Papers

Retrieval-augmented generation (RAG) can improve the specificity and grounding of large language model responses, but its effect is not uniformly beneficial in single-turn mental-health question answering, where user queries often combine emotional distress, treatment concerns, and safety-sensitive needs. We study when retrieval helps or hurts mental-health QA, and whether a lightweight selective retrieval policy can better control this trade-off. We operationalize retrieval need using three draft-conditioned utility dimensions: psychoeducational need, coping need, and response specificity, together with a rule-based safety trigger. Following psychotherapy-grounded RAG systems such as coTherapist, we construct a compact and controllable guideline corpus comprising coping-strategy, psychoeducational, and safety resources. We fine-tune an instruction-tuned generator on MentalChat16K using QLoRA and compare Closed-book, Always Retrieval, and Selective Retrieval settings on CounselBench-Eval and CounselBench-Adv. Experiments show that retrieval is not uniformly beneficial in this domain. Always Retrieval improves specificity but lowers overall quality and introduces additional safety-sensitive failures. Selective Retrieval preserves closed-book behavior for low-need cases while avoiding the additional degradation caused by unconditional retrieval, supporting the view that retrieval activation is a safety-sensitive control decision.

A model that inherits one-line memories may pull one archived source record before acting; a directive in the store can steer that pull: a pointer, a criterion or both. Across sixteen registered studies (179,352 attempts) we measured where the request goes under each form; every result is descriptive, with registered intervals, no mechanism claim. A length-matched criterion exceeded a bare id on six direct-provider models (D) and failed its registered superiority rule on a nine-model OpenRouter panel (E). On generated worlds (K2-K5): the two registered signatures held on Opus 5 and Fable 5.1, Fable 5 followed the same sign, Haiku 4.5 reversed, and Sonnet 5, the GPT-5.6 endpoints and GPT-6 Astra lay near zero (K2). With a defensive adapter at five gains, the 70B rule for a gain-dependent change of the composite - criterion contrast was not met (K3 and K4); under the 8B attenuation rule (0.95 intervals: slope below zero; change beyond the margin), the 8B change of -17.5 [-26.7, -8.1] did not meet it on 36 families (K4) and at registered power on 337 families -16.6 [-19.4, -13.8] did (realised one-sided error at the margin 1.8 to 3.2% per corner of a finite grid, nominal 2.5%, not a uniform-error guarantee; K4's status stands; K5, first ladder), while a second SecAlign++ adapter under the imposed Meta-SecAlign template did not (-11.8 [-14.3, -9.3]; K5, second ladder); no NOT-MET is a statement that the contrast was unchanged; their difference (+4.7 [+2.3, +7.2]) describes two fixed execution paths, licenses no superiority, equivalence or 'significant difference' claim; nothing follows from the statuses differing (K5). Intervals describe family-reweighting stability conditional on the execution, not reproducibility across engine executions; audit replays were neither substituted for nor averaged into outcomes; no missingness gate fired and directional completions changed no status.

Retrieval-Augmented Generation (RAG) has made dense retrieval over large document collections a standard building block. Organizations increasingly outsource vector indexes to untrusted clouds, exposing proprietary corpora and user queries. Cryptographic protection is challenging because each query searches corpus-scale state, causing computation, correlated randomness, and communication to grow with the corpus. At million-document scale, a naive secure implementation takes minutes and about 90 GB of communication per query. Even recent optimized systems require 10–22 seconds. We propose Spruce (Scalable Private Outsourced Retrieval Using Compact Embeddings), which co-designs representations with the cryptographic protocol. Spruce learns compact binary codes that preserve candidates for full-precision reranking, replacing corpus-wide embedding scoring with efficient Hamming-distance computation under two-server multi-party computation (MPC). A corpus-calibrated fixed-radius protocol avoids multi-round candidate selection while preserving retrieval quality. Spruce also provides private cluster pruning, which trades minor quality loss for substantially less computation, and a one-core owner-operated dealer that removes cloud OT preprocessing bottlenecks. Across four corpora containing 383K–5.42M documents, Spruce preserves the original search quality with median candidate sets of only 382–1,952. At 10 Gbps inter-server bandwidth, full scans take 0.21–2.97 seconds, $4.8$–$6.7\times$ faster than the closest measured prior work. Private pruning takes 0.06–1.09 seconds, achieves $13.1$–$22.9\times$ speedups, and retains $93.9\%$–$97.3\%$ of full-float NDCG. On the largest corpus, pruning and the dealer jointly improve sustained throughput by $31.5\times$ at 1 Gbps per link.

Sequential recommender systems model user behavior as item ID sequences, while recent generative methods cast recommendation as a language modeling task using large language models (LLMs). While this paradigm incorporates rich textual semantics, it introduces a fundamental mismatch: LLMs operate on text tokens, whereas recommender systems depend on discrete item indices. This misalignment often leads to hallucinations in generative recommendations. Existing methods attempt to bridge this gap by learning item vocabularies in Euclidean space, but they struggle to model the inherent long-tail distribution of real-world catalogs, where a small number of head items dominate, and a vast number of tail items reflect users' niche preferences. To address this issue, we introduce Hyperbolic Residual-Quantized Variational AutoEncoder (HypRQ-VAE), the first framework to learn item indexing in hyperbolic space. HypRQ-VAE leverages the unique properties of hyperbolic geometry, whose exponential volume expansion naturally accommodates the power law structure of user-item interactions. This allows the model to encode rich textual semantics while preserving the representational fidelity of sparse, long-tail items. Experiments on three benchmark datasets show that HypRQ-VAE significantly improves the performance of recommendation, particularly in recommending tail items. Our analysis attributes these gains to the superior capacity of hyperbolic space to model item hierarchies and sparsity in generative recommendation. Our code and data are available at: https://github.com/wulongfeng/HypRQ-VAE.

Scientific literature synthesis agents increasingly rely on proprietary online services, limiting reproducibility, privacy, and offline deployment. To address this challenge, we introduce SciLENS Scientific Localized Evidence Navigation and Synthesis), a fully local autonomous agent framework operating on a dual-tier infrastructure indexing approximately 12 million academic records. SciLENS pioneers the integration of structural visualization as an actionable tool within the reasoning loop, enabling the agent to compress complex citation topologies into validated data-driven charts and thereby mitigate context exhaustion during macro-level synthesis. To train the agent without human annotation, we develop an automated data synthesis pipeline that extracts multi-hop subgraphs from a citation knowledge graph, verified by cross-model consensus among 20 frontier models. The agent is subsequently aligned through a reverse-decomposition rubric strategy that provides fine-grained process rewards for early planning and strict evidence grounding. Evaluations across six scientific benchmarks encompassing standard QA, citation accuracy, factual reasoning, and structural synthesis demonstrate that SciLENS significantly outperforms open-source baselines and achieves performance comparable to GPT-5.2 and Gemini-3.0-pro. Our source code and data are released at https://github.com/LQgdwind/SciLENS.

Large Language Models (LLMs) have recently emerged as powerful backbones for recommendation. To better elicit their capabilities, reasoning has been widely incorporated to help LLMs interpret rich textual signals and improve recommendation accuracy. However, explicitly generating intermediate reasoning traces often incurs substantial computational costs, which limits practical deployment in real-world recommender systems. To address this challenge, we propose SelfDR, a Self-Distillation from Reasoning framework for LLM-based Recommendation. SelfDR distills an LLM's own reasoning-enhanced predictions to produce recommendations directly, improving recommendation effectiveness while maintaining inference efficiency. All components in the framework are built on the same base LLM, without relying on any external models. Specifically, the teacher recommender is constructed by training a reasoner with downstream performance as the reward, enabling it to generate targeted rationales that are later incorporated into the teacher's input. A student recommender for direct recommendation, with the same underlying model, then learns from the teacher through self-distillation with a dynamic weighting strategy. Extensive experiments on three public datasets validate the effectiveness, rationality, and efficiency of SelfDR. Codes are available at https://github.com/JiangDeccc/SelfDistillation.

Large language models (LLMs) are effective rerankers, but pointwise reranking repeatedly processes the same document across different queries, causing substantial redundant document-side computation. We propose DoPR, a compressed document prefix framework that decouples offline document processing from online reranking. DoPR first selects query-independent document representations and converts them into compressed document prefix states, which are precomputed offline and reused whenever the document is retrieved. During online reranking, the model scores each query-document pair by processing only the query and scoring token, with document information supplied by the stored prefix states. This design reduces online cost through both document-side compression and cross-query prefix-state reuse. Experiments on TREC DL, BEIR, and BRIGHT with Qwen3 models from $0.6$B to $8$B show that DoPR achieves up to 8.0$\times$ online document-side memory reduction and up to 8.04$\times$ latency speedup, while retaining 97.1%-99.5% of the average NDCG@10 of matched full-document rerankers.

Unified modeling has become a major direction for industrial click-through rate (CTR) prediction. Existing approaches typically unify sequential and non-sequential signals at the token level, model their interactions in a shared backbone, and increase model capacity to improve scaling behavior. However, this division originates from legacy feature-engineering practice and is misaligned with the underlying decision process. User behavior is inherently a sequence of homogeneous context units; at the level of input organization, historical behavior and the current request differ only in whether their outcomes are observed or remain to be predicted. Treating them as heterogeneous signals obscures structural dependencies within the user's decision context, limiting both scaling efficiency and prediction quality. This limitation is particularly pronounced in context-rich scenarios such as e-commerce shelves and waterfall feeds. To address this, we propose UniCon, a unified context-centric modeling architecture that treats the request context as the basic modeling unit and organizes history and prediction targets as homogeneous context units. Intra-context attention captures local coupling among items within a context (Locality), while inter-context attention models the dynamic evolution of decision states across contexts (Dynamics). This organization bridges the structural gap between history and target and supports more effective scaling of unified CTR models. Context-unit-level sequence compression further reduces deployment overhead. On Meituan search advertising, UniCon improves offline AUC by 0.0139 over a strong production baseline and achieves statistically significant online lifts of 3.09% in RPM, 2.07% in CTR, and 2.95% in revenue.

Deciding whether two supplier records belong to the same corporate family is a prerequisite for spend consolidation, credit exposure aggregation and sanctions screening. It is usually treated as entity matching, but the tasks differ: a family link connects records that are deliberately different entities, and the evidence often appears in neither record. We introduce CorpFam, a public benchmark of 54,864 candidate pairs over 10,307 corporate families, derived from 6,638,350 US federal award records in which every supplier self-reports its ultimate parent to a government registry. Pairs are stratified by name visibility: whether the names are identical after normalisation, share a distinctive token, or share none. Because strata have positive rates from 10.2% to 97.3%, we report per-stratum recall, base-rate invariant, rather than F1, which is not. The strongest of 5 matchers recovers 100.0% of identical pairs and 4.2% of invisible ones; no method exceeds 4.7% on the latter. The failure begins before matching. Blocking decides which pairs a matcher sees, and we evaluate 7 schemes spanning phonetic keys, attribute keys that ignore the name, and semantic nearest neighbours. None reaches three percent on invisible pairs, and their union recovers 6.8%. 93.2% of these links never enter the candidate set, so no matching-stage improvement can reach them. The links are real: against SEC Exhibit 21 subsidiary schedules, which share no provenance with procurement registration, 64.2% of invisible links are corroborated, against 0.16% under permuted parents and 0.41% against the same parent's wrong exhibit: two unrelated nulls agreeing to within 0.25 points. Corporate-family resolution is a retrieval problem misfiled as a matching problem; the intervention point is candidate generation, not ranking. The benchmark, adjudication log, and code reproducing every number are released.

Libraries and archives manage large collections with limited staff and computing budgets, yet common benchmarks do not systematically test their bibliographic work. They need to know which methods work for their tasks and what those methods require to run. SHELF, the Synthetic Harness for Evaluating LLM Fitness, addresses this gap. It is a Python system that turns labelled taxonomies, writing specifications, and a generation budget into controlled benchmark data and evaluation tasks. This first release contains 62,899 model-written documents based on Library of Congress vocabularies, with tasks for classification, clustering, retrieval, pair classification, and instruction retrieval. We compare TF, TF-IDF, BM25, popular encoders, and, on subject classification only, zero-shot decoders; each method appears only on tasks that support it. Subject classification reaches 0.8887, while genre-form classification reaches only 0.2605, and several pair and clustering tasks remain near chance. Sparse methods remain competitive on classification, while TF-IDF is the fastest measured arm in the subject timing experiment. SHELF also varies bibliographic facets independently and can generate new, verifiably unseen documents after a model's training cutoff. Comparisons with LCSHBench and Project Gutenberg show that model rankings transfer more reliably than absolute scores, but SHELF scores do not estimate accuracy on production catalogue data. We release all source code and data under permissive licenses on GitHub and Hugging Face.

Selecting a retrieval model for a production RAG system requires reliable comparative evaluation, but obtaining relevance judgments at scale is expensive and difficult to repeat as new candidate systems arrive. We study pooled LLM evaluation, in which an LLM judges the union of documents retrieved by the current set of candidate systems, and the pool is then expanded incrementally as new systems are introduced by judging only the new documents they contribute. These judgments are reused to evaluate all systems on a common basis. We validate this approach on four retrieval benchmarks with 11 systems spanning dense, sparse, and hybrid configurations, and deploy it to compare 62 retrieval configurations for a financial news QA system. Pooled LLM rankings correlate strongly with gold-standard evaluation across datasets, and 97% of pairwise system orderings are preserved once bootstrap uncertainty in the qrels is taken into account. In production, document overlap yields 65-80% judgment reuse and up to 4.9x lower evaluation cost, allowing teams to benchmark new retrieval candidates without re-judging previously assessed documents. These results suggest pooled LLM evaluation is a practical and cost-effective workflow for incremental retrieval model selection in deployed systems.

Sequential recommendation models are foundational to modern personalized services, yet their effectiveness varies substantially across heterogeneous user environments. In particular, static one-pass recommenders often perform well on common behavior patterns but degrade on operationally challenging user groups, such as users with longer histories or less mainstream item profiles. To address this limitation, we propose DS-Frame, an adaptive fast–slow inference framework for sequential recommendation. DS-Frame combines a Fast System for efficient routine prediction, a Slow System for iterative latent refinement, and a learned selector that routes each sample under a controllable computation budget. Experiments on five real-world datasets show that DS-Frame consistently improves representative sequential recommendation backbones, with larger gains on challenging groups and effective accuracy–efficiency trade-offs. This highlights the potential of adaptive inference for more efficient and robust recommendation. Code is available at \href{https://github.com/ZichenYuan233/Recommender-System-as-Slow-and-Fast-Thinkers}{this link}.

Recommender-system experiments often rely on a single random training seed, assuming that run-to-run stochasticity has limited impact on evaluation conclusions. This assumption is risky, as a training seed may influence several algorithm-dependent mechanisms, including parameter initialization, mini-batch ordering, dropout, masking, latent sampling, and training-time negative sampling. We examine this assumption by fixing the data partition and varying the training seed across hyperparameter configurations. We analyze seed effects at three levels: user-level metric sensitivity, validation-based model selection and recommendation-list agreement. Results show that seed variation is often detectable. Its impact depends on whether configurations are clearly separated, whether validation results transfer to test, and whether similar scores lead to similar top-$k$ lists. Findings suggest that reporting single-seed results can overstate the stability of recommender system evaluation, and that training seeds should be treated as part of the evaluation protocol rather than as incidental implementation noise.

Document Visual Question Answering (DocVQA) often leverages Retrieval-Augmented Generation (RAG), where late-interaction encoders are commonly used to identify document pages relevant to a user query, before answer generation by a Large Vision-Language Model (LVLM). Existing approaches typically retrieve a fixed top-$k$ number of pages regardless of query complexity, which increases LVLM latency and may degrade answer accuracy. We introduce ViSAR (Visual Semantic Activation Retrieval), a training-free adaptive-$k$ retrieval method for late-interaction visual document retrieval. ViSAR operates directly in the embedding space to construct a query-conditioned page-level similarity matrix that highlights query-relevant semantics and dynamically determines the number of pages to retrieve. Across multiple encoders and LVLMs, ViSAR retrieves compact, query-adapted page sets that reduce RAG latency by up to 58.7%, while maintaining or improving answer accuracy compared with fixed top-$k$ and adaptive retrieval heuristics. Furthermore, we show that the similarity matrix structure correlates with answer accuracy, suggesting future directions for retrieval quality-aware document understanding.

While existing work on LLM authorship attribution (AA) has made progress, available benchmarks remain limited, often focusing on English, controlled settings, or relatively outdated models, with the few multilingual studies considering only relatively short texts. We introduce MultiGhostBench, a multilingual benchmark comprising 928 books generated by five recent LLMs across six languages and three scripts, with an average length of approximately 59K words per book. The benchmark supports evaluation under domain, author, and language shifts. Evaluation of representative AA methods shows that no single method consistently performs best across settings, and performance generally degrades under distribution shifts. Transformer-based detectors can retain generator-related information across languages, although transfer effectiveness varies by language pair, whereas statistical and fingerprint-based detectors are more language-dependent. We envision MultiGhostBench as a valuable resource for the development and evaluation of robust AA methods. The dataset and code can be found at https://github.com/GrecoMT/MultiGhostBench.

Large language models have enabled natural language interfaces for structured databases, but generated queries may still contain syntactic errors, violate database schemas, or fail during execution. Test-time inference strategies improve generation reliability without additional training, but existing approaches often use fixed inference budgets and uniform refinement strategies, leading to unnecessary computation across questions with different complexity levels. In this work, we investigate adaptive test-time inference for Text2Cypher and introduce two strategies: adaptive trace budgeting, which dynamically adjusts the candidate generation budget based on question difficulty, and selective execution-guided refinement, which applies correction only when additional inference is expected to be beneficial. Experiments on Gemma-2-9B and Qwen-2.5-7B show that adaptive trace budgeting reduces the average generation budget by 30.7% and wall-clock inference time by 21-25% while maintaining comparable generation quality. Selective refinement preserves nearly all execution success gains of full refinement, reducing execution success by only 0.2-0.5% while avoiding unnecessary refinement for simpler questions. Experiments show that a single correction model (Gemma-4) effectively refines outputs from a different model family, suggesting refinement transfers across families.

Generative engine optimization (GEO) enables content producers to increase the visibility of their web pages in generative search engines, but the same techniques can deliver targeted misinformation when adversaries publish ordinary-looking GEO-optimized documents that victim large language models (LLMs) retrieve and synthesize into distorted answers. No existing benchmark evaluates defenses against this threat under controlled conditions. Therefore, we present Counter-GEO-Bench, a defense benchmark that pairs 247 human-verified, quality-gated queries with information-preserving and information-distorting GEO rewrites, and evaluates defenses on attack success rate (ASR), false positive rate, and answer quality across three victim LLMs. Under Counter-GEO-Bench, three off-the-shelf defenses (Granite Guardian, Llama Guard 3, and NeMo Self-Check Fact-Checking) reduce ASR by at most 5.7% relative, while Granite Guardian's reduction is not statistically significant. Safety-taxonomy guardrails target policy violations, while GEO misinformation passes through them as fluent informational content. To this end, a lightweight benchmark baseline, C-GEO Guard, is proposed, reducing ASR by 47.6% relative with near-zero utility loss, which proves threat tractable.

Publishing research data is widely expected to increase its reuse and to inspire new research. In the social sciences, data from surveys, interviews, polls, and statistics are primary resources for research. There is a long tradition to collect and offer research data in data archives and online repositories. Researchers use these systems to identify data relevant to their research. However, especially in data search, users' complex information needs seem to collide with the capabilities of data search systems. The search capabilities, in turn, depend to a high degree upon the metadata schemes used to describe the data. In this research, we conducted an online survey with 72 social science researchers who expressed their individual information needs for research data like they would do when asking a colleague for help. We analyzed these information needs and attributed their different components to the categories: topic, metadata, and intention. We compared these categories and their content to existing metadata models of research data and the search and filter opportunities offered in existing data search systems. We found a mismatch between what users have as a requirement for their data and what is offered on metadata level and search system possibilities.

Reasoning-intensive retrieval addresses queries whose relevance cannot be identified by surface-level matching, thereby requiring multi-step reasoning. Because relevant documents rarely appear in the initial candidate set, retrieval systems suffer from the bounded recall problem. Existing methods iteratively expand a candidate pool at the document level over a corpus graph, examining each neighbor in isolation and drifting toward a narrow region of the corpus. To address this problem, we propose Group-Aware Adaptive Retrieval for Evidence Navigation (GAREN), which explores the corpus graph through group-level expansion. GAREN organizes documents into semantically coherent and distinguishable groups based on their connections in the corpus graph. The information in each group indicates what can be accessed by expanding through it, providing guidance beyond individual document-level signals. At each iteration, GAREN uses a group-level navigator to identify promising expansion directions, retrieves documents from the selected groups, and applies a document-level reranker to evaluate the updated candidate set. Extensive experiments show that GAREN achieves up to 8.0% improvement over the strongest baseline on BRIGHT. The source code is available at https://github.com/KJ12124/GAREN

Serving useful recommendations under distribution shift is crucial for balancing utility and risk in out-of-distribution (OOD) recommendation. However, most existing OOD methods improve ranking or construct counterfactual candidates without controlling the proxy-label false discovery rate (FDR) of the served set. In this work, we formulate OOD serving as the $\alpha$-Valid Counterfactual Recommendation ($\alpha$-VCR) problem to retain candidate support learned from counterfactual supervision while controlling proxy-label FDR, and propose GenCAR, which couples preference-grounded counterfactual supervision with calibrated set selection. In particular, GenCAR fixes the stable-preference representation while intervening on the environmental factor, grounds offline large language model proposals through preference anchors and trust-radius filtering, and uses conformal $p$-values for Benjamini–Hochberg selection. We theoretically bound conditional counterfactual approximation error and prove finite-sample, distribution-free control of proxy-label FDR under exchangeability and positive regression dependence, with a Benjamini–Yekutieli guarantee under arbitrary dependence. Extensive experiments audit realized proxy false discovery proportions and demonstrate that GenCAR consistently enhances OOD candidate recovery across diverse benchmarks.

Multimodal Recommender Systems (MRSs) typically rely on a flawed "modality harmony" assumption, presuming that multimodal features are inherently beneficial and strictly aligned with users' collaborative interaction patterns. However, modality-topology conflicts are ubiquitous in real-world scenarios due to deceptive visual clickbaits and mismatched semantics. Blindly integrating these noisy modalities inevitably pollutes the pristine collaborative space, causing severe representation distortion. To address this, we propose Orthogonal purification and topology-guided MoE for conflict-aware multimodal Recommendation (OrthoRec). At its core, OrthoRec introduces Collaborative-Guided Orthogonal Purification (CGOP), which geometrically decouples multimodal features into directions parallel and orthogonal to a pure collaborative anchor. By adaptively truncating the orthogonal noise with an energy-preserving normalization, CGOP rectifies deceptive semantic directions while preserving the modality's intrinsic representation capacity. Furthermore, we design a Topology-Aware Routing Mixture-of-Experts (TAR-MoE). Guided by the collaborative topology, TAR-MoE employs decoupled sigmoid gating to break the zero-sum bottleneck of traditional softmax attention, autonomously determining the injection scale for each purified modality. Finally, a safe-SSL objective is introduced to dynamically penalize the forced contrastive alignment of contradictory pairs. Experiments on three real-world Amazon datasets show that OrthoRec consistently outperforms competitive recent baselines and exhibits improved robustness under modality noise and item sparsity.

Most vector databases rely on graph-based indexes, notably HNSW and Vamana, for approximate nearest neighbor search. With embedding models widely adopted, the datasets these databases store grow rapidly. At a fixed accuracy, how does search cost scale with dataset size? The prevailing answer is poly-logarithmic growth. Yet the claim is proven only under special conditions and asserted without proof for the indexes used in practice. It is also largely untested: standard benchmarks measure cost at one dataset size, not across sizes. We put the claim to the test. The answer depends on the scale itself. While the dataset size $N$ is small relative to the data's intrinsic dimensionality, search cost grows as $N^c$ for a constant $0<c<1$. We call this scaling the Sublinear Power Law. Once $N$ is large enough, growth slows to subpolynomial, consistent with the poly-logarithmic claim. The Sublinear Power Law appears on every dataset, mostly up to its full size, at every recall target, query hardness level, and index configuration we test. The transition to subpolynomial growth appears on the two datasets that grow large enough relative to their intrinsic dimensionality. One mechanism underlies both behaviors: a dataset's intrinsic dimensionality grows with its size until the data resolves its underlying distribution. Higher intrinsic dimensionality packs more vectors into the query neighborhood the search must examine. We present a unifying theory of beam-search cost that explains our observations. For exact and bounded-degree constructions, we prove the Sublinear Power Law and the eventual transition to poly-logarithmic scaling, and derive the scale at which it occurs. We also develop models that predict the power-law exponents for any recall target and index configuration. These models give a principled way to navigate trade-offs among search cost, insertion cost, and recall as data grows.

Data-centric agents repeatedly perform a discovery step before planning or execution: identifying the data objects relevant to a task. Yet successful discovery outcomes are typically discarded rather than reused. We introduce persistent discovery context, a lightweight memory layer that stores prior intent-to-object mappings and reuses them to augment future retrieval. Across three structured data environments, persistent discovery context consistently improves retrieval quality over metadata-only search, remains effective with automatically generated memories, and exposes a reproducible interference failure mode. In lexically sparse domains, memory-only retrieval can even outperform metadata-based retrieval. These findings suggest that discovery outcomes constitute a useful form of reusable context for data-centric agents.

Generative Point-of-Interest (POI) recommendation, autoregressively generating a target POI's semantic ID (SID), holds great promise for Location-Based Services, where a recommendation helps only if the user can reach it. Yet, existing methods operate within an interest space defined by behavior sequences and collaborative signals, where geography enters only as a textual attribute of the SID, leaving no explicit mechanism to learn or preserve how urban places are related by distance, direction, and reachability; their predictions are thus behaviorally plausible yet far from the user's real-time location. We argue that such services require injecting real urban spatial knowledge into the interest space, rather than inferring geography from behavior alone. Hence, we propose SPAR, a unified framework whose three synergistic stages jointly construct, cultivate, and preserve urban spatial knowledge: (1) at the tokenization level, Spatially-Intrinsic SID (SI-SID) explicitly encodes longitude–latitude coordinates into a sinusoidal geospatial embedding and fuses it with the textual semantic embedding, producing identifiers via RQ-Kmeans that are simultaneously semantically and geographically consistent; (2) at the cognition level, Multi-Granular Geospatial CPT (MG-CPT) continually pre-trains the base LLM on 25 curated geospatial datasets organized into three tiers of basic attributes, pairwise relations, and city-scale navigation, so that scattered POIs cohere into a connected urban space; and (3) at the adaptation level, Task-Vector Anchored SFT (TV-SFT) anchors the acquired spatial knowledge as a frozen parameter-space task vector to prevent its catastrophic forgetting during behavioral fine-tuning, thereby fusing the two spaces. Extensive quantitative and visualization experiments on two public and four industrial-scale datasets demonstrate the effectiveness of SPAR.

Point-of-interest (POI) localization matches user-provided storefront close-ups to the same shops in wide, geo-tagged vehicle-mounted street views. POIs may change while the surrounding scene stays similar, so scene-level recognition alone cannot establish POI identity. Differences in target scale and capture domains further challenge matching. We introduce POI-Loc, to our knowledge the first benchmark dedicated to this asymmetric, fine-grained POI localization task. Many visual place recognition methods represent each image with a single global vector, which tends to dilute fine-grained features of small storefronts amid background clutter. We propose GLAM (Global-to-Local Asymmetric Matching) to combine global and local evidence. In stage one, a single attention-pooled query probe is matched against compact reference region tokens via learnable soft top-k interaction, with the resulting local similarity fused with global similarity for retrieval. Stage two reuses query region tokens before attention pooling and stored reference tokens for mutual-nearest-neighbor re-ranking. GLAM surpasses both global and two-stage baselines on Recall@1/5/10 and mAP, with about $5\times$ smaller re-ranking features and $280\times$ lower per-pair matching cost than FoL. The benchmark and code will be released at https://github.com/roadhan/glam.

Large language model question answering over power-grid models must respect a fixed context budget. We introduce seed-anchored graph rendering, a deterministic method that prioritizes query-local graph evidence without adding method-specific tuned or learned parameters beyond the shared hop bound and context budget. The method provides a checkable condition under which predefined seed-local answer-bearing render units are preserved in a greedy bounded-context prefix. We evaluate the approach on Common Information Model (CIM) network models exchanged through the Common Grid Model Exchange Standard (CGMES). On two budget-binding CGMES encodings, naive descriptions-first rendering retains local evidence for every single-hop item but only 0.12 and 0.00 of multi-hop items, whereas seed-anchored rendering retains all such evidence. On a preregistered fresh 100-item bank from the SmallGrid topology family, accuracy rises from 0.450 to 0.970 under a fixed 8,000-character context budget. Under a common retrieval and rendering pipeline, the standards-native seed-anchored graph matches or exceeds extracted graph representations produced by LightRAG, Microsoft GraphRAG, and HippoRAG, while avoiding LLM graph-construction tokens. The results are specific to the evaluated CIM/CGMES models, reader, and context budget; they concern budget-bounded retrieval rather than general question answering.

Dense document images often contain many fine-grained visual and textual entities whose relevance depends on a user query. Standard vision-language retrievers encode cropped regions with a single vector, which can mix distinct entity signals and obscure the evidence needed for fine-grained retrieval. We call this failure mode Semantic Dilution and quantitatively show that it degrades entity-level retrieval as a function of entity density. To mitigate it, we propose SAGE, a training-free framework that parses semantic entities from dense document images, represents them as hierarchical graph nodes with multi-vector embeddings, and retrieves query-relevant evidence through iterative entity-level subgraph matching. We also introduce DEAR, a dataset of 1,055 query–image pairs sourced from product detail pages, where each query requires retrieving and comparing multiple fine-grained entities from visually dense inputs across four question types of increasing complexity. Experiments show that SAGE substantially reduces semantic dilution and outperforms patch-level and OCR-based retrieval baselines on DEAR, achieving a Recall@3 of 0.849 and a generation score of 2.746 on multi-entity visual comparison queries. Our code is available at https://github.com/All4Nothing/SAGE.

Democratizing data access through natural language is a crucial goal for modern enterprises, but the practical adoption of Text-to-SQL is critically hindered by real-world complexities: 1. Obscure and large database schemas, 2. Ineffective retrieval of relevant tables and columns due to structured setting of schemas and vague user query, 3. Generation of syntactically or logically flawed SQL due to a lack of robust validation and correction mechanism. To address these systemic challenges, we introduce Reflect-SQL, a novel framework for Text to SQL, grounded in multi-stage self-reflection approach to develop understanding of obscure schema using a knowledge base, setup a process for effective retrieval and system to generate syntactically/semantically SQL. Instead of a single-pass attempt, our system employs an LLM-as-a-judge driven scoring mechanism within interconnected feedback loops to iteratively refine the results at every stage. A feedback-driven retrieval loop refines the user's natural language query, while a synthesis loop validates and corrects the SQL and finally, an entailment loop optimizes the end-to-end process and continuously enriches the knowledge base. By integrating these layers of reflection, Reflect-SQL bridges the critical gap between user intent and complex data. On the challenging BIRD benchmark, our framework achieves an execution accuracy of 72.03%, significantly outperforming state-of-the-art baselines, demonstrating a major leap in reliability for enterprise applications.

In product entity resolution, relationship definitions constantly evolve with business needs, yet adapting to each change traditionally requires slow, costly human annotation that is often noisy and carries no reasoning. Large vision-language models (VLMs) prompted zero-shot can adapt to a new definition immediately and supply the reasoning that human labels lack, but their cost and latency are prohibitive at production scale. We present MERGED, a distillation framework that transfers not just labels but structured reasoning from large teacher VLMs into a compact 7B-parameter student, requiring no human annotation for training. Multiple teachers label each product pair and articulate the reasoning behind their decision: agreement pairs supply supervised fine-tuning, while disagreements are resolved by a meta-judge into preference pairs for Direct Preference Optimization. Evaluated against human-labeled ground truth on a multilingual e-commerce dataset, the resulting student improves PR-AUC by 13.79% over the same backbone trained on human labels and surpasses the larger Qwen2.5-32B-VL baseline by 6.32% at 6x lower cost, while also yielding tighter label-reasoning consistency (over 10% above Qwen2.5-32B-VL). Moreover, re-applying MERGED from an existing checkpoint adapts to a new relationship definition with only 10K samples, improving PR-AUC by 6.97% over zero-shot and outperforming from-scratch training. MERGED enables rapid adaptation to evolving relationship definitions, supporting a new one in days rather than months, at a cost and latency suitable for large-scale industrial deployment.

Embedding-based code retrieval is a core component of coding agents and retrieval-augmented code generation, where retrieving correct code matters more than retrieving lexically similar code. Existing code-retrieval benchmarks do not plant controlled, execution-verified single-edit variants of each query's canonical implementation in the search pool, leaving the question of whether embeddings can functionally discriminate correct from near-clone-but-incorrect code unanswered in a retrieval setting. Resolving this requires a benchmark whose search pool itself contains the relevant counterfactuals – execution-verified buggy variants near-identical to each canonical – so that a retriever's rank ordering can be directly tested for functional discrimination rather than topical or identity overlap. We introduce ExecRetrieval, 939 Python tasks each paired with one execution-verified canonical implementation and up to four execution-verified buggy distractors, each generated by a mechanical mutation making a single targeted edit, and evaluate 23 dense embedding configurations plus BM25 under provider-native invocation with paired McNemar tests and query-level bootstrap intervals. With near-clone counterfactuals in the pool, the top hosted system reaches exec@10 = 1.00 but only exec@1 = 0.331; rank-1 misses are paired buggy variants 91.5-99.4% of the time across the four leading systems, and the canonical scores below at least one of its four paired distractors in 67-78% of queries on the leading systems. The full dataset, execution oracle, embedding matrices, environment snapshot, and pairwise statistical tests are released at the URL in Appendix D.