Papers

Mobile AI acts as a visual oracle, empowering users to snap a picture of something and ask for information. Snap-and-ask retrieval is now one of the most common entry points for mobile AI, yet photos are often blurry, while text questions may be short or mistyped. Existing benchmarks only test on clean inputs or do not isolate paired robustness in snap-and-ask retrieval. Therefore, we introduce SnapBench, the first paired benchmark for robust snap-and-ask multimodal retrieval, spanning 1,145 queries, 9,085 gallery items under 53 controlled corruption conditions with human annotations. We evaluate 16 multimodal retrievers, covering dual-tower encoders and embedding-based VLMs. Results show that image corruptions substantially degrade retrieval, while text corruptions mainly affect text-only retrieval and have limited impact on joint retrieval. Clean image-only retrieval often outperforms joint retrieval, indicating the coarse-text drag and the lack of cross-modal fallback under noisy inputs. SnapBench provides a controlled testbed for evaluating robust retrieval in snap-and-ask scenarios. We further propose MOOR (Modality-anchored, Outlier-aware, Optimal Reweighting), a simple adaptive fusion approach, highlighting the need for reliability-aware modality calibration in snap-and-ask retrieval.

Vision-language retrieval with CLIP-style dual encoders achieves strong cross-modal performance, yet practical accuracy often hinges on localized semantic distinctions where top-ranked near misses differ from the true match by a single critical detail. Hard-sample mining can select confusable candidates but cannot construct corrected counterparts; synthetic augmentation can generate novel samples but, without conditioning on actual model failures, targets irrelevant dimensions of hardness. We observe that a top-ranked false positive is a counterfactual scaffold—sharing most of the query's semantics while differing in a localized failure-causing residual. Minimally correcting this residual yields a hard positive of the ground truth in the same modality; the corrected and unedited versions form a hard negative pair that straddles the decision boundary, producing complementary pull–push supervision. We introduce RePair, guided by three principles—Validity, Minimality, and Locality—which mines false positives bidirectionally, applies LLM-guided counterfactual editing, and trains with a local hard-pair contrastive objective. On Flickr30K and COCO30K, RePair outperforms controlled augmentation baselines with only 107K synthetic samples—26%–75% fewer than comparable methods—confirming failure-conditioned repair is more data-efficient than error-agnostic augmentation.

Off-policy evaluation (OPE) of ranking policies is challenging be- cause selecting and ordering multiple items from a candidate set makes the number of possible rankings grow combinatorially with the number of candidates and the ranking length. Consequently, Inverse Propensity Scoring (IPS), whose importance weight is the full-ranking probability ratio under the evaluation and logging policies, can have excessive variance. Independent IPS (IIPS) and Reward Interaction IPS (RIPS) reduce variance by imposing fixed assumptions on how users browse rankings, but may introduce bias when those assumptions mismatch actual behavior. Adaptive Inverse Propensity Scoring (AIPS) addresses this trade-off by adap- tively marginalizing importance weights over the actions that affect each position-wise reward. It attains minimum variance within a class of unbiased IPS-based estimators when the true user be- havior model is observed. However, its estimation accuracy may still degrade for longer rankings, and AIPS does not use a reward model for residual correction. We propose Adaptive Doubly Robust (ADR), which combines adaptive importance weighting with re- ward regression through a control-variate correction. We establish its unbiasedness when the true user behavior model is observed and characterize a sufficient condition under which it reduces vari- ance relative to AIPS. Across synthetic experiments with 10,000 simulations per condition, ADR improves mean squared error over AIPS and conventional ranking OPE estimators across a range of logged-data sizes and ranking lengths.

Graph collaborative filtering relies on item–item graphs whose edges are used for positive smoothing, under the implicit assumption that stronger edges encode more of the same relation as weaker ones. We show that this assumption fails for a practically important class of graphs: those whose edge weights come from a choice model. On such graphs, strong and weak edges encode qualitatively different relations, which we call an edge spectrum. Specifically, strong edges concentrate on the in-slate competitors of clicked items, exactly the pairs that the within-slate ranking gradient pushes apart, while weak edges do not. We formalize this as a sign mismatch between the smoothing operator and the ranking gradient, and prove that co-click graphs cannot exhibit the same misalignment by construction. This diagnosis explains three empirical observations on MIND and EB-NeRD: (i) drop-in choice-derived operators do not beat co-click, despite indexing structurally distinct neighborhoods; (ii) uniform scalar fixes (sign flip, in-slate margin loss) fail predictably, because the misalignment lives in the graph, not in the loss; (iii) only edge-magnitude-aware operators, with the regime boundary located by the diagnosis rather than by tuning, recover the predicted ordering. The neighbor cutoff $k$ is therefore a semantic switch, not a sparsification hyperparameter. Our claim concerns which interventions fail or succeed and why, not absolute headline gains, which the diagnosis itself predicts to be small under the attenuated propagation channel we observe. We turn the diagnosis into a reusable protocol practitioners can run before deploying any choice-derived item-side operator. Code: https://github.com/kyomusso/Edge-Spectrum-in-CF.

Large audio-language models (LALMs) demonstrate growing music-understanding capabilities, but whether their responses are grounded in acoustic evidence remains unclear. Musical language often involves abstract concepts whose acoustic evidence is difficult to define and evaluate precisely. We introduce MusicGroundingBench, a controlled benchmark of algorithmically generated piano audio with exact symbolic alignment, comprising three-note and two-bar settings. We evaluate two complementary capabilities: grounding, which localizes the acoustic evidence for a musical query, and understanding, which answers questions about the same excerpts. Our experiments show that cross-modal fine-tuning enables models to learn each capability, but adding grounding supervision does not consistently improve understanding across backbones. We further test whether understanding requires listening through audio-ablation controls that remove or replace the input audio, and use attention analysis to examine whether grounding supervision shifts attention toward note boundaries. Meanwhile, the two evaluated LALMs show limited zero-shot grounding even for basic musical concepts, highlighting grounded music understanding as an important open challenge.

Most agentic question answering (QA) systems do an important part of their semantic work at the worst possible time: every time someone asks a question. When a corpus contains revisions, drafts, revocations, deletions, and sources with different levels of authority, the model must reconstruct the governed current state on every read - then throw that work away and repeat it on the next query. This is a bit like a database that rebuilds a materialized view every time someone reads from it. We present ingest-time fact compilation, an architecture that performs this work when corpus data is ingested or changed. Raw passages are rephrased into self-contained facts; rules governing revisions, deletions, effective dates, and source trust are resolved once; and the resulting state is stored as typed records carrying source and revision provenance. At query time, an inexpensive model reads the compiled record instead of reconstructing it from noisy candidates. In a controlled synthetic experiment across five seeds, the same low-cost model produced the correct value, source, and revision in only one of 30 trials under query-time reconstruction, but in all 30 trials from the compiled substrate, at 12.89 times lower mean read cost per question. On simpler revision questions both architectures were exact, but the compiled path used 21.6 times fewer tokens. A separate test found that fact rephrasing roughly halved verbose Federal Reserve dialogue while preserving high source entailment, but left concise Wikipedia prose essentially unchanged. These results support a narrow but practical claim: resolving a corpus state once can make subsequent QA cheaper and more reliable for inexpensive models. We release the open source, MIT-licensed implementation and experimental artifacts.

Reasoning-intensive retrieval remains difficult for small models. Compact public ColBERTs are usually trained on general-purpose corpora and underperform reasoning-tuned 150M+ baselines on BRIGHT~\cite{bright} by several nDCG@10 points. However, no public reasoning-tuned ColBERT exists at edge scale. We introduce SmallReason-ColBERT, a 32M late-interaction retriever that closes much of this gap with three components: a varied-length contrastive warmup on ReasonIR-VL, a hard-negative contrastive polish on merged ReasonIR-HQ and BGE-Reasoner data, and a single-layer per-query-token importance head trained on top of the frozen base. The head is trained with an un-normalised weighted MaxSim score and evaluated with its length-normalised form. In a controlled re-training, replacing this training objective with the symmetric normalised score causes the loss to stall and costs $3.59$ nDCG@10. The full recipe reaches 21.41 mean nDCG@10 on BRIGHT, within $1.21$ of the 150M Reason-ModernColBERT (22.62) and above all $\le 33$M ColBERTs we evaluate. Through ablations on capacity, initialisation, and score variants, we further show that the learned head outperforms fixed IDF weighting and that simply thresholding the learned gates is harmful. https://github.com/DataScienceUIBK/SmallReason-ColBERT

Recent topic models leverage pretrained embeddings, but neural architectures produce latent representations without grounding in specific texts, and clustering-based pipelines assign representative documents only post hoc, relying on absolute distances distorted by hubness and anisotropy in high-dimensional spaces. We introduce MARETopic, a training-free framework that casts topic discovery as rank-based prototype selection. After projecting embeddings onto a low-dimensional manifold, MARETopic builds ranked lists encoding ordinal neighborhood structure. A greedy algorithm selects exactly K exemplar documents, real corpus texts, whose neighborhoods cover the corpus. Two variants share this criterion. MARETopic$_\text{Corr}$ scores candidates with a query performance predictor and a rank correlation measure, leading Purity and NMI on the two benchmarks with the most categories, ahead of both neural and clustering-based topic models. MARETopic$_\text{Diff}$ scores them with a rank-based diffusion matrix, needs neither measure, and runs 1.7 to 1.9 times faster. Without a single gradient update, MARETopic leads topic coherence on two of three datasets. A novel inter-topic Maximal Marginal Relevance step raises vocabulary diversity at little cost in coherence. Our code is available at https://github.com/thcastilho/maretopic.

Qualitative research is widely used in the human and social sciences, characterized by a deep understanding of phenomena through the interpretation of meanings and contexts. Among qualitative data analysis methods, content analysis stands out as a consolidated technique, which allows for the systematic description and interpretation of textual contents. However, as data volume increases, the time required for organization, reading, and categorization becomes a significant challenge, potentially delaying research development. Therefore, this work aimed to develop a web application to support content analysis, based on Bardin's methodology, targeted at academic researchers. The methodology adopted a mixed approach, combining bibliographic research on content analysis with semi-structured interviews with four experienced researchers, aiming to identify real needs and requirements. Based on these inputs, the Cortex software was developed to assist the researcher in the pre-analysis and material exploration stages, generating suggestions for indices, indicators, and categories with full traceability to original documents. The results demonstrate that Cortex is capable of guiding the researcher through the entire methodological workflow, from corpus configuration to results exportation, acting as a methodological collaborator without replacing the researcher's interpretative autonomy.

An LLM agent calling a production API cannot distinguish a query that matched nothing from a query the server did not understand. Both return HTTP 200 with a parsable body, no exception to catch and no field to branch on. We ask what predicts which one occurred, and what it does to the agent. Auditing 721,320 parameters across 2,501 independently published OpenAPI documents, we find that 7.5% declare an enumeration and 15.2% declare any machine-checkable constraint at all, while 40.1% of documents state at least one constraint in prose that their schema does not encode. Executing 219 schema-derived perturbations against live commercial endpoints from 27 vendors, reached through a single aggregation layer (Monid) that publishes a schema and returns a run identifier for every call, we find that constraint form, not vendor identity, predicts honesty: machine-checkable constraints yielded an honest error in 111 of 111 cases, prose-only constraints failed silently in 44 of 61 (p = 2e-13). Twelve models across eight families then met these endpoints on ordinary tasks. A vocabulary that the description merely exemplifies was missed by every model on 88 of 88 attempts, while vocabularies written out in full were used correctly 88 to 91% of the time. Running the full agent loop, models detected the resulting silent failure in 12% of cases, repaired it in 0%, asserted a false negative to the user in 41%, and invented a figure in 12%. Promoting the vocabulary into the schema removes the failure, from 88 of 88 to 0 of 89. The fix is one line of schema rather than a better model. Code, schemas, perturbation sets, agent transcripts and per-call run identifiers are released at https://github.com/Jasper0122/silentprobe.

Industrial recommenders give new content initial views through budgeted exploration, then use early performance to decide further delivery. On many short-video platforms, exploration is the primary way new videos reach viewers. Viewer-side tests measure consumption; the published budget objectives we review omit creator response. We analyze four experiments on a major short-video platform. An eight-month creator ablation finds production exploration raises videos posted per creator by 8.55% and creators posting at least once by 7.10% relative to a minimal floor. A budget-matched reallocation raises creator participation with no detectable short-run viewer-side change. A year-long viewer ablation finds 1.74% more video views but 2.13% less view time. A delivered view creates immediate feed value, can trigger organic take-up, and can induce creator supply. Take-up and supply replenish a shared corpus, creating two measurement limits. Viewer-side A/B tests cancel the corpus effect when both arms consume the same corpus. Giving each arm its own corpus avoids cancellation, but turnover still controls the horizon. If the corpus turns over at rate w per posting cycle, a t-cycle experiment expresses at most wt of the eventual corpus effect. More users reduce noise but do not speed turnover. Before the corpus path visibly bends, data cannot distinguish a modest fast effect from an arbitrarily large slow one, so a valid confidence interval may lack a finite upper endpoint. As predicted, the three-week co-diverted experiment cannot determine the sign of the eventual corpus effect. Within the window, it identifies the direct feed effect, and an exploratory cohort analysis detects organic lift after exploration ends. The experiments establish a positive creator response, measure the gross corpus flow visible within three weeks, and show the design and duration needed to identify total value.

Online platforms increasingly rely on multimodal recommender systems to rank products, media, and other Web content. Existing methods usually inject visual and textual features into item representations or build homogeneous graphs from modality-level similarity, but the resulting signals can remain misaligned with the recommendation objective. We study this semantic gap from a knowledge-integration perspective: multimodal content should be interpreted together with user behavior before it is used to construct recommendation graphs or adjust rankings. We propose AgentMMRec, an agent-based multimodal recommendation framework with two coordinated roles. The Integrator Agent infers behavior- and multimodal-aware user preferences and item properties from training interactions and item content, then stores them in a reusable knowledge memory. The Utilizer Agent consumes this memory to refine modality-specific item-item graphs, construct behavior-aware homogeneous graphs, and rerank candidate lists under a frozen evaluation-time memory. This design differs from direct LLM feature augmentation and pure LLM reranking because the generated knowledge is first converted into graph structure and model representations before recommendation. Experiments on three Amazon multimodal recommendation datasets show that AgentMMRec consistently improves Recall and NDCG over recent multimodal baselines, remains effective under sparsity and item cold-start settings, and can transfer its constructed knowledge to existing backbones.

Workout recommender systems aim to help gym users complete effective and engaging training sessions. However, recommending exercises alone is insufficient, as a practical system must also determine appropriate sets, repetitions, and training loads, while adapting to user behavior such as skipping exercises. Existing approaches typically consider only a subset of these factors, limiting their applicability in real-world settings. In this paper, we extend workout recommendation from exercise selection to full workout prescription. We propose a reinforcement learning (RL)-based framework with four environments: exercise-only and full-prescription settings, each with and without skip-based interaction. The full-prescription environments recommend exercises, sets, repetitions, and load, while the skip-enabled environments use user skipping behavior for online personalization. Experiments with synthetic users show that modeling the full prescription task leads to higher rewards and greater user engagement than exercise-only recommendation, highlighting the importance of realistic workout planning in personalized gym recommender systems.

FISICA is a body-assessment and recommendation service running in production. One standing session with two photographs returns foot-loading measures, posture coordinates, a driven 3D avatar, a visual report, and ranked shoe and exercise candidates. Measurement comes from a purpose-built scale carrying 634 force-sensitive elements on a 1 cm grid and four load cells, and a rule-based evaluator controls every recommendation while a language model only explains the stored result. The method contribution is the avatar. Instead of mapping a measured angle onto a rig through a tuned gain, we measure the avatar with the same function used on the subject and solve until the two agree, on a sampling-invariant spinal metric that separated a normal from a kyphotic record by 7.2 degrees against 0.9 degrees for a single-joint formulation. In production, general APIs respond at a 0.023 s median, plantar-pressure analysis at 0.45 s, and recommendation at 2.16 s to 2.26 s with the rule-based portion under one second in every trial. The served keypoint graph reaches 0.960 PCK@0.2 on public data, and the catalog holds 699 shoes with 10,500 typed facts. An approved study supplies the radiographic reference for the validation still ahead.

Many Internet of Things (IoT) services provide Representational State Transfer (REST) APIs, which require client developers to implement applications that conform to the corresponding API specifications. When client programs contain API misuse, developers debug them based on error responses. However, such responses are often insufficient for identifying the root cause, requiring developers to repeatedly communicate with the server. Retrieval-Augmented Generation (RAG) is a promising approach for providing large language models (LLMs) with external knowledge. However, in automated repair of REST API misuses, it remains unclear how specifications should be stored in a RAG database. This study evaluates how different configurations for organizing API specifications affect RAG-based repair of REST API misuse. We constructed 11 RAG configurations with different database structures and compared their repair rates with a baseline method. For evaluation, we used REST API misuse cases collected from real-world repositories. The results show that, in the studied datasets, the baseline method achieved a repair rate of 54.3%, whereas a RAG-based method using four databases achieved a maximum repair rate of 88.6%. These results indicate that organizing specifications according to version and content type can be an effective design choice for RAG-based REST API misuse repair.

Deploying large language models for legal question answering raises challenges that general-purpose leaderboards do not capture, particularly for low-resource languages and under hard operational constraints. We report on building and operating a retrieval-augmented (RAG) legal assistant for Uzbek that must run in two regimes: a managed cloud service that maximizes answer quality within a per-token cost ceiling, and an on-premises deployment for clients whose legal data may not leave their infrastructure, restricting us to open-weight models on limited local hardware under latency constraints. Because no evaluation existed for this setting, we build two domain benchmarks: a retrieval benchmark of 178 expert-annotated legal queries with gold provision spans, and an end-to-end benchmark of 504 expert-curated question–answer pairs scored by an LLM judge whose ratings we validate against human judgments and against an independent-family judge. Applying these benchmarks under each regime, we find the open-versus-proprietary gap is small and cheaply closed by fine-tuning. Therefore, we train UTE-1, which is a state-of-the-art text embedder among open models for Uzbek. We also demonstrate that closing the performance gap via fine-tuning is both impractical due to the intensive hardware demands of long-context legal Q&A and unnecessary, given that legal acts change frequently. We support this by reporting a negative result from a QLoRA experiment. We distill practical guidance for similar deployments, drawn from a system serving real users in production. We release our benchmarks, evaluation code and the fine-tuned embedder (UTE-1) \href{https://metric-ai-lab.github.io/Uzbek-Legal-RAG/}{at this https URL} to support future work on low-resource legal NLP.

The online culinary ecosystem is increasingly populated by recipe content generated, modified, or summarized by Large Language Models (LLMs). While often plausible, such outputs may contain hallucinated ingredients, misrepresented quantities, or culturally implausible combinations, limiting their suitability for downstream applications and knowledge graph construction. In this paper, we present a semi-automated soundness assessment workflow for validating structured recipe data extracted and augmented by LLMs from informal culinary sources. Developed as part of FKG(.in), a knowledge graph of Indian food, the pipeline identifies and addresses common failure modes, including structural inconsistencies, semantic and logical incoherence, and deviations from the source text, through a multi-stage process combining formal grammars, vocabulary-based checks, statistical heuristics, Set Transformer-based coherence modeling, and retrieval-based verification. Although evaluated on Indian recipes, the proposed methods are applicable to broader multilingual and multicultural culinary domains. We provide a practical, auditable, and application-agnostic framework for validating LLM-augmented recipe data, thereby strengthening the foundations of machine-readable food knowledge infrastructures in the era of LLM-generated content.

Generative recommendation retrieves items by autoregressively generating semantic identifiers, but beam search may discard a target before its complete identifier is generated. Our preliminary analysis across three benchmarks shows that most missed targets are pruned within the first two decoding steps, highlighting the importance of early prefix retention. However, retaining a target's first-token branch alone is insufficient if its continuation is pruned at the next step: aligning only the first-token distribution improves first-token survival but yields little gain in complete-path retention. This observation motivates joint supervision of early branches and their continuations. We propose Prefix Alignment with Temporal History (PATH), which aggregates transition statistics from the training corpus over recent interactions with exponential decay to construct history-conditioned two-token prefix targets. PATH aligns the model's joint predictions with these targets through forward KL divergence, using a chain-rule decomposition that enables unbiased Monte Carlo estimation. At inference time, PATH reuses the transition statistics for pointwise mutual information (PMI) calibration, reranking completed candidates relative to global prefix frequency. Experiments on Beauty, Instruments, and Yelp demonstrate the effectiveness of PATH, showing consistent improvements in recommendation performance and higher full-SID survival rates.

Exogenous events can temporarily change the relevance of items in recommender systems, but these shifts are often not visible in historical interaction data until after users have already responded. In book recommendation, movie adaptations provide a clear example of such events: the release of a movie based on a book can temporarily increase attention to the source text and change its relevance for some readers. We examine this phenomenon using a large Goodreads dataset matched to movie release dates. We find a clear spike in readership around the release month, and then we evaluate existing recommendation models to understand how they rank movie-adapted books around the movie release date.

The advances in visual information modeling and representation during the last decades are remarkable, mainly supported by Convolutional Neural Networks, Transformer-based, and Foundation Models. Despite this progress, critical challenges regarding the nature of similarity assessment and model transparency have been neglected. A primary concern is the Geometric Gap, where traditional pairwise measures fail to capture the intrinsic geometry of the dataset manifold. Furthermore, the Interpretability Gap persists, as representations often lack alignment with human cognition. Therefore, how to provide interpretability to representations while maintaining low dimensionality and high effectiveness in downstream tasks remains an open challenge. In this paper, we propose a novel unsupervised framework that integrates Manifold Learning strategies with Rank-based Interpretable Graph Embeddings. Our approach effectively bridges these gaps by first characterizing the contextual information of the dataset through manifold analysis and subsequently generating sparse, self-explainable embeddings. The proposed approach employs a flexible formulation, allowing different Manifold Learning and Representation Learning strategies. Extensive experimental evaluation across diverse datasets and features demonstrates that our Context-Aware representations not only provide intrinsic interpretability and dimensionality reduction but also maintain or enhance effectiveness in downstream tasks, specifically in image retrieval and semi-supervised classification using Graph Convolutional Networks (GCNs).

In a data-driven world, efficiently organizing and mapping relationships between objects is crucial. Graphs are powerful tools for modeling these connections, being widely used in social networks, telecommunications, and biology. However, graph-based methods often face high computational costs, particularly in memory and space usage. To address this, graph embedding techniques, also referred to as Network Representation Learning, encode graph information into lower-dimensional representations while preserving structural aspects. Traditional methods, however, lack interpretable dimensions. RaDE (Rank Diffusion Embedding) introduces a new approach using rank-based information, with a key step being the selection of a representative subset of nodes to provide interpretability for its dimensions and improve retrieval tasks. Despite its potential, RaDE's original proposal did not fully explore the effectiveness of representative subset selection across different classes or evaluate embeddings in tasks like classification and clustering. Inspired by RaDE, this work introduces GRaCE (Graph and Rank-based Contextual Embeddings), a fully unsupervised framework that generates interpretable embeddings by leveraging robust rank-based measures for representative subset selection and node embedding. GRaCE surpasses RaDE and Original Features across diverse datasets, including textual and image collections, excelling in retrieval, classification, and clustering tasks, considering state-of-the-art Transformer models as feature descriptors and Graph Convolutional Networks models in classification tasks.

We report an exploratory, single-seed ablation of TALH (Adaptive Latent Hybrid), a decoder-only language model with parallel Multi-head Latent Attention (MLA) and a custom recurrent state-space (SSM) branch. Five variants, spanning 117–217M estimated active parameters per token, are trained from scratch on a FineWeb sample for the same number of optimisation steps and tokens. In this specific setup, removing the SSM branch gives the largest degradation in validation perplexity (MLA-only PPL 315), whereas removing MLA has a much smaller effect (SSM-only PPL 239). A dense-FFN hybrid obtains PPL 231, compared with 240 for the tested top-2 ternary-MoE hybrid, while using 3.87 GB less peak training memory. We also preserve a preliminary Apple M3 timing observation: among the five unoptimised implementations, MLA-only has the flattest measured time-to-first-token curve from 512 to 2,048 prompt tokens, although the dense Transformer is much faster in absolute terms. Because the runs are single-seed, parameter counts are unmatched, the evaluation stream may overlap the training source, and raw repeated timing records are unavailable, these results support implementation-specific hypotheses rather than general conclusions about MLA, SSMs, or mixture-of-experts models.

Retrieval-augmented question answering is increasingly used to help people navigate public-benefits eligibility, yet the documents these systems retrieve from are written in agency register while intended users often ask questions in plain, informal, or non-native English. We show that this register mismatch can turn a high-performing retrieval system into an inequitable one. We construct a controlled benchmark of 51 publicly documented federal benefit-eligibility rules and 25 information needs, each phrased in both agency register and plain user register while keeping the gold passage fixed. Across BM25, TF-IDF, and a term-graph retriever, formal-register evaluation is nearly perfect (Recall@5 96-100%), but plain-register retrieval collapses (Recall@5 36-44%). For BM25, Recall@1 falls from 84% to 16% and Recall@5 from 100% to 44%, a 56-point equity gap on identical information needs. We trace the mechanism to a measurable vocabulary gap: formal queries share 0.63 of their content terms with the gold passage, while plain queries share only 0.11, a 5.9x reduction. A deliberately simple, auditable plain-to-formal lexicon bridge recovers much of the failure, lifting plain-query BM25 Recall@5 from 44% to 80%. The contribution is not a new retriever; it is an evaluation protocol, benchmark, mechanistic diagnosis, and transparent mitigation for a high-stakes social-impact failure mode that standard retrieval evaluation hides.

Effective skill identification is crucial for LLM agents to reuse task-solving experience for complex task execution, which however remains challenging as agents must accurately identify useful skills from a large skill gallery and invoke them on demand. Existing studies have explored various strategies for skill retrieval and invocation, but largely overlook the natural discrepancy between task descriptions and skill documents. To be specific, our empirical study reveals that task descriptions mainly specify task objectives, whereas skill documents describe the required capabilities and procedural guidance for achieving the objectives. Such a so-called Task–Skill Misalignment (TSM) problem would cause desirable skills to be neglected during retrieval or remain unselected during invocation, thus hindering task execution. As a remedy, inspired by human prospective cognition, we propose SkillDreamer, a prospective skill identification framework to mitigate TSM. In brief, SkillDreamer first infers execution-required capabilities and imagines their realizations as pseudo skills, thereby translating task objectives into execution-oriented guidance for skill retrieval. As execution proceeds, SkillDreamer reevaluates the capabilities required by emerging subtasks to determine which skills should be invoked, thereby aligning skill invocation with evolving execution needs. Extensive experiments on three representative agent skill benchmarks not only verify the effectiveness of SkillDreamer in mitigating TSM, but also demonstrate its generalizability across diverse retrievers and LLM agents. The code will be released upon acceptance.

Matching users to interest categories at scale is central to personalized shopping, but the task is challenging in large e-commerce platforms, where label spaces continually evolve and user-interest signals are sparse and long-tailed. Autoregressive language models are appealing because their world knowledge and semantic priors over descriptors generalize across extreme label spaces and accommodate multiple valid label assignments. Yet under teacher-forced fine-tuning, inference-time predictions become part of the conditioning context: early errors steer later outputs toward co-occurring labels, over-generating near-correlates and missing unrelated true interests. We present MERIT, a framework for user-interest propensity modeling that mitigates this exposure bias through a self-correction objective. A permutation-invariant multi-target loss over shuffled mixtures of gold and mined hard-negative labels exposes the generator to erroneous prefixes while preserving the efficiency of teacher-forced training. This training objective concentrates supervision at classification positions, yielding propensity-aligned hidden states powering a lightweight scorer for bidirectional retrieval (interests for users and users for interests). On a proprietary e-commerce dataset with 250k+ interest categories, MERIT improves global recall by at least 11.9% and average Hit@k by 6.1%. In production A/B tests, it achieves +0.26% gain in user conversion.

Technical operations teams resolve large volumes of incidents by synthesizing fragmented evidence from ticket text, historical cases, system logs, and technical documentation. Existing automation often relies on monolithic generation without explicit evidence modeling or provenance, making outputs difficult to verify when critical signals are sparse across sources. We propose ASTRA, an agentic system for ticket resolution in which a central orchestrator coordinates three specialist information-gathering agents and drives a judge-orchestrator refinement loop to produce evidence-backed troubleshooting reports. TicketSimilarityAgent retrieves relevant historical precedents through dense retrieval and LLM reranking; LogAgent distills hundreds of thousands of log lines into structured, quote-grounded findings using deterministic filtering and constrained LLM analysis; and DomainKnowledgeAgent retrieves relevant technical knowledge via the Model Context Protocol (MCP). Their outputs are transformed into a claim-evidence representation linking each claim to a verbatim source passage, assigning a support level, and preventing cross-attribution. A JudgeAgent scores the report on five criteria, while the OrchestratorAgent converts low scores into targeted follow-up queries for bounded iterative refinement. Evaluated on 987 real-world telecom fault tickets across seven product lines, ASTRA achieves a mean quality score of 4.13/5.0, with 59.9% of reports identifying the fault area at the component-family level or better. Relevance and Clarity scores are 4.88 and 4.94, respectively, while fabricated technical details remain below 3% of error cases. Stratification by fault type reveals that hardware faults remain substantially harder than software or configuration faults (Cohen's d=0.80), pointing to a fundamental limitation of text-based evidence channels for hardware fault diagnosis.

Institutional investors search visually dense pitch decks, board packs, and diligence materials that change hourly near deal closing. OCR followed by figure verbalisation is costly to refresh at this scale and can lose chart detail. We present PULSAR, a production vision-first retrieval system deployed at Mubadala Investment Company. PULSAR indexes page images with a frozen ColPali-style backbone and uses a pooled two-stage late-interaction index: compact page summaries support initial retrieval, followed by exact MaxSim rescoring over a finer pooled representation. On ViDoRe V3, this design reduces median vector-search latency by 15.1 times against an unpooled configuration with less than 0.01 absolute NDCG@10 and Recall@10 loss; production median vector-search latency is 156 ms. Under concurrent load, the pooled index sustains approximately 88 times higher QPS than an unpooled index. The event-driven ingestion path is estimated to be approximately 20 times cheaper per page than the OCR+verbalisation baseline it replaced. Since March 2026, PULSAR has served 78 thousand documents and approximately 2.4 million pages across more than 3,000 deals. At the production top K, it more than doubles answer-fact recall over the OCR+verbalisation baseline.

Legal decisions on asylum applications consist of long, complex, and heterogeneous documents, covering narrative applicant interviews, original decisions, and additional supporting materials. If an application is rejected, a critical question in processing an appeal is whether the credibility of the information in the original application was a factor that determined the original decision. In this paper, we present the QUEST system (Query and Extraction System for Topics) to extract and identify factors relating to credibility assessments in two datasets of Danish asylum application appeals. QUEST frames this problem as an information retrieval task, combining synthetic query generation, topic extraction, and relevance assessment to identify information related to credibility indicators in appeals board application materials. In addition to standard retrieval evaluation metrics, we propose a new type of domain-specific assessments distinct from the traditional relevance to evaluate the performance of the tested systems with respect to credibility factors. In this way, we obtain insights about how well automatic methods can return answers for different types of indicators appearing in asylum appeals. Our results indicate that there is an increased challenge when estimating performance using credibility-based relevance assessments, thus pointing to the difficulty of the task.

Multimodal sequential recommendation (MSR) improves recommendation by incorporating heterogeneous information such as text, images, and user interactions. However, existing MSR methods often fail to capture user-level preference heterogeneity and dataset-level modality bias, limiting their adaptability across users and datasets. To address this issue, we propose Sequence-Guided Universal Multimodal Prioritization Calculation Framework (SG-UMP), a plug-and-play plugin for enhancing multimodal information processing in MSR. SG-UMP includes a Module Combiner for flexible multimodal processing and a Module Router for dynamic module ordering, enabling adaptation to both user preferences and dataset characteristics. Experiments on four real-world datasets show that SG-UMP consistently improves recommendation performance across different backbones and multimodal settings. The code is available at https://github.com/esemsc-xz524/SG-UMP .

As digital publishers face the challenge of managing massive content catalogs, the ability to effectively embed relevant video within text-based articles has become critical for both monetization and user retention. However, manual selection is impractical for large scale publishers, especially when navigating their own extensive video libraries or the entire global Dailymotion catalog. In this paper, we present the "Contextual Video Matching" system, a solution that automatically matches relevant videos with text-heavy web pages and articles. By leveraging Large Language Models (LLMs) and textual embeddings, we provide a scalable solution for publishers to efficiently combine video content with their articles. We discuss in detail the motivations, architecture, evaluations, and deployment of this system within Dailymotion's production environment. Since its launch, the system has been adopted by hundreds of publishers, significantly increasing user engagement and enriching user experiences with highly relevant video content.