跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+979新增论文
+781其中 AI 相关
+119已有论文修订
+95新增 AI 导读

arXiv 最近完成扫描:2026-11-20 08:48 · 全站最近入库:2026-11-20 19:58各信源进度 →

统计口径与数据新鲜度

统计 2026-11-20 06:40 — 2026-11-21 06:40(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-20 08:48最近发现新增版本:2026-11-20 19:58库内论文最新更新时间:2026-11-20 02:59

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

28,961 已收录论文24,030 AI 相关8,912 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 545 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-20 08:46(北京时间)
新论文 · v1arXiv:2611.20702

When the Bill Falls: Reusing the Old Index Defers,Never Avoids, Re-encoding in Complete Embedding-Model Migration

Nannan Zhao, Jiameng Zhang, Taoyu Zhong, Hao Wang, Bo Wang

原文摘要

Upgrading the embedding model behind a vector index requires re-encoding the corpus. Recent work reuses the old index through adapters or window re-ranking. Across sixteen model pairs on passage and code corpora, plus eleven on Natural Questions (NQ), we find three results. First, reuse never avoids re-encoding: if migration must complete, every strategy eventually encodes every document; reuse only changes when that work is paid and the quality served meanwhile. Second, adapters should be compared with the old model, not only the new one. Of five adapter variants, only same-family LLM bridges beat the old model, and only by a negligible margin. Third, window re-ranking cost can be priced from the old index alone; pre-registered replays on all 8,841,823 passages confirmed the prediction on all three tested pairs. Workload skew also matters: across 36 further pre-registered Zipf replays, the planned family matched on 32. These findings yield three pre-migration rules: whether reuse is worthwhile, which strategy to use, and how deep a window to take. The depth rule priced every window that reached target quality to within 4 parts in 10^5 of the corpus. On NQ, a pre-committed plan selected the family on 10 of 11 pairs and priced every window exactly. Quality inputs transferred less reliably across corpus scale, so the model gap and window quality loss must be measured on the serving index.

cs.IR · cs.CL

版本更新 · v2arXiv:2606.27058

UniFormer: Efficient and Unified Model-Centric Scaling for Industrial Recommendation

Bo Chen, Jinlong Jiao, Tijian Hu, Ruihao Zhang, Yanzhi Liu, Chenghou Jin, Qinglin Jia, Baixuan He 等 16 位作者

原文摘要

Recently, substantial progress has been made in industrial recommendation through component-centric model scaling, where individual components such as behavior modeling, feature interaction, or task modeling are independently scaled to improve model capacity. Although recent methods such as HyFormer and OneTrans further explore cross-module co-scaling by jointly modeling behavior and interaction, their designs are still confined to the feature space and lack a unified model-centric scaling framework over the overall modeling space. In this paper, we propose UniFormer, an efficient and unified model-centric scaling framework for industrial recommender systems. To improve efficiency, UniFormer decomposes the overall modeling space into feature and task spaces, which are modeled by stacked Feature-space Interaction Modules and Task-space Interaction Modules, respectively. Moreover, UniFormer introduces semantic-based tokenization scheme to enable user-item decoupling, thereby achieving request-level inference acceleration. To prevent preference collapse, UniFormer employs multi-sequence cross-attention to separately capture heterogeneous behavior patterns, followed by the self-attention to enhance interaction modeling. Besides, dedicated multi-view FFNs are introduced to support flexible and scalable parameter scaling across different modeling components. Extensive online A/B testing in two production scenarios, Kuaishou and Kuaishou Lite, shows that UniFormer consistently improves user engagement and interaction metrics, achieving gains of +0.101%/+0.260% in App Stay Time and +0.729%/+1.113% in Watch Time, respectively.

cs.IR

版本更新 · v2arXiv:2607.24402

CogRec: Structure-Cognitive Fast-and-Slow Reasoning for Generative Recommendation

Xiang Liu, Jingsong Su, Shuqi Zhao, Pengbo Mo, Yiming Qiu, Huimu Wang, Mingming Li, Jiao Dai 等 10 位作者

原文摘要

Generative recommendation reformulates next-item prediction as identifier sequence generation, yet existing methods mainly treat the generated identifiers as target sequences to be memorized, leaving their hierarchical and relational structure underused for explicit reasoning. We propose CogRec, a structure-cognitive fast-and-slow reasoning framework that grounds intermediate reasoning in the same SID topology used for target generation. CogRec augments the vertical SID hierarchy with intra-layer semantic graphs and introduces SID Routing to represent recommendation reasoning through layer-wise Match, LateralJump, and Explore operations. These operations provide a structural interpretation of fast localization and slow navigation. A supervised multi-stage pipeline aligns newly introduced SID tokens, establishes direct generation, and trains natural-language and SID Routing branches from a shared checkpoint. Experiments on three Amazon Reviews 2014 categories show dataset-dependent effects of SID Routing. Routing improves Hit@10 and NDCG@10 over its direct-generation counterpart on Beauty and Sports, while reducing both metrics on Toys. These results characterize the benefits and dataset-dependent behavior of SID-grounded intermediate reasoning for generative recommendation. Code is available at https://github.com/caskcsg/CogRec.

cs.IR

版本更新 · v3arXiv:2603.22916

GateSID: One Gate to Balance Semantic and Collaborative Signals across the Item Lifecycle

Hai Zhu, Yantao Yu, Lei Shen, Bing Wang, Xiaoyi Zeng

原文摘要

In recommender systems, collaborative signals work well for popular items but are nearly absent for cold-start ones. Semantic content helps new items, yet over-reliance on it blurs the behavioral distinctions of popular items. Existing Semantic ID (SID) based methods handle this trade-off only partially, because their fusion of the two signals is rigid and item-agnostic. We present GateSID, a ranking framework that takes SID from an upstream tokenizer and adapts how each SID is used to the item's maturity through a single shared gate. The gate drives three mechanisms. Gating-Fused Shared Attention (GFSA) fuses the attention distributions of the two modalities rather than their representations. Gate-Regulated Contrastive Alignment (GRCA) strengthens semantic-behavioral alignment for items with weak collaborative evidence and relaxes it when collaborative structure is reliable. Gate-Regulated Semantic Resolution (GRSR) uses coarse shared SID prefixes for items with weak behavioral evidence and opens deeper SID levels when the gate indicates more reliable collaborative evidence. On a large-scale industrial dataset, GateSID leads the strongest SID-enhanced baseline by +0.27% CTR AUC and +0.36% CTCVR GAUC, with gains balanced across item groups rather than trading cold-start for popular-item performance; on the public Taobao-MM benchmark, it also leads overall and gives the strongest new-item results under both DIN and RankMixer backbones. In a one-month online A/B test, GMV improves by 2.6%, CTR by 1.1%, and Order by 1.6%, at a cost of less than 5 ms of additional latency. We also provide a systematic analysis of SID design choices, covering embedding type, codebook configuration, and fusion strategy.

cs.IR

版本更新 · v4arXiv:2602.23530

Unified Ranking Across Retrieval Channels: A Large-Scale E-Commerce Approach

Aditya Gaydhani, Guangyue Xu, Dhanush Kamath, Ankit Singh, Alex Fengjie Li

原文摘要

Large-scale e-commerce search and recommendation systems rely on multiple specialized retrieval channels to surface diverse products, each optimized for a distinct objective. A key challenge is merging candidates from these heterogeneous channels into a single ranked list under strict latency constraints while optimizing business KPIs such as user conversion. Static rank-fusion methods such as Reciprocal Rank Fusion (RRF) and Weighted Interleaving use the same fusion behavior across queries, limiting their ability to adapt to context or model interactions among channel signals. We reformulate multi-channel fusion as a channel-aware learning-to-rank problem and propose a unified ranking model trained on a scalar, funnel-weighted graded-relevance label that combines clicks, add-to-carts, and purchases. The model incorporates channel-specific scores together with recent aggregate query-item behavior to capture near-term population demand. The evaluated channel set combines conventional relevance retrieval with query-conditioned, recommendation-oriented channels emphasizing newness, trending demand, regional relevance, and seasonality. A randomized online experiment shows a 2.85% increase in user conversion over the deployed weighted-interleaving control. The model achieves a p95 reranking-stage latency below 50 ms and has been deployed on Target.com.

cs.IR

版本更新 · v4arXiv:2602.02343

Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics

Ziwen Xu, Chenyan Wu, Hengyu Sun, Haiwen Hong, Mengru Wang, Yunzhi Yao, Longtao Huang, Hui Xue 等 12 位作者

原文摘要

Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model's valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md.

cs.CL · cs.AI · cs.CV · cs.IR · cs.LG

版本更新 · v3arXiv:2603.04836

Beyond Text: Text-Image Fusion for Scalable Two-Tower Retrieval

Qujiaheng Zhang, Guangyue Xu, Fengjie Li

原文摘要

E-commerce customers make purchase decisions by jointly considering product text and visual information, yet most industrial retrieval systems rely almost exclusively on textual signals. Deployed at scale on CPU-based nearest-neighbor infrastructure, these systems leave substantial visual evidence untapped. In this work, we study unified text-image fusion for two-tower retrieval models at a large-scale e-commerce platform, training on 20M query-item pairs. We demonstrate that domain-specific fine-tuning and modality-specific query alignment between user queries and both product text and image modalities are crucial for effective multimodal retrieval. Building on these insights, we propose a mixture-of-modality-experts fusion architecture with a bilinear interaction network to capture cross-modal complementary information. Our model achieves up to 4.86% improvement in nDCG@1 for desirability and 2.36% for relevance over a strong text-only baseline, while preserving the pre-computable item embedding design required for production deployment.

cs.IR

新论文 · v1arXiv:2611.20449

MAGE: Meta-path Adaptive Graph Evidence for Transferable Multimodal Graph Learning

Haiyang Jiang, Tong Chen, Xinyi Gao, Hao Wang, Hongzhi Yin

原文摘要

Multimodal graphs integrate textual and visual attributes with relational structures, providing a rich representation of complex real-world data. To enable the broad applicability and reuse of graph models, it is crucial for models trained on source graphs to generalize to unseen graph domains. However, this is challenging because cross-modal interactions among neighbors and the neighboring information needed for prediction can vary across target nodes and graph domains. This gives rise to two key challenges: adaptive cross-modal neighborhood fusion, which captures and integrates cross-modal interactions among neighbors to accommodate variations across nodes and domains; and target-specific evidence selection, which identifies the neighboring textual and visual information needed for each target node to support LLM-based prediction. In light of these challenges, we propose MAGE (Meta-path Adaptive Graph Evidence Prompting), a framework for transferable multimodal graph learning. MAGE represents textual and visual node views separately in a heterogeneous graph and introduces two complementary components. Meta-path Adaptive Multimodal Propagation (MAMP) adaptively aggregates neighborhood information along heterogeneous meta-paths and compresses it into graph-aware soft tokens. Evidence-Guided Instruction Tuning (EGIT) selects relevant neighboring evidence for each query and incorporates its graph-aware representations into instructions for a frozen multimodal LLM. Together, these components support transfer from multiple source graphs to unseen target graphs without target-domain parameter updates. Experiments with three source graphs and four unseen target graphs demonstrate strong cross-domain transfer performance on node classification and link prediction, without requiring parameter updates on the target domains.

cs.IR

新论文 · v1arXiv:2611.20443

Overview of the Plagiarism Detection Task at PAN 2026

André Greiner-Petter, Yun-Ang Wu, Yuka Kitamura, Kon Woo Kim, Maik Fröbe, Bela Gipp, Akiko Aizawa, Martin Potthast

原文摘要

The generative plagiarism detection task at PAN 2026 aims at retrieving the original scientific sources for given automatically generated textual plagiarism. We created two novel datasets of synthetically generated plagiarism following the premise of a malicious actor trying to generate papers with minimal effort while evading detection of undisclosed re-use. We utilized four widely available LLMs: GPT-OSS 120B, Llama 3.1 8B, Qwen 3.5 27B FP8, and DeepSeek-R1 Distill Llama 8B. In this task overview paper, we outline the creation of this dataset and summarize and compare the results of all participants with three baselines. For this year's task, we received 26 submissions from six teams that significantly outperformed our baselines.

cs.IR

新论文 · v1arXiv:2611.20388

What Unlinked Aggregates Still Prove: Ranker Comparisons from Exact and Differentially Private Item Counts

Chang Xu, Hao Wang, Gedi Zhou

原文摘要

Privacy controls can prevent an evaluator from linking purchases to the recommendations users received, while a merchant can still release complete item-level purchase counts for the evaluation cohort. We ask when these counts identify which recommender performs better. The counts constrain who could have bought each item. We bound the performance difference over all assignments consistent with the counts. A reduced transportation network handles full catalogs and can certify attainable endpoints. If the interval excludes zero, it identifies the better recommender for that cohort. With differentially private counts, one confidence guarantee applies jointly to all comparisons from the same release. In two e-commerce cohorts of about 20,000 and 100,000 users, exact counts narrowed every interval compared with bounds that ignored the counts. Most decisions involved a recommender that favors popular items. For NDCG@10, which rewards ranking purchased items higher, each setting combined one of three personalized pairs with a pattern and rate of missing links. Exact counts settled 14 of 168 settings in the smaller cohort and none of 81 in the larger cohort. In the smaller cohort, exact counts broken down by user group settled two personalized pairs when 90% of users remained linked, at the cost of revealing more about individual purchases. Whole-cohort private releases with one count per item retained much of the narrowing but settled no personalized pair in either cohort. Grouping less popular items into bins let private counts settle one personalized pair in the smaller cohort.

cs.IR

新论文 · v1arXiv:2611.20378

Straight to the Product: What Online Retailers Miss When Shoppers Ask AI First

Michael Iannelli, Alan Ai

原文摘要

When people ask an AI assistant what to buy, what happens next at the store? We link the same people's assistant conversations to their later browsing at twenty large online retailers. People go to the product the assistant named: among shoppers who had not visited the store in the past week, visits to the named product within a day rise from 0.02 to 2.07 per 100 answers (+2.0 points [1.7, 2.4]), and on days with a product-page visit at the store, 44% show a product from their own answer, against 1.5% for a matched answer given to someone else. The result replicates in a separate five-month period. Few arrive through the assistant's link: 89% of visits carry no sign of it, and 54% follow a web search, the same share as their usual visits. When a search comes just before the first visit, it names the answer's product 28% of the time (1% in earlier weeks), as if the search box were an address bar. The chat names the product; the store logs a search. Once there, shoppers keep browsing. We propose product follow-through, the share of answers followed by a visit to the product they named, as a way to evaluate shopping assistants and agents by what people do next: it needs no referral data or provider logs and can be computed for each assistant.

cs.IR · cs.CY · cs.HC

新论文 · v1arXiv:2611.20319

How did data sharing evolve in the decade since the FAIR Principles?

Sheeba Samuel, Daniel Mietchen

原文摘要

A decade after the publication of the FAIR Guiding Principles, what have we learned about data sharing, and what do data availability statements (DAS) and other materials accrued over that period tell us about how data-sharing practices, attitudes, and infrastructures evolved? Here, we present a large-scale longitudinal analysis of over 6.8 million biomedical articles from the PMC Open Access Subset, spanning 2011 to 2025. Using an automated pipeline, we extracted and classified DAS, repository links, accession numbers, and supplementary materials, assigning each article a data-sharing score and independently FAIR-assessing a subset of cited dataset URLs. We find that explicit DAS adoption rose from ~2% in 2011 to 68% in 2025, while "available on request" statements grew in parallel from near zero to over 19%. The corpus growth amplifies both patterns: articles with a DAS containing an explicit repository mention rose from 132 in 2011 to 94,697 in 2025, while "available on request" articles rose from 77 to 173,474 over this span. Structural biology and genomics consistently achieved the highest data-sharing scores, while clinical research and public health ranked lowest. Gold Open Access journals showed a small advantage over subscription journals, while Diamond Open Access journals outperformed both. Internationally collaborative articles scored higher than domestic ones. A gender gap emerged: female first-authored articles' scores declined more than male first-authored ones. The analysis further covered additional dimensions, from compliance with funder/journal mandates and JATS XML tagging standards to computational reproducibility signals and the co-occurrence of DAS with ethics statements. A decade on, the gap between declaring data availability and actually sharing machine-readable, reusable data persists across fields, infrastructures and demographics.

cs.IR · cs.DL

版本更新 · v2arXiv:2610.19608

Forecasting Brand Mentions in LLM Answers: Frozen Predictions Evaluated from Hours to Weeks

Dmitrij Żatuchin

原文摘要

Companies and agencies now monitor which brands large language models (LLMs) name when buyers ask for recommendations. Answers vary between runs, so a single check can mislead and repeated checks cost money. We extend an earlier 20-day replication to ask how well past brand mentions predict later ones over seven weeks. We froze probabilistic forecasts before each of four prospective collection waves of Gemini 3.1 Pro answers to the same 54 buyer questions: nine question families in English, Estonian and Finnish, each for two markets, with four of 12 tracked brands per question. The waves were collected between 1.5 hours and 52 days after the history used by the earliest forecast, and each scored 648 brand and answer labels from 162 answers. A smoothed per-question history achieved a Brier score of 0.0599 at 1.5 hours, against 0.1131 for a per-brand rate and 0.25 for a constant 0.5. A pooled rate clipped to 0.05--0.95, named as the challenger before its first test, scored 0.0259 after 8 hours, 0.0269 after 30 days and 0.0328 after 52 days. Of the brand and question cells named in all or none of the earlier answers, 94.5% to 97.5% kept that pattern in each later wave, and none switched from always to never or back. With a search instruction, 100 of 162 answers returned search activity, compared with 1 of 540 earlier Pro answers; mentions stayed predictable (Brier 0.0563), and Pro history predicted a smaller model's answers less well (0.0787). Human coding agreed with 298/300 sampled automatic labels; two coders agreed on 99/100 shared items (kappa = 0.98). The results cover one provider family, a fixed candidate set and lexical mentions. They support per-question history as the baseline that any forecasting or monitoring method for LLM brand mentions should beat.

cs.IR · cs.CL

版本更新 · v2arXiv:2611.08804

Generative Search and Recommendation at Shopify

Juhana Kangaspunta, Khalil Hajji, Darshan Fofadiya, Kyumin Park, Michael Garner, Anthony Maltsev, Ryan Synk, Jose Sepulveda 等 11 位作者

原文摘要

Search and recommendation at Shopify require retrieving relevant products from a changing catalog of billions of products under strict latency constraints. We present SID-LLM, which casts both tasks as generating semantic IDs: short sequences of content-derived tokens representing products. We jointly pre-train SID-LLM on catalog and behavioral data, giving search and recommendation a shared representation and training objective. We show that reinforcement learning further improves judged search relevance. Furthermore, distillation from task-specific experts improves recommendation quality over joint supervised fine-tuning without a statistically detected loss in search quality. Finally, for efficient serving, we pair multi-token prediction with a lightweight ranker, improving retrieval quality over our autoregressive baseline while reducing latency. We evaluate SID-LLM both as a standalone retriever and as a component of Shopify's production search and recommendation systems. As a standalone search system, SID-LLM achieves quality comparable to Shopify's existing multi-model search stack while meeting production latency requirements. Integrating SID-LLM into the search and recommendation stacks provides significant gains in both offline evaluations and online A/B tests. Together, these results demonstrate the practical value of semantic-ID generation as a unified approach to product discovery at scale.

cs.IR

版本更新 · v2arXiv:2606.25787

Cited Web Sources in LLM Answers about Brands: A Multilingual Audit

Dmitrij Żatuchin

原文摘要

Displayed citations make part of generative search observable, but they do not reveal its complete retrieval process. This audit covers three archived datasets containing 167,551 URL-bearing citation entries across 128 brands, 13 languages and 12 home markets. In the principal Nordic-Baltic component, 3,863 of 21,033 registrable domains account for 80% of 131,514 citations. Wikipedia leads in 11 of 12 query languages. A recovered, separate brand-associated website classification retains 131,512 entries: 108,410 (82.4%) fall outside its brand-associated category and 23,102 (17.6%) inside. This classification does not establish legal ownership. In the full 57-brand Polish aggregate, four recruitment hosts receive 637 citations against 297 for Polish Wikipedia. The results show why normalization, classification rules and denominators belong in citation audits. They describe displayed links without estimating causal effects on brand reputation.

cs.IR · cs.CL · cs.CY

版本更新 · v3arXiv:2603.27910

GAAMA: Graph Augmented Associative Memory for Agents

Swarna Kamal Paul, Shubhendu Sharma, Nitin Sareen

原文摘要

AI agents that talk with users across many sessions need long-term memory to stay coherent and personalized. Existing approaches use either flat retrieval-augmented generation (RAG) that ignores how memories connect, or entity-centric knowledge graphs. Entity-centric graphs split each message into separate triples, which loses its context, and in conversations the speakers become hubs linked to almost every message, so propagation through them tells retrieval little. We present GAAMA, a graph-augmented associative memory that organizes conversations around topics. It keeps verbatim episodes, uses an LLM to distill them into atomic facts and topic concepts, and then synthesizes higher-order reflections. These four node types are linked by five kinds of structural edges, and topic concepts supply the cross-cutting paths, so mega-hubs do not form. Retrieval blends nearest-neighbor similarity search with Personalized PageRank whose transitions depend on edge type. We also introduce GRAFT (Graph Repair by Augmenting Facts & Topology), an offline repair step: for questions whose retrieved evidence fell short, a bounded ReAct agent searches memory again and records the grounded answer as a fresh fact, so later questions find it directly without extra query-time cost. On LoCoMo10, GAAMA achieves the highest average score of all compared methods (79.1%) and is competitive in every question category; GRAFT adds a small further gain. In a controlled comparison, replacing GAAMA's concept-linked facts with entityrelation triples lowers reward by 3.0 pp. On MemoryArena, GAAMA also has the best average across the three datasets.

cs.AI · cs.IR · cs.MA

版本更新 · v2arXiv:2611.06044

Do Search-Augmented Recommenders Use the Latest Search Episode?

Jiandong Ding

原文摘要

When a user returns to a recommendation feed after searching, does the latest search episode help rank the next item? Aggregate ranking gains leave open whether search-augmented recommenders use the latest search at prediction time or benefit from broader search state and representation learning. FocalSwap is a counterfactual audit that masks or replaces the latest episode or the full search history while holding the target, candidates, recommendation history, and trained scorer fixed. A planted-signal test and GenSAR's native search-item prediction task verify that FocalSwap detects known search dependence. In proxy analyses, KuaiSAR's routing gain falls below the shuffled threshold, and Qilin's clearest matched-search advantage appears at high taxonomy overlap. The primary comparisons for SESRec and UniSAR do not establish a positive ranking gain from the latest search episode alone. For the full search history, the estimated contribution depends on the control. On 66,139 shared instances across three UniSAR checkpoints, neutral masking consistently lowers mean reciprocal rank (RR). Replacing the history with observed searches matched on category and activity profile changes 44.1 to 44.8 percent of target ranks, yet the original history has little average advantage. In the primary assignment, matched replacements yield mean RR 0.002770 to 0.006604 higher than neutral masking across checkpoints. The adjusted intervals for this difference lie above zero under the primary assignment and all five alternative assignments per checkpoint. Search audits should specify which inputs changed and what replaced them, and report both how often ranks move and how much the original search helps.

cs.IR · cs.LG

新论文 · v1arXiv:2611.20034

EchoSpec: Training Search Agents to Weigh Evidence Above Echoes

Shihao Yang, Can Zeng, Xinzhe Li

原文摘要

Search agents trained with reinforcement learning decide what to retrieve, when to search again, and when to answer. In real corpora, one claim is copied and paraphrased across many pages while the primary source settling a question appears once. We show that reinforcement-learned search agents grant echoes the credence of independent sources and lose their ability to abstain. Across five base models from 3B to 32B and seven released RL agents on multi-hop Wikipedia questions with primary evidence intact, eight paraphrased false-claim echoes overturn 58-73% of previously correct answers, while exact-match RL drives deleted-evidence abstention below 15%. Conditioning on retrieval localizes the failure to evidence acquisition: once primary evidence enters context, agents overrule eight echoes as reliably as one, whereas echoes fill top ranks and halt search early, so larger budgets, forced search turns, or passive passage filtering alone fail to recover displaced evidence. We introduce EchoSpec, specifying reliability across five corpus transformations with known target answers and training in two stages. Verification warm-start teaches multi-hop bridge-entity recovery past echoes, after which group RL optimizes an evidence-gated $λ$-reward with an outcome-independent acquisition bonus and scale-preserving advantages retaining the optimal rule $λ/(1+λ)$. Across Qwen2.5-3B, Qwen2.5-7B, and Llama-3.1-8B backbones on nine evaluation suites, EchoSpec cuts eight-echo deference by 30-44 points on Qwen2.5-7B down to 29-31%, nearly quadruples eight-echo reach to 26-27% with zero-shot transfer to unseen BGE and BM25 retrievers, and raises standalone QA precision to 47-55%. Code is publicly available at https://github.com/unihsy/EchoSpec.

cs.IR

新论文 · v1arXiv:2611.19974

Take a Step Back: A Dual-Granularity Raw Feature Empowered Intermediate Fusion Graph Convolutional Network for Multimodal Recommendation

Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Wenhao Yuan, Ziyue Peng, Puzhen Wu, Shujie Li 等 11 位作者

原文摘要

Although multimodal recommendations have shown the ability to learn user preferences and item properties by leveraging explicit multimodal information, they remain constrained by two critical limitations: low-quality representation learning and modality fusion. For the former, existing solutions primarily focus on multi-granularity representation learning; however, we argue that training such representations is difficult due to the inherent data sparsity problem in recommendation scenarios. For the latter, existing solutions often rely on contrastive learning to align representations of different modalities to facilitate modality fusion, yet alignment-only objectives may over-compress valuable modality-specific information. To this end, we suggest \underline{``Take a Step Back"} and propose a \textbf{D}ual-granularity raw feature empowered \textbf{I}ntermediate fusion \textbf{G}raph \textbf{C}onvolutional \textbf{N}etwork (\textbf{DIGCN}) for multimodal recommendation. DIGCN leverages the multimodal understanding capabilities of Large Language Models (LLMs) to construct dual-granularity raw features directly. Furthermore, DIGCN introduces an intermediate fusion graph convolutional network equipped with dynamic cross-modality weight correction, designed to capture cross-modality signals and achieve effective modality fusion. Comprehensive experiments on four real-world datasets validate that DIGCN consistently outperforms a variety of strong baselines. Additional analyses confirm its superiority under varying levels of data sparsity and cold-start scenarios. Notably, DIGCN demonstrates high extensibility, capable of further benefiting from additional contrastive learning-based alignment tasks.

cs.IR

新论文 · v1arXiv:2611.19967

DetouR: Plugging Association Routes into Search Agents for Multi-hop Question Answering

Tingwei Chen, Yunxiao Shi, Zhendong Chu, Zhiying Deng, Min Xu, Qingsong Wen

原文摘要

Search agents answer multi-hop questions by interleaving reasoning with retrieval, deciding at each step what to search next. Existing agents follow semantic routes: they infer what is missing and phrase it as a query, reaching only what the model can name. Yet the decisive page is often one the question never names, frequently just one hyperlink away from evidence already retrieved. Following such association routes reaches more gold pages than semantic search under the same budget. Exploiting them raises two challenges: obtaining the association structure without costly offline graph construction, and keeping linked evidence in the reader's limited context, where relevance- and fusion-based selection discards it. We propose DetouR, a plug-in that reuses the links arriving with retrieved pages and adds no search or LLM call. Link Rerouting replaces near-repeat queries with routes to linked pages, chosen by a lightweight learned scorer; Chain-aware Context Selection runs personalized PageRank over these links so that evidence chains reach the reader intact. Across four agents, DetouR improves EM by 4.0 points on 2Wiki and 1.4 on HotpotQA on average. We also find that heavier LLM models struggle with this non-semantic evidence, adding little over these simple selectors or even falling behind. In summary, our work reveals association routes as a cheap complement to semantic routes that current search agents overlook: added as a plug-in, they improve every agent we test, in replay and in live runs, at no extra search or LLM cost.

cs.IR

新论文 · v1arXiv:2611.19926

A Demonstration System for Faithful Temporal Question Answering

Zhen Jia, Xinghong Hou, Chenyang Jin

原文摘要

Temporal question answering (TQA) aims to answer questions that involve temporal constraints or require temporal information. We propose Faith, a novel TQA approach that operates over heterogeneous sources and ensures explainable answers by presenting traceable and faithful evidence. In Faith, we design a new recursive answering mechanism transforming implicit temporal constraints into explicit conditions. Additionally, we visualize the system functionality to track how the answers are computed and the supporting evidence is derived. The code and video of the demonstration system are publicly available.

cs.IR

新论文 · v1arXiv:2611.19900

Attribute or Abstain: Evaluating Evidence-Grounded Retrieval on the Low-Resource African Business Web

Bakary Gibba, Muhammad Abba Mandara, Karim Ismail Abd-Ellah Mohamed

原文摘要

Retrieval-augmented systems that research companies are judged by how much they find, yet in markets where public evidence is sparse, a confident answer without support costs more than no answer. We introduce AfriBizClaims, a benchmark of 150 company-research queries across six African markets in English, French and Portuguese, including queries about fictitious companies that a correct system must decline. Each claim a system makes is judged against the page it cites by an AI judge, checked against a second AI judge from another model family and against two human annotators, and abstentions are scored against answerability labels. We compare BM25 over a pooled corpus, standard web retrieval-augmented generation, and a multi-step agent pipeline (Konexium V1), all using the same language model. The two simple baselines are well attributed: 98% of their judged claims are supported by the page they cite, and they decline every field of every fictitious company. Counting claims about a company other than the one asked about as errors, the agent pipeline's per-query precision is 68%, and it never answers a question about a named company correctly: it substitutes other firms, including for 5 of the 24 fictitious names, and it is the least stable between runs and the most expensive. In the pipeline we evaluated, more steps between the query and the evidence did not buy more grounded answers. We will release the queries, system outputs, labels and short evidence excerpts.

cs.CL · cs.IR

新论文 · v1arXiv:2611.19849

NoLBERT-Calls: Reading Earnings Calls Without Hindsight

Kevin Foley, Jonathan Hartadi, Shivesh Prakash, Hino Wirachapong Suwanphibun, Swapnil Vatsal

原文摘要

Language models are now a standard tool for measuring what managers say on earnings calls, but the encoders behind these measures were usually pretrained on text from the same years a backtest covers. Holding out a test period does not remove that hindsight, because it sits in the features. We adapt NoLBERT, an encoder pretrained only on text published between 1976 and 1995, to earnings calls without letting it read anything dated after 31 December 2023. The calls supply the training signal themselves: every analyst question is followed by management's answer, and the model learns to pick that answer out from the other answers of the same call. We test on every US call from January 2025 to October 2026 (20,323 calls, 254,575 questions) and make every modeling choice on calls from 2024. The adapted model, NoLBERT-Calls-2023, ranks the true answer first for 78.0% of questions (95% interval 77.8-78.2; chance 8.0%), against 26.3% before adaptation. It is 1.0 point ahead of a call-adapted ChronoBERT, 3.4 points ahead of a call-adapted FinBERT, and 18.7 and 19.8 points ahead of the best off-the-shelf retriever and BM25. A year-by-year test over 2013-2026 finds no look-ahead premium: relative to a model adapted only through 2012, the released model does no better on the years it was trained on than on years it never saw. The price of the cutoff is compute: under one shared 150-minute recipe NoLBERT adapts least well of four encoders, and its lead comes from a longer, staged adaptation of about 13 GPU-hours.

cs.CL · cs.IR · cs.LG · q-fin.GN

新论文 · v1arXiv:2611.19847

Reflection Is Not Always Worth It: Budgeted Reflection Allocation for Recommendation Agents

Jinfeng Xu, Zheyu Chen, Ziyue Peng, Jianheng Tang, Zheng Lin, Wenhao Yuan, Jian Chen, Shujie Li 等 9 位作者

原文摘要

This work aims to make recommendation agents more efficient by reducing unnecessary LLM reflection while preserving full-catalog ranking and inference reranking. Since BRAVE changes only how training-time reflection is allocated, it does not introduce new user data requirements or additional LLM exposure beyond the base agent protocol. The main potential risk is that a poorly calibrated allocator may assign less reflection to users whose preferences are sparse, niche, or under-represented, which could weaken personalization for those users or indirectly affect item exposure. This risk can be mitigated through deployment-time auditing and conservative allocation design. In practice, reflection budgets should be evaluated not only by aggregate Recall or NDCG, but also across user-history lengths, item popularity groups, and application-relevant cohorts when such analysis is permitted. The allocator should rely on task-side ranking evidence rather than sensitive attributes, and fallback policies can reserve a small portion of the budget for random or fairness-aware coverage when subgroup robustness is important. Privacy risks can be controlled with the same protections required by the base recommendation agent: minimizing user text sent to the LLM controller, anonymizing user identifiers, retaining only necessary interaction evidence, and applying standard access-control and logging policies. Overall, BRAVE is designed as a cost-aware allocation layer, and its ethical deployment depends on validating that the saved computation does not come at the expense of vulnerable users or responsible exposure.

cs.IR

版本更新 · v6arXiv:2509.11080

Membership Inference Attacks on Recommender System: A Survey

Jiajie He, Xintong Chen, Wenjin Liu, Xinyang Fang, Min-Chun Chen, Yuechun Gu, Keke Chen

原文摘要

Recommender systems (RecSys) have been widely applied across e-commerce, finance, healthcare, and social media and have become increasingly influential in shaping user behavior and decision-making, underscoring their growing impact across domains. Since RecSys heavily relies on user data, its privacy concerns are significant and need to be addressed urgently. Recent studies on membership inference attacks (MIAs) in RecSys highlight this need. MIAs aim to infer whether a user or an interaction record was used to train a target RecSys model. Successful MIAs can lead to severe privacy breaches, such as revealing sensitive aspects of users' lifestyles. MIAs in RecSys have features distinct from those targeting classification models or large language models. However, no systematic survey on this topic has yet been conducted. We present the first comprehensive survey on RecSys MIAs, exploring their taxonomy, design principles, evaluation methods, and defense mechanisms. Based on the summary of existing studies in this area, we also outline several promising future research directions. This survey aims to raise awareness of privacy risks among RecSys researchers, practitioners, and users, and promote privacy protection practices in RecSys design.

cs.IR · cs.AI · cs.CR

新论文 · v1arXiv:2611.19770

Exact Sparse Assignment with Costly Target Embeddings: Fixed-Realization Certificates and an Empirical Audit

Hamidreza Hasani Balyani, Mahtab Movahhedrad, Seyed Pouyan Mousavi Davoudi

原文摘要

Embedding-based one-to-one association is common in vision and data integration: detections are associated with tracks and records are linked across databases. Such pipelines often solve a sparse optional assignment whose edge scores add a cheap term to an inner product of neural embeddings. When target embeddings are costly, one can acquire them selectively and stop once a certificate proves that the current assignment is optimal. We give such a certificate for unit-norm embeddings. It bounds the unknown part of every reassignment by a support function and maximizes the resulting robust gains with a second sparse assignment; a shared owner-centred bound dominates any fixed-centre triangle bound. We make precise what is certified when network features are not batch-invariant: the realized values of the acquired rows and every unit-ball completion of the rest, not a second forward pass in a different batch. Auditing against a separately batched cache tests a different property; on our first pilot it produced a spurious $2.49\times10^{-6}$ objective failure. We audit the corrected protocol on four public text entity-matching benchmarks with two frozen encoders; no image data is used. All 96 certificate simulations and 120 online selective timings were exact on their own realization (largest objective gap $6.82\times10^{-13}$). At group size 8 the shared certificate acquires on average 60% of candidate-active targets, against 74% (mean centre), 69% (MEB) and 60% (zero centre): it beats the mean and MEB centres but not the zero centre, and savings concentrate in one benchmark (7% acquired). The prespecified runtime claim is contradicted: selective execution is +0.510 s slower per run (dataset-cluster 95% interval [+0.248, +0.772]). Uniform-microbatch re-encodings preserved the optimum in only 5 of 8 encoder-dataset cells, and 45 of 120 online runs disagreed with an eager cache at near-ties.

cs.LG · cs.IR

新论文 · v1arXiv:2611.19752

StaleBench: Isolating Stale Retrieval in Agent Memory, and Pricing the Fix

S M Mahbub Murshed

原文摘要

An agent that stores facts over time eventually stores several versions of the same fact. Asked a value-agnostic question - "which team does X play for now?" - a similarity retriever cannot prefer the current version, because every version matches the entity the query names. StaleBench isolates this stale-retrieval failure on 892 real updating facts, with currency labels from the source (Wikidata validity years via TempLAMA), never from a retrieval score. Our central result is that retrieval metrics can rank memory policies backwards: a retention policy and its retain-everything control return the identical statement at rank 1 for every test query, yet differ by 29.5 and 65.6 points in reader accuracy on two readers when no date is shown. Contamination@5, the metric we add, ranks them correctly where current@1 and current@5 do not. Withdrawing the lexical anchor that statements carry costs deployed keying mechanisms tens of points of freshness; entity resolution on both the write path and the query recovers most of the loss. Three deployed systems (Mem0, Graphiti, LangMem) and a reimplemented ERASE each handle supersession differently, and no single scoring convention describes all four. We release the data, a frozen evaluator, and reference baselines.

cs.CL · cs.AI · cs.IR

新论文 · v1arXiv:2611.19655

User First Intelligence: Three Measurements of Search and AI-Answer Analytics (SEO and GEO/AEO), and a Knowledge-Graph Framework for Query Discovery Beyond Search Volume

Venkata Pagadala

原文摘要

People say what they want to know in many places besides a search box, and the tools that search and AI-answer optimization (SEO and GEO/AEO) rely on see only part of that demand, through models reported as counts. This paper reports three measurements of such instruments and a framework. On the area-code section of a large US telecom provider's site, two commercial analytics platforms differed on estimated monthly visits by 7.5x to 27.1x (8.6x to 29.9x worldwide); across twelve sections of the domain the ratio ran from 0.20x to 7.24x with no detectable relation to section age, while they agreed on which section is larger in 60 of 66 pairs. On 37,193 reviews of 378 telecom stores linked to a 42,994-node keyword taxonomy, review demand and search volume agreed on branches to a varying degree (Spearman 0.47 to 0.75; 0.91 post hoc) and barely on topics (0.18), on links two blind AI annotators judged two-thirds correct; on the 372 area codes, a public count of Wikipedia requests for each code and vendor volume agreed roughly on ranking (0.45 over 240 articles, gone once keywords matched per code are held constant) and not on magnitudes, predictions fixed first. In one vendor's index of AI answers naming or citing the largest companies of each S&P 500 sector, of four predictions written first, one held, one in direction only and two failed; video and community sites took a median 21% of citation incidences on Google's AI Overviews and AI Mode against 0.6% to 9.2% on four other surfaces, in every sector, and in education answers in each of 13 countries read later. Estimates and sources need measuring per instrument and surface. User First Intelligence (UFI), built on that premise, collects what people ask beyond the website, links it to an industry ontology and scores where an organization does not answer; an open implementation is released, its score not evaluated. No causal claim is made.

cs.IR

版本更新 · v5arXiv:2601.00821

Fidelity Before Structure: Verbatim Chunks Beat Lossy Artifact Extraction in Long-Conversation LLM Memory

Tao An

原文摘要

A growing class of conversational-memory systems compresses dialogue history into structured artifacts (extracted facts, decisions, or events). The premise is that distilled structure retrieves better than raw text. We test this premise with a controlled ablation. Within one fixed retrieval--rerank--reasoning pipeline, we swap only the stored representation (LLM-extracted typed artifacts versus verbatim conversation chunks) and hold the model, retriever, reranker, and judge constant. Verbatim chunks win by 15.9 points on LoCoMo (43.9% vs. 28.0%) and by 22.0 points on LongMemEval-S (67.4% vs. 45.4%). A semantic graph does not recover the gap, and six confound controls reproduce the effect. The mechanism is lossy distillation rather than structure per se. Accuracy tracks how much source text survives in the store, and the extracted-artifact pipeline does not beat naive RAG in overall accuracy (chunks do abstain worse; see Limitations). For the extraction designs we test, structured memory should augment verbatim text rather than replace it. Adding artifacts alongside chunks costs no detected accuracy, while substituting them forfeits the gap. Code and data are available at https://github.com/tao-hpu/cog-canvas.

cs.AI · cs.CL · cs.IR

版本更新 · v2arXiv:2512.16795

From Evidence to Judgment: Trace-Augmented Reasoning under Conflicting Evidence

Shubham Mishra, Atharv Sharma, Dhruv Kumar, Harsh Sharma

原文摘要

Retrieved documents often disagree, contain outdated information, or fail to support any answer, and retrieval-augmented generation (RAG) systems have no reliable way to decide which documents to trust. We treat this as a reasoning problem and test it separately from answer generation. We release 943 training and validation examples and a disjoint 736-example benchmark, and fine-tune four language models to write a structured think trace before answering: the model judges each document, infers the conflict pattern, and writes a grounded answer or a justified refusal. Our evaluation framework, CATS (Conflict-Aware Trust Score), scores grounded refusal, behavioural adherence, factual grounding, and single-truth recall; a human study validates factual grounding and single-truth recall (kappa=0.78 and 0.79). The think trace improves the answer/refuse decision: under all three prompts, untuned models score at most 0.429 Refusal F1 (F1 with refusal as the positive class) and trace-tuned models at least 0.826. On the answerable examples a model answers, factual grounding under the 40-word prompt is at most 0.11 for untuned and 0.86-0.88 for trace-tuned models. Cutting the prompt from 1,965 words to 40 leaves trace-tuned scores within the benchmark's measurement noise (+/-0.034). Answer-only fine-tuning trails trace supervision on every metric under the minimal prompt.

cs.CL · cs.AI · cs.CY · cs.IR