Papers.

Shattering the Ephemeral Storage Cost Barrier for Data-Intensive Serverless Workflows

Utilizing Large Language Models for Ablation Studies in Machine Learning and Deep Learning

Decentralized Adaptive Ranking using Transformers

Practical Federated Learning without a Server

Turing.jl: A General-Purpose Probabilistic Programming Language

Probabilistic programming languages (PPLs) are becoming increasingly important in many scientific disciplines, such as economics, epidemiology, and biology, to extract meaning from sources of data while accounting for one's uncertainty. The key idea of probabilistic programming is to decouple inference and model specification, thus allowing the practitioner to approach their task at hand using Bayesian inference, without requiring extensive knowledge in programming or computational statistics. At the same time, the complexity of problem settings in which PPLs are employed is steadily increasing, both in terms of project size and model complexity, calling for more flexible and efficient systems. In this work, we describe Turing.jl, a general-purpose PPL, which is designed to be flexible, efficient, and easy to use. Turing.jl is built on top of the Julia programming language, which is known for its high performance and ease-of-use. We describe the design of Turing.jl, contextualizing it within different types of users and use-cases, its key features, and how it can be used to solve a wide range of problems. We also provide a brief overview of the ecosystem around Turing.jl, including the different libraries and tools that can be used in conjunction with it. Finally, we provide a few examples of how Turing.jl can be used in practice.

Are Your Fairness Metrics Accurate? A Semi-Supervised Approach to Improving Fairness Estimates Under Sample Selection Bias

Exploring Intelligent Dynamic Resource Provisioning for Elastic Massive MIMO vRAN

The computational load in a Virtualized Radio Access Network (vRAN) varies rapidly and massive Multiple-Input Multiple-Output (MIMO) makes it even more dynamic. Existing massive MIMO vRAN systems commonly adopt static resource provisioning and thus often result in either under- or over-provisioning of compute resources. In this paper, we explore the feasibility of a Dynamic Resource Provisioning (DRP) in a massive MIMO vRAN to dynamically allocate Central Processing Unit (CPU) cores for varying computational loads. We investigate a Reinforcement Learning (RL)-based DRP algorithm that provisions optimal number CPU cores in an automated manner. To experimentally demonstrate its viability, we integrate our RL-based DRP into a real-time fully-software massive MIMO vRAN prototype system. Our early experimental results, however, do not show any significant advantage of an RL-based solution over a simple Steepest Gradient Descent (SGD)-based DRP. We believe that this surprising result is largely due to the fact that current open-source vRAN systems cannot sufficiently reflect the highly complex nature of the real vRAN. We further discuss the way ahead, challenges, and opportunities in this research direction.

Designing for an AI-Augmented Journaling Experience: Balancing Guidance and Autonomy for Deeper Emotional Insight.

Flo: A Semantic Foundation for Progressive Stream Processing

Streaming systems are present throughout modern applications, processing continuous data in real-time. Existing streaming languages have a variety of semantic models and guarantees that are often incompatible. Yet all these languages are considered “streaming”–what do they have in common? In this paper, we identify two general yet precise semantic properties: streaming progress and eager execution. Together, they ensure that streaming outputs are deterministic and kept fresh with respect to streaming inputs. We formally define these properties in the context of Flo, a parameterized streaming language that abstracts over dataflow operators and the underlying structure of streams. It leverages a lightweight type system to distinguish bounded streams, which allow operators to block on termination, from unbounded ones. Furthermore, Flo provides constructs for dataflow composition and nested graphs with cycles. To demonstrate the generality of our properties, we show how key ideas from representative streaming and incremental computation systems—Flink, LVars, and DBSP—have semantics that can be modeled in Flo and guarantees that map to our properties.

15th Temporal Web Analytics Workshop (TempWeb) Overview

CORD: Low-Latency, Bandwidth-Efficient and Scalable Release Consistency via Directory Ordering

Increasingly, multi-processing unit (PU) systems (e.g., CPU-GPU, multi-CPU, multi-GPU, etc.) are embracing cache-coherent shared memory to facilitate inter-PU communication. The coherence protocols in these systems support write-through accesses that place the data directly at the LLC to enable efficient producer-consumer communications pervasive in AI/ML workloads. Moreover, release consistency has emerged as the standard memory model in such systems due to its programming simplicity and ability to support high performance. In today’s multi-PU systems, the source processor that issues the writes also orders them to enforce release consistency, even for write-through accesses. Unfortunately, such source ordering of write-through operations results in unnecessary communications between the source processor and the LLC directory, incurring significant performance, interconnect traffic, and energy overheads for multi-PU applications. To eliminate such communication, we present cord 1, a novel cache coherence protocol that orders write-through accesses directly at the cache directory. cord employs several novel mechanisms to minimize the metadata required for ordering traffic while efficiently scaling to multiple directories. Evaluations atop the gem5 simulator show that compared to source ordering, cord improves application performance by \( 24\% \) and reduces traffic by \( 13\% \) on average while incurring \( \lt 1\% \) storage, area, and power overheads. Compared to hand-optimized message-passing implementations, cord observes a mere \( 3\% \) performance overhead and \( 6\% \) more traffic on average with a significantly simpler programming model.

Relax: Composable Abstractions for End-to-End Dynamic Machine Learning

Dynamic shape computations have become critical in modern machine learning workloads, especially in emerging large language models. The success of these models has driven the demand for their universal deployment across a diverse set of backend environments. In this paper, we present Relax, a compiler abstraction for optimizing end-to-end dynamic machine learning workloads. Relax introduces a cross-level abstraction that encapsulates computational graphs, loop-level tensor programs, and external library calls in a single representation. Relax also introduces first-class symbolic shape annotations to track dynamic shape computations globally across the program, enabling dynamic shape-aware cross-level optimizations. We build an end-to-end compilation framework using the proposed approach to optimize dynamic shape models. Experimental results on LLMs show that Relax delivers performance competitive with state-of-the-art systems across various GPUs and enables deployment of emerging models to a broader set of emerging environments, including mobile phones, embedded devices, and web browsers.

pulse: Accelerating Distributed Pointer-Traversals on Disaggregated Memory

Caches at CPU nodes in disaggregated memory architectures amortize the high data access latency over the network. However, such caches are fundamentally unable to improve performance for workloads requiring pointer traversals across linked data structures. We argue for accelerating these pointer traversals closer to disaggregated memory in a manner that preserves expressiveness for supporting various linked structures, ensures energy efficiency and performance, and supports distributed execution. We design pulse, a distributed pointer-traversal framework for rack-scale disaggregated memory to meet all the above requirements. Our evaluation of pulse shows that it enables low-latency, high-throughput, and energy-efficient execution for a wide range of pointer traversal workloads on disaggregated memory that fare poorly with caching alone.

Quantitative Evaluation of Using Large Language Models and Retrieval-Augmented Generation in Computer Science Education.

Domain specific ontologies from Linked Open Data (LOD)

Logical and probabilistic reasoning tasks that require a deeper knowledge of semantics are increasingly relying on general purpose ontologies such as Wikidata and DBpedia. However, tasks such as entity disambiguation and linking may benefit from domain specific knowledge graphs, which make it more efficient to consume the knowledge and easier to extend with proprietary content. We discuss our experience bootstrapping one such ontology for IT with a domain-agnostic pipeline, and extending it using domain-specific glossaries.

Speaker Role Identification in Clinical Conversations

247 UCCT-BCMA-1: a first-in-human, non-viral CRISPR-engineered CAR T cell therapy targeting BCMA in relapsed/refractory multiple myeloma

Social Context Matters for Turn‐Taking Dynamics: A Comparative Study of Autistic and Typically Developing Children

FairXAI - A Taxonomy and Framework for Fairness and Explainability Synergy in Machine Learning

Predictive and Reinforcement Learning-Based Framework for Cloud Resource Optimization

The article presents a new optimization framework for dynamically managing resources in a cloud environment that combines predictive analytics and RL to optimally manage operational cost, system performance, and compliance with SLAs in multi-tenant cloud settings. Our framework comprises a predictive tier featuring resource demand forecasting based on regression models and an optimization tier that employs real-time resource allocation using Deep Q-Networks (DQN). Experiments conducted with real-world data from Microsoft Azure and MIT’s Supercloud have shown that the framework reduces over-provisioning and SLA violations while improving cost-efficiency. The automated dynamic resource allocation strategy proposed in this study outperforms traditional static allocation methods by reducing resource wastage and enhancing SLA compliance, demonstrating the viability of the approach in sophisticated multi-tenant cloud environments. This blended approach improves the efficiency of resource utilization while maintaining flexible and economical resource control. The findings accelerate the integration of AI-powered models, such as predictive analytics and reinforcement learning, with cloud resource management in response to the evolving challenges of cloud infrastructure complexity. We show that the proposed framework addresses the problems of achieving high-performance, scalable, and cost-effective optimization of cloud resources

A Formalization of Measurement-Commuting Unitaries

Hybrid classical-quantum algorithms promise efficient use of near-term quantum devices by strategically partitioning computation between classical and quantum resources. A central challenge in this paradigm is identifying quantum operations that can be executed independently of specific measurement outcomes, thereby reducing quantum-classical feedback and enabling circuit cutting, early measurement, and simulation optimizations. This paper introduces a formal framework for analyzing when unitary operations commute with projective measurements. We characterize a class of such unitaries, termed eigenvalue permutations, that preserve measurement outcomes and thus commute with measurement in a well-defined sense. These unitaries allow for early measurement without additional computational overhead, offering a new primitive for optimizing hybrid circuits. We discuss theoretical foundations, provide structural results, and outline implications for hybrid execution strategies, circuit cutting, and classical simulation.

SwarmSearch: Decentralized Search Engine with Self-Funding Economy

Enhancing Personal Healthcare through AI-Driven DNA Analysis and Public Health Data Integration

This article presents a privacy-preserving framework that integrates AI-driven DNA analysis with publicly available health data to deliver personalized healthcare recommendations. The model synthesizes private genomic profiles with research repositories such as PubMed and the NIH Genome Database, ensuring that insights are both individualized and grounded in current medical evidence. A dual-enclave architecture separates private DNA information from public research inputs, while large language models (LLMs) serve as verification agents to cross-check outcomes against peer-reviewed literature. Key genetic markers, including MTHFR, COMT, MTR, and AHCY, are analyzed to provide recommendations on supplementation, stress management, and lifestyle modifications. By employing a zeroknowledge design and encrypted processing, sensitive data remains protected while benefiting from continuous recalibration against emerging evidence. Unlike existing frameworks that remain descriptive or limited to static recommendations, the proposed system enables dynamic, realtime updates that adapt to evolving genomic and clinical knowledge. This approach advances personalized medicine by offering actionable, evidence-based interventions while maintaining the confidentiality and integrity of individual health records.

Human-in-the-Loop Evaluation Pipelines for Production LLM Systems

Large Language Models (LLMs) have become essential tools in a wide variety of domains; however, inherent limitations such as hallucinations, ambiguities, domain-specific errors, and others often create significant problems when used in the context of productions. We present a Human-in-the-Loop (HITL) assessment pipeline in this work that is an integration of human feedback, LLM-based adjudication system, and observability tools and expect to increase the reliability, safety and alignment of the results produced by the LLM with the business needs. It implements the use of human raters to create evaluation rubrics, review model outputs and provide feedback, and an adjudicator of the LLM automates the scoring process to ensure scalability. Empirical case studies on the customer support and document quality assurance have shown that the hybrid approach brings improvements to the accuracy of task completion, customer satisfaction, and policy compliance. Moreover, the manuscript outlines the best practices regarding the scaling of the HITL workflows, handling high-risk cases, and continuously refining the system performance through the continuous feedback processes. In this way, the suggested solution will provide a scalabilitybased, low-cost paradigm to enrich the usage of LLM applications and, at the same time, maintain the alignment with the business requirements and safety level.

Federated Edge Intelligence for Low-Latency Intrusion Detection in Smart Homes

The increasing number of smart home devices has created critical security concerns that call for sophisticated intrusion detection systems. This research develops a distributed, AI-based smart home Intrusion Detection System (IDS) which utilizes machine learning, federated learning, and edge intelligence for anomaly detection in IoT devices, all while safeguarding user privacy. At the core of the system is federated learning which permits group-based training without exposing raw data, safeguarding privacy, mitigating datasharing issues, and lowering network load. Applying public IoT intrusion datasets for evaluation showed that the IDS have a strong detection rate while incurring low processing costs. Because of its distributed nature, the system is able to span a large number of IoT devices and provide real-time responses to emerging threats. This study resolves major concerns such as device diversity, data confidentiality, delays, and system expandability, providing smart home systems with reliable and practical security that strengthens threat detection and maintains optimal system performance.

Leveraging Multiple LLM Evaluators for Scalable and Fair Language Model Assessments

Augmenting Hate Speech Detection with Hybrid CNN-RNN Models and Dataset Distribution Techniques

Identifying hate speech is an important element in controlling toxicity and allowing people to function in digital spaces. In this work, we scale hybrid CNN-RNN architectures using augmentation and carefully balancing datasets to achieve better classification accuracy of hate speech. Unlike previous works that grappled with the one-sided persistent fight against imbalanced datasets, our work builds balanced datasets through synthetic oversampling like SMOTE, under sampling, and modern augmentation strategies that create stronger distributions for training. Moreover, we integrate other state-of-the-art machine and deep learning systems, including pre-trained language models and explainable AI. Explainable AI techniques helped enable transparency by providing insights into the flagged content which builds trust. This work achieves better accuracy (up to 0.908) and F1 scores (up to 0.914) while maintaining computational efficiency for real-time deployment. Bias mitigation as well as ethical frameworks for adaptation within torture communities strengthens ethical concerns. The societal interest motivates development into large-scale, multi-ligneous, big-impact hate-speech detection frameworks.

Framework for Fairness in Machine Learning Using Detecting and Mitigating Bias in AI Algorithms

The importance of artificial intelligence (AI) and machine learning (ML) is on the rise as they are increasingly being used in critical decision-making in areas including healthcare, finance, and criminal justice. Their effectiveness, however, is often compromised as these systems replicate and even worsen pre-existing biases from historical data or from the designs of models, ultimately causing unfair and undesirable results. This paper describes a systematic analysis for bias detection and mitigation in machine learning algorithms. The approach includes preprocessing of data, algorithmic changes, and changes during post-processing for improvement of anticipated inequalities while keeping the performance of the model intact. The model's applicability has been illustrated in the areas of healthcare, finance, and criminal justice with case studies demonstrating the need for a trade-off between the model's accuracy and fairness. Besides, AI ethics are also advanced through transparency methods like SHAP and GradCAM, which strengthen trust in AI systems as ethical principles are ensured in the deployment of the AI systems. Though these achievements are significant, challenges on the issue of scaling, fairness definitions, and contextual variations still persist with regard to the application of AI for fairness promoting purposes, suggesting interdisciplinary directions as ways to mitigate the challenges.

Cross-Examiner: Evaluating Consistency of Large Language Model-Generated Explanations

Large Language Models (LLMs) are often asked to explain their outputs to enhance accuracy and transparency. However, evidence suggests that these explanations can misrepresent the models' true reasoning processes. One effective way to identify inaccuracies or omissions in these explanations is through consistency checking, which typically involves asking follow-up questions. This paper introduces, cross-examiner, a new method for generating follow-up questions based on a model's explanation of an initial question. Our method combines symbolic information extraction with language model-driven question generation, resulting in better follow-up questions than those produced by LLMs alone. Additionally, this approach is more flexible than other methods and can generate a wider variety of follow-up questions.

Towards a Single ASR Model That Generalizes to Disordered Speech.

Speech Recognition with LLMs Adapted to Disordered Speech Using Reinforcement Learning.

Optimizing Federated Learning in the Era of LLMs: Message Quantization and Streaming

Federated Learning (FL) offers a promising solution for training machine learning models across distributed data sources while preserving data privacy. However, FL faces critical challenges related to communication overhead and local resource constraints, especially in the era of Large Language Models (LLMs) with billions of parameters. The sheer size of these models exacerbates both memory and communication constraints, making efficient transmission and processing essential for practical deployment. NVIDIA FLARE, an open-source SDK for federated learning, addresses these challenges by introducing advanced communication capabilities. Building upon existing solutions for large object streaming, we enhance FL workflows for LLMs through two key techniques: message quantization and container/file streaming. Quantization reduces message size, while streaming enables efficient memory management, improving scalability and integration with existing workflows. These advancements significantly enhance the robustness and efficiency of FL with LLMs, ensuring better performance in real-world federated learning scenarios.

Learning to Drive from a World Model

Most self-driving systems rely on hand-coded perception outputs and engineered driving rules. Learning directly from human driving data with an end-to-end method can allow for a training architecture that is simpler and scales well with compute and data. In this work, we propose an end-to-end training architecture that uses real driving data to train a driving policy in an on-policy simulator. We show two different methods of simulation, one with reprojective simulation and one with a learned world model. We show that both methods can be used to train a policy that learns driving behavior without any hand-coded driving rules. We evaluate the performance of these policies in a closed-loop simulation and when deployed in a real-world advanced driver-assistance system.

Secure Federated XGBoost with CUDA-accelerated Homomorphic Encryption via NVIDIA FLARE

Federated learning (FL) enables collaborative model training across decentralized datasets. NVIDIA FLARE's Federated XGBoost extends the popular XGBoost algorithm to both vertical and horizontal federated settings, facilitating joint model development without direct data sharing. However, the initial implementation assumed mutual trust over the sharing of intermediate gradient statistics produced by the XGBoost algorithm, leaving potential vulnerabilities to honest-but-curious adversaries. This work introduces "Secure Federated XGBoost", an efficient solution to mitigate these risks. We implement secure federated algorithms for both vertical and horizontal scenarios, addressing diverse data security patterns. To secure the messages, we leverage homomorphic encryption (HE) to protect sensitive information during training. A novel plugin and processor interface seamlessly integrates HE into the Federated XGBoost pipeline, enabling secure aggregation over ciphertexts. We present both CPU-based and CUDA-accelerated HE plugins, demonstrating significant performance gains. Notably, our CUDA-accelerated HE implementation achieves up to 30x speedups in vertical Federated XGBoost compared to existing third-party solutions. By securing critical computation steps and encrypting sensitive assets, Secure Federated XGBoost provides robust data privacy guarantees, reinforcing the fundamental benefits of federated learning while maintaining high performance.

STARVESPAM: Mitigating Spam with Local Reputation in Permissionless Blockchains

Screening Dyslexia Using Visual Auditory Computer Games and Machine Learning

Automatic quantification of syntactic complexity in natural spontaneous speech of people with primary progressive aphasia

Patient-Specific In Vivo Gene Editing to Treat a Rare Genetic Disease

Author response for "Microfluidic capillary transit velocity as a functional measure for sickle cell disease and in vitro-derived red blood cells"

Author response for "Microfluidic capillary transit velocity as a functional measure for sickle cell disease and in vitro-derived red blood cells"

Author response for "Microfluidic capillary transit velocity as a functional measure for sickle cell disease and in vitro-derived red blood cells"

Automated speech and language markers of longitudinal changes in psychosis symptoms

Principled approaches for extending neural architectures to function spaces for operator learning

Deep learning has achieved remarkable success in computer vision and natural language processing, where tasks are commonly formulated as mappings between finite-dimensional representations. Many scientific problems, however, including those governed by partial differential equations, are naturally posed on infinite-dimensional function spaces. This mismatch has limited conventional neural networks from achieving comparable success in scientific applications. Here we identify and distil key principles for constructing practical neural architectures for mappings between function spaces. Neural operators provide a principled extension of neural networks to such settings, offering a path towards bringing deep learning’s transformative impact to science. Because deep learning’s success has relied heavily on architectural refinements, extending these advances to neural operators allows operator learning to benefit from refined designs. Guided by the principles we outline, we propose a recipe for converting popular neural architectures into neural operators with minimal modifications. We also discuss practical steps for making these models effective. This perspective offers a systematic bridge between finite-dimensional network design and operator learning for scientific applications. Berner et al. show how to adapt popular neural networks into discretization-agnostic neural operators that learn from continuous scientific data, enabling scientific simulations that generalize more re

Measurement and clinical interpretation of CRISPR off-targets

Optimizing generative AI by backpropagating language model feedback

Unique signatures in verbal fluency task performance in schizophrenia and depression

ULKB Logic: A HOL-based framework for reasoning over knowledge graphs

Large Language Models and Data Quality for Knowledge Graphs

Bias, machine learning, and conceptual engineering

Large language models (LLMs) such as OpenAI’s ChatGPT reflect, and can potentially perpetuate, social biases in language use. Conceptual engineering aims to revise our concepts to eliminate such bias. We show how machine learning and conceptual engineering can be fruitfully brought together to offer new insights to both conceptual engineers and LLM designers. Specifically, we suggest that LLMs can be used to detect and expose bias in the prototypes associated with concepts, and that LLM de-biasing can serve conceptual engineering projects that aim to revise such conceptual prototypes. At present, these de-biasing techniques primarily involve approaches requiring bespoke interventions based on choices of the algorithm’s designers. Thus, conceptual engineering through de-biasing will include making choices about what kind of normative training an LLM should receive, especially with respect to different notions of bias. This offers a new perspective on what conceptual engineering involves and how it can be implemented. And our conceptual engineering approach also offers insight, to those engaged in LLM de-biasing, into the normative distinctions that are needed for that work.

Advancing AI Incidents Classification: Leveraging LLMs with Strategic Prompting