Faster Distributed Inference-Only Recommender Systems via Bounded Lag Synchronous Collectives

Recommender systems are enablers of personalized content delivery, and therefore revenue, for many large companies. In the last decade, deep learning recommender models (DLRMs) are the de-facto standard in this field. The main bottleneck in DLRM inference is the lookup of sparse features across huge embedding tables, which are usually partitioned across the aggregate RAM of many nodes. In state-of-the-art recommender systems, the distributed lookup is implemented via irregular all-to-all (alltoallv) communication, and often presents the main bottleneck. Today, most related work sees this operation as a given; in addition, every collective is synchronous in nature. In this work, we propose a novel bounded lag synchronous (BLS) version of the alltoallv operation. The bound can be a parameter allowing slower processes to lag behind entire iterations before the fastest processes block. In special applications such as inference-only DLRM, the accuracy of the application is fully preserved. We implement BLS alltoallv in a new PyTorch Distributed backend and evaluate it with a BLS version of the reference DLRM code. We show that for well balanced, homogeneous-access DLRM runs our BLS technique does not offer notable advantages. But for unbalanced runs, e.g. runs with strongly irregular embedding table accesses or with delays across different processes, our BLS technique improves both the latency and throughput of inference-only DLRM. In the best-case scenario, the proposed reduced synchronisation can mask the delays across processes altogether.

Real-Time Machine Learning for Embedded Anomaly Detection

The spread of a resource-constrained Internet of Things (IoT) environment and embedded devices has put pressure on the real-time detection of anomalies occurring at the edge. This survey presents an overview of machine-learning methods aimed specifically at on-device anomaly detection with extremely strict constraints for latency, memory, and power consumption. Lightweight algorithms such as Isolation Forest, One-Class SVM, recurrent architectures, and statistical techniques are compared here according to the realities of embedded implementation. Our survey brings out significant trade-offs of accuracy and computational efficiency of detection, as well as how hardware constraints end up fundamentally redefining algorithm choice. The survey is completed with a set of practical recommendations on the choice of the algorithm depending on the equipment profiles and new trends in TinyML, which can help close the gap between detection capabilities and embedded reality. The paper serves as a strategic roadmap for engineers deploying anomaly detection in edge environments that are constrained by bandwidth and may be safety-critical.

Identifying Features Associated with Bias Against 93 Stigmatized Groups in Language Models and Guardrail Model Safety Mitigation

Large language models (LLMs) have been shown to exhibit social bias, however, bias towards non-protected stigmatized identities remain understudied. Furthermore, what social features of stigmas are associated with bias in LLM outputs is unknown. From psychology literature, it has been shown that stigmas contain six shared social features: aesthetics, concealability, course, disruptiveness, origin, and peril. In this study, we investigate if human and LLM ratings of the features of stigmas, along with prompt style and type of stigma, have effect on bias towards stigmatized groups in LLM outputs. We measure bias against 93 stigmatized groups across three widely used LLMs (Granite 3.0-8B, Llama-3.1-8B, Mistral-7B) using SocialStigmaQA, a benchmark that includes 37 social scenarios about stigmatized identities; for example deciding wether to recommend them for an internship. We find that stigmas rated by humans to be highly perilous (e.g., being a gang member or having HIV) have the most biased outputs from SocialStigmaQA prompts (60% of outputs from all models) while sociodemographic stigmas (e.g. Asian-American or old age) have the least amount of biased outputs (11%). We test if the amount of biased outputs could be decreased by using guardrail models, models meant to identify harmful input, using each LLM's respective guardrail model (Granite Guardian 3.0, Llama Guard 3.0, Mistral Moderation API). We find that bias decreases significantly by 10.4%, 1.4%, and 7.8%, respectively. However, we show that features with significant effect on bias remain unchanged post-mitigation and that guardrail models often fail to recognize the intent of bias in prompts. This work has implications for using LLMs in scenarios involving stigmatized groups and we suggest future work towards improving guardrail models for bias mitigation.

FC-MIR: A Mobile Screen Awareness Framework for Intent-Aware Recommendation based on Frame-Compressed Multimodal Trajectory Reasoning

Identifying user intent from mobile UI operation trajectories is critical for advancing UI understanding and enabling task automation agents. While Multimodal Large Language Models (MLLMs) excel at video understanding tasks, their real-time mobile deployment is constrained by heavy computational costs and inefficient redundant frame processing. To address these issues, we propose the FC-MIR framework: leveraging keyframe sampling and adaptive concatenation, it cuts visual redundancy to boost inference efficiency, while integrating state-of-the-art closed-source MLLMs or fine-tuned models (e.g., Qwen3-VL) for trajectory summarization and intent prediction. We further expand task scope to explore generating post-prediction operations and search suggestions, and introduce a fine-grained metric to evaluate the practical utility of summaries, predictions, and suggestions. For rigorous assessment, we construct a UI trajectory dataset covering scenarios from UI-Agents (Agent-I) and real user interactions (Person-I). Experimental results show our compression method retains performance at 50%-60% compression rates; both closed-source and fine-tuned MLLMs demonstrate strong intent summarization, supporting potential lightweight on-device deployment. However, MLLMs still struggle with useful and "surprising" suggestions, leaving room for improvement. Finally, we deploy the framework in a real-world setting, integrating UI perception and UI-Agent proxies to lay a foundation for future progress in this field.

DPSR: Differentially Private Sparse Reconstruction via Multi-Stage Denoising for Recommender Systems

Differential privacy (DP) has emerged as the gold standard for protecting user data in recommender systems, but existing privacy-preserving mechanisms face a fundamental challenge: the privacy-utility tradeoff inevitably degrades recommendation quality as privacy budgets tighten. We introduce DPSR (Differentially Private Sparse Reconstruction), a novel three-stage denoising framework that fundamentally addresses this limitation by exploiting the inherent structure of rating matrices -- sparsity, low-rank properties, and collaborative patterns. DPSR consists of three synergistic stages: (1) \textit{information-theoretic noise calibration} that adaptively reduces noise for high-information ratings, (2) \textit{collaborative filtering-based denoising} that leverages item-item similarities to remove privacy noise, and (3) \textit{low-rank matrix completion} that exploits latent structure for signal recovery. Critically, all denoising operations occur \textit{after} noise injection, preserving differential privacy through the post-processing immunity theorem while removing both privacy-induced and inherent data noise. Through extensive experiments on synthetic datasets with controlled ground truth, we demonstrate that DPSR achieves 5.57\% to 9.23\% RMSE improvement over state-of-the-art Laplace and Gaussian mechanisms across privacy budgets ranging from $\varepsilon=0.1$ to $\varepsilon=10.0$ (all improvements statistically significant with $p < 0.05$, most $p < 0.001$). Remarkably, at $\varepsilon=1.0$, DPSR achieves RMSE of 0.9823, \textit{outperforming even the non-private baseline} (1.0983), demonstrating that our denoising pipeline acts as an effective regularizer that removes data noise in addition to privacy noise.

Microsoft Academic Graph Information Retrieval for Research Recommendation and Assistance

In today's information-driven world, access to scientific publications has become increasingly easy. At the same time, filtering through the massive volume of available research has become more challenging than ever. Graph Neural Networks (GNNs) and graph attention mechanisms have shown strong effectiveness in searching large-scale information databases, particularly when combined with modern large language models. In this paper, we propose an Attention-Based Subgraph Retriever, a GNN-as-retriever model that applies attention-based pruning to extract a refined subgraph, which is then passed to a large language model for advanced knowledge reasoning.

CIRR: Causal-Invariant Retrieval-Augmented Recommendation with Faithful Explanations under Distribution Shift

Recent advances in retrieval-augmented generation (RAG) have shown promise in enhancing recommendation systems with external knowledge. However, existing RAG-based recommenders face two critical challenges: (1) vulnerability to distribution shifts across different environments (e.g., time periods, user segments), leading to performance degradation in out-of-distribution (OOD) scenarios, and (2) lack of faithful explanations that can be verified against retrieved evidence. In this paper, we propose CIRR, a Causal-Invariant Retrieval-Augmented Recommendation framework that addresses both challenges simultaneously. CIRR learns environment-invariant user preference representations through causal inference, which guide a debiased retrieval process to select relevant evidence from multiple sources. Furthermore, we introduce consistency constraints that enforce faithfulness between retrieved evidence, generated explanations, and recommendation outputs. Extensive experiments on two real-world datasets demonstrate that CIRR achieves robust performance under distribution shifts, reducing performance degradation from 15.4% (baseline) to only 5.6% in OOD scenarios, while providing more faithful and interpretable explanations (26% improvement in faithfulness score) compared to state-of-the-art baselines.

Computable Gap Assessment of Artificial Intelligence Governance in Children's Centres:Evidence-Mechanism-Governance-Indicator Modelling of UNICEF's Guidance on AI and Children 3.0 Based on the Graph-GAP Framework

This paper tackles practical challenges in governing child centered artificial intelligence: policy texts state principles and requirements but often lack reproducible evidence anchors, explicit causal pathways, executable governance toolchains, and computable audit metrics. We propose Graph-GAP, a methodology that decomposes requirements from authoritative policy texts into a four layer graph of evidence, mechanism, governance, and indicator, and that computes two metrics, GAP score and mitigation readiness, to identify governance gaps and prioritise actions. Using the UNICEF Innocenti Guidance on AI and Children 3.0 as primary material, we define reproducible extraction units, coding manuals, graph patterns, scoring scales, and consistency checks, and we demonstrate exemplar gap profiles and governance priority matrices for ten requirements. Results suggest that compared with privacy and data protection, requirements related to child well being and development, explainability and accountability, and cross agency implementation and resource allocation are more prone to indicator gaps and mechanism gaps. We recommend translating requirements into auditable closed loop governance that integrates child rights impact assessments, continuous monitoring metrics, and grievance redress procedures. At the coding level, we introduce a multi algorithm review aggregation revision workflow that runs rule based encoders, statistical or machine learning evaluators, and large model evaluators with diverse prompt configurations as parallel coders. Each extraction unit outputs evidence, mechanism, governance, and indicator labels plus readiness scores with evidence anchors. Reliability, stability, and uncertainty are assessed using Krippendorff alpha, weighted kappa, intraclass correlation, and bootstrap confidence intervals.

Efficient Optimization of Hierarchical Identifiers for Generative Recommendation

SEATER is a generative retrieval model that improves recommendation inference efficiency and retrieval quality by utilizing balanced tree-structured item identifiers and contrastive training objectives. We reproduce and validate SEATER's reported improvements in retrieval quality over strong baselines across all datasets from the original work, and extend the evaluation to Yambda, a large-scale music recommendation dataset. Our experiments verify SEATER's strong performance, but show that its tree construction step during training becomes a major bottleneck as the number of items grows. To address this, we implement and evaluate two alternative construction algorithms: a greedy method optimized for minimal build time, and a hybrid method that combines greedy clustering at high levels with more precise grouping at lower levels. The greedy method reduces tree construction time to less than 2% of the original with only a minor drop in quality on the dataset with the largest item collection. The hybrid method achieves retrieval quality on par with the original, and even improves on the largest dataset, while cutting construction time to just 5-8%. All data and code are publicly available for full reproducibility at https://github.com/joshrosie/re-seater.

SecureCode v2.0: A Production-Grade Dataset for Training Security-Aware Code Generation Models

AI assistants produce vulnerable code in 45% of security-relevant scenarios, introducing flaws into production systems at scale. Yet existing secure coding datasets fall short. They lack incident grounding, don't provide the scale modern training requires, and miss the operational security context developers need for production deployments. We present SecureCode v2.0, a production-grade dataset of 1,215 security-focused coding examples that passed structural validation and expert security review. Every example ties to actual documented security incidents with CVE references, provides vulnerable and secure implementations, demonstrates concrete attacks, and includes defense-in-depth operational guidance. The dataset covers 11 vulnerability categories (complete OWASP Top 10:2025 plus AI/ML Security Threats) across 11 languages (Python, JavaScript, Java, Go, PHP, C#, TypeScript, Ruby, Rust, Kotlin, and YAML for infrastructure-as-code). Our quality assurance framework ensures complete incident grounding. Each example includes SIEM integration strategies, infrastructure hardening recommendations (Docker, AppArmor, WAF configurations), and testing approaches using language-appropriate frameworks. The dataset uses a 4-turn conversational structure mirroring actual developer-AI interactions, escalating from basic implementations to advanced security considerations and defense-in-depth guidance. Our contributions: (1) 1,215 rigorously validated examples split into 989 training, 122 validation, and 104 test sets, (2) an automated validation framework ensuring dataset consistency, (3) a 4-turn conversational structure capturing realistic security workflows, (4) comprehensive operational security guidance with SIEM integration strategies, (5) complete language-specific implementation fidelity, and (6) open-source release of data, validation tools, and benchmarking protocols.