏽‍💻 github.com/noah-daniels...

arxiv.orgKnowing When the Evidence Is Enough: Requirement-Directed Search with Horn-Closure CertificatesAfter each Web search, a search-augmented model must decide whether its evidence suffices: answer, search again, or abstain. Prevailing systems leave this to holistic model judgement, which can accept...

Knowing When the Evidence Is Enough: Requirement-Directed Search with Horn-Closure Certificates Decides whether to answer, search again, or abstain using a per-query requirement hypergraph and Horn closure instead of having a model judge the evidence. 📝 arxiv.org/abs/2610.34684

arxiv.orgSLIDE: Search-to-List Item Distillation for Single-Rollout Generative RetrievalGenerative retrieval (GR) typically relies on item-level beam search to retrieve candidates, leaving cross-item dependencies unmodeled and limiting direct optimization of list-level objectives. A natu...

SLIDE: Search-to-List Item Distillation for Single-Rollout Generative Retrieval Distills a teacher model's offline beam-search results into list-wise targets, so a smaller generative retrieval model can produce a full ranked list in one greedy pass. 📝 arxiv.org/abs/2610.34764

arxiv.orgOcten-WebGR: Scaling Generative Retrieval to 200 Million Web PagesNavigational search aims to find a single target page among hundreds of millions of web pages. Click tables work well for queries seen before but cannot generalise to new formulations, while embedding...

Octen-WebGR: Scaling Generative Retrieval to 200 Million Web Pages Introduces a generative retriever that indexes 201 million web pages with two-token identifiers for navigational search. 📝 arxiv.org/abs/2610.35256 👨🏽‍💻 github.com/Octen-Team 🤗 huggingface.co/Octen

arxiv.orgUnified Heterogeneous Feature Interaction via Effective-Rank Scaling for Recommender SystemsLarge-scale industrial ranking models must jointly capture heterogeneous feature interactions, long-term user behaviors, and multi-scenario dependencies under stringent online serving constraints. Exi...

Unified Heterogeneous Feature Interaction via Effective-Rank Scaling for Recommender Systems Introduces a ranking model that combines behavior sequence modeling and feature interaction in a single backbone for multi-scenario recommendation. 📝 arxiv.org/abs/2610.35318

arxiv.orgTowards a Unified Framework of Optimal Input-Side Dimensional Configuration for Deep CTR PredictionDeep click-through rate (CTR) prediction has achieved substantial success through increasingly expressive feature interaction architectures. However, predictive performance also depends on how input-s...

Towards a Unified Framework of Optimal Input-Side Dimensional Configuration for Deep CTR Prediction Introduces a framework that jointly selects samples, features, and embedding units for CTR prediction, using optimal transport to align the levels. 📝 arxiv.org/abs/2610.35695

Do Similarity Measures Measure Similarity? How User Activity Creates Phantom Correlation in Recommender Systems Shows that differences in user activity alone make items look similar, even when users share no taste. 📝 arxiv.org/abs/2610.35765 👨🏽‍💻 github.com/Mfavier91/Ph...

arxiv.orgACARE: Agent-Context-Aware Reranker for Iterative SearchIn agentic search, a document can match an agent-generated sub-query without helping answer the user's original question. Conventional rerankers score documents against the sub-query alone, overlookin...

ACARe: Agent-Context-Aware Reranker for Iterative Search Introduces a compact reranker for search agents that scores documents using the original question and reasoning trace, not just the agent's sub-query, improving retrieval and cutting searches. 📝 arxiv.org/abs/2610.35840

arxiv.orgRetriever-Free Retrieval: Navigating Documents with Decision ModelsWhile language models have become dramatically more capable, retrieval still largely relies on lexical matching or similarity in fixed embedding spaces. This creates a capability asymmetry: what the r...

Retriever-Free Retrieval: Navigating Documents with Decision Models Retrieves by walking down a corpus hierarchy, with a decision model scoring all candidates at each level in a single pass, without generating tokens. 📝 arxiv.org/abs/2610.35868

arxiv.orgBeyond Post-Hoc Rewards: Privileged Reasoning Exploration for Multimodal EmbeddingsAdvances in Multimodal Large Language Models (MLLMs) have driven substantial progress in multimodal embeddings. Recent methods exploit generative capabilities by augmenting embedded inputs with self-g...

Beyond Post-Hoc Rewards: Privileged Reasoning Exploration for Multimodal Embeddings Uses training-time positive and negative candidates to guide reasoning exploration for multimodal embeddings, with input-only inference. 📝 arxiv.org/abs/2610.32474

arxiv.orgTGR-EvoDAG: A DAG-based Multi-Agent Code Evolution Framework for Generative Recommendation on TencentLarge Language Model (LLM) agents have enabled automated code evolution for algorithm discovery. However, current recursive self-improvement (RSI) research for sequential recommendation faces a dilemm...

TGR-EvoDAG: A DAG-based Multi-Agent Code Evolution Framework for Generative Recommendation on Tencent Tencent introduces a multi-agent framework that evolves generative recommendation code as a DAG of branches, with GPU cost planning for each pipeline stage. 📝 arxiv.org/abs/2610.32326

arxiv.orgMICE to Meet You: From Re-ranking Only to End-to-End RetrievalLate-interaction models offer state-of-the-art efficiency-effectiveness trade-offs for neural retrieval by enabling document representations to be pre-computed and efficiently indexed. Yet, their rela...

MICE to Meet You: From Re-ranking Only to End-to-End Retrieval Introduces MICE-P, which indexes MICE document representations with PLAID to build an end-to-end retrieval pipeline with cross-encoder-like effectiveness. 📝 arxiv.org/abs/2610.31932 👨🏽‍💻 github.com/xpmir/mice

arxiv.orgFrom Item Scores to Semantic Tokens: Test-Time Collaborative Adaptation for Generative RecommendationGenerative recommenders using semantic identifiers predict items through successive token decisions, but beam search may irreversibly prune promising item paths. This limitation is particularly conseq...

From Item Scores to Semantic Tokens: Test-Time Collaborative Adaptation for Generative Recommendation Introduces a training-free way to use discriminative item scores to guide beam search in a frozen generative recommender before prefixes get pruned. 📝 arxiv.org/abs/2610.32098

arxiv.orgPersonalQuery-Benchmark:Evaluating Retrieval Robustness to Personalized User Expression DifferencesExisting personalized information retrieval methods mainly model differences in user needs through user preferences and query intent, while paying little attention to the personalized expression diffe...

PersonalQuery-Benchmark: Evaluating Retrieval Robustness to Personalized User Expression Differences Introduces a benchmark where personalized query syntax and writing errors shift rankings for a fixed information need. 📝 arxiv.org/abs/2610.32164 👨🏽‍💻 github.com/cpxjazz123/P...

arxiv.orgSystem One models Are Not Better RerankersSystem One models, or decision models, answer a question about an input with a calibrated probability instead of generated text. Within weeks of the first release, in September 2026, they were used as...

System One Models Are Not Better Rerankers Finds that decision (System One) fine-tuning never beats plain reranker fine-tuning on the same Qwen backbones, and its calibration advantage depends on the question wording. 📝 arxiv.org/abs/2610.32318 👨🏽‍💻 github.com/KanishkaGD/d...

arxiv.orgAgentic Memory Management Is the Next Frontier of Information RetrievalInformation retrieval has long been focusing on connecting people with information they need to complete their tasks. Recently, the meaning of this has changed dramatically as more and more AI agents ...

Agentic Memory Management Is the Next Frontier of Information Retrieval Presents a perspective paper arguing that managing an AI agent's memory is a natural extension of IR. Also shares a memory framework, a comparison with classic IR, and a research agenda. 📝 arxiv.org/abs/2610.32464

Knowledge Distillation across Item and Code Spaces for Parallel Generative Recommendation Proposes two objectives that distill a sequential recommender's item scores into a parallel semantic ID model, matching item distributions or code marginals. 📝 arxiv.org/abs/2610.32943

arxiv.orgKnowledge Distillation across Item and Code Spaces for Parallel Generative RecommendationGenerative recommendation with semantic IDs represents each item as a sequence of discrete codes derived from its semantic representation and predicts these codes to generate recommendations. Existing...

arxiv.orgSeparating Thinking from Spelling for Efficient Generative RecommendationGenerative recommenders retrieve items by writing their semantic IDs (SIDs) code by code under trie-constrained beam search, an approach now deployed on large web platforms. Standard autoregressive in...

Separating Thinking from Spelling for Efficient Generative Recommendation Presents a GR system where the backbone reads the user history once into a compact concept and a light decoder runs the entire semantic ID beam search, which cuts serving cost. 📝 arxiv.org/abs/2610.32980

amazon.scienceSafe and sufficient caching: Per-request validated evidence reuse for production retrieval-augmented generationProduction Retrieval-Augmented Generation (RAG) systems serve repeated queries over evolving document corpora under user-specific access constraints. Standard pipelines perform full retrieval and gene...

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 and retrieval cost with zero invalid reuse. 📝 www.amazon.science/publications...

arxiv.orgTransX Technical Report: Scaling Sequence Transduction for RecommendationWe present TransX, a sequence-transduction framework and co-designed infrastructure for scaling industrial recommendation with explicit control over model-compute allocation and execution amortization...

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

arxiv.orgLearning When to Think: Utility-Guided Adaptive Reasoning for Efficient Text RetrievalReasoning-conditioned embeddings can improve retrieval by progressively refining semantic constraints. However, additional reasoning may introduce over-generalized or semantically noisy abstractions t...

Learning When to Think: Utility-Guided Adaptive Reasoning for Efficient Text Retrieval Trains a retrieval model to reason step by step and stop once extra reasoning stops helping, cutting generation cost and latency. 📝 arxiv.org/abs/2610.21030 👨🏽‍💻 github.com/yingtaoluo/A...

arxiv.orgBeyond Retrieval Failure: Understanding Why LLMs Miss Evidence in Multi-Hop RAGMulti-hop retrieval-augmented generation (RAG) requires an agent to retrieve evidence for a sequence of dependent sub-questions. When a wrong answer lacks a required passage, the passage may be inacce...

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