Immersive formats such as 360° and 6DoF point cloud videos require high bandwidth and low latency, posing challenges for real-time AR/VR streaming. This work focuses on reducing bandwidth consumption and encryption/decryption delay, two key contributors to overall latency. We design a system that downsamples point cloud content at the origin server and applies partial encryption. At the client, the content is decrypted and upscaled using an ML-based super-resolution model. We integrate this pipeline into an adaptive bitrate (ABR) streaming client, which enables the system to dynamically select among multiple resolution levels and apply Super-Resolution (SR) to enhance delivered quality. We evaluate our approach through a full streaming deployment on CloudLab where we compare SR-aided ABR streaming against a standard ABR baseline under varying cache configurations. Our results show that the SR-aided approach raises the average delivered resolution from approximately 18% to 31% of full density under constrained network conditions, a 76% relative improvement over the ABR-only baseline. Additionally, we demonstrate a nearly linear reduction in bandwidth consumption and encryption/decryption overhead with lower downsampling resolutions while maintaining the ability to effectively reconstruct the original full-resolution point clouds with inference times as low as 68 ms for a 157K-point frame.
When analyzing networks, Betweenness Centrality is a method of measurement that can be used to find the frequency at which a node is accessed when calculating the shortest path between various nodes. The more often a node is accessed, the greater the node’s Betweenness Centrality score becomes. This metric allows for numerous practical applications, from gene identification for diseases to analyzing the movement of data through informational infrastructures. With modern day computing power, we are able to utilize Large Language Models (LLMs) to calculate Betweenness Centrality on large-scale networks. However, the performance of LLMs may not be quick nor accurate when presenting a response for Betweenness Centrality graph problems. Therefore, we propose a GPS-BC module that has the capability to formulate accurate solutions for Betweenness Centrality problems using tailored algorithms. The problems contain several combinations of the constraints on (A) node color, (B) edge/path weights, and (C) node degree. For all the constrained problems, GPT-5.2 was observed to have an average accuracy of 62% when reporting Betweenness Centrality scores. The accuracy decreases when increasing the complexity of the constraints imposed on a problem. Additionally, GPS-BC algorithms were found to be, on average, [Formula: see text] faster than the observed processing time for GPT-5.2. Through this approach of using validated algorithms in our GPS-BC module, we are able to process Betweenness Centrality calculations with efficiency and ensured accuracy.
Predicting student grades plays a critical role in providing individualized academic guidance and identifying students at risk of failing to complete their studies. While previous studies on student performance prediction have often relied on student demographics or attempted to model students’ entire academic careers, we focus on predicting future grades using both individual academic records and transcripts of similar students. Our proposed framework leverages peer information to model student grade trajectories and aligns them with the academic records of previously enrolled students who followed similar academic paths. This approach enhances prediction accuracy while maintaining adaptability across diverse curricula. Furthermore, to address data scarcity and privacy concerns, we present a methodology for generating synthetic datasets using generative language models, enabling the creation of realistic academic records without compromising student privacy. Our evaluation of sequential academic progression modeling includes four temporal models: RNN, LSTM, GRU, and Transformer. LSTM achieved the best performance, closely followed by GRU and RNN, while the Transformer performed slightly worse.
Finding shortest paths in graphs with complex constraints is a fundamental problem in optimization with applications ranging from network routing to robotics. Traditional approaches either rely on exhaustive enumeration, which has exponential complexity, or generic constraint solvers that fail to exploit the problem structure. We present a constraint-aware optimization framework Generative Problem Solver for the Shortest Path Problem (GPS-SP) that categorizes constraints by their computational properties and applies specialized algorithms to each category. Our framework enables automatic problem generation and constraint composition, thereby forming a generative problem solver for shortest path problems. It supports five constraint categories: (A) hard filters for static graph filtering, (B) regular constraints expressible as finite automata, (C) budget constraints on accumulated resources, (D) global graph properties, and (E) combinatorial constraints. We prove the completeness and optimality of our approach for all category combinations and demonstrate a [Formula: see text] average speedup ([Formula: see text] median) over exhaustive search on 1500 benchmark problems and a [Formula: see text] average speedup ([Formula: see text] median) over GPT-5.2 on 1280 benchmark problems with 54.2% agreement rate. This work provides both theoretical guarantees and practical efficiency for constraint shortest path problems and demonstrates a blueprint for extending the framework to other combinatorial optimization domains.
Combinatorial optimization and particularly Vehicle Routing Problems (VRPs), as well as trip or itinerary planning, remain a fundamental challenge in operations research. However, the advent of Large Language Models (LLMs) introduces a paradigm shift toward more adaptable, context-aware, and human-centric optimization frameworks. The study adopted a multidisciplinary literature review-driven approach to examine the intersection of LLMs and combinatorial optimization within the ambit of VRPs. The scope spans from professional practitioners using specialized optimization software to commuters and tourists navigating the built environment. The paper explores how LLMs may simultaneously strengthen underlying optimization processes and improve user interfaces — in particular by enhancing transparency, interactivity, and explainability. The study identified various opportunities for LLMs to enhance routing systems by, for example, enabling nuanced translation of (implicit) user input into personalized and context-aware route generation. The integration of Retrieval Augmented Generation (RAG) may enhance data fidelity and temporal relevance, while hybrid solver configuration and automated heuristic generation, as well as self-debugging mechanisms, are amongst the proposed application areas of LLMs. Advances in semantic routing may be especially pertinent for active travelers who rely on nuanced environmental attributes such as lighting, terrain, and perceived safety when selecting routes. Effectively capturing and incorporating such complex, multi-objective preferences into a route recommender system may support a broader adoption of sustainable, non-motorized modes of transport.
This study proposes the development of an academic domain-specialized Small Language Model (SLM), designed to be used in combination with a general-purpose Large Language Model (LLM), with the objective of constructing a knowledge graph through the highly accurate extraction of detailed technical elements, such as research objectives, methodologies and experimental targets from scholarly articles. The proposed academic domain-specialized SLM is developed through continuous pretraining using collections of survey papers within the target domain and is further optimized to specialize in extracting domain-specific technical elements through attention steering techniques. In the knowledge graph construction process, the SLM is employed as a fact-checking mechanism for outputs generated by a general-purpose LLM. In this paper, we conducted experiments on extracting technical elements in two fields: Fake News Detection in Computer Science and Bibliometrics in Science and Technology Management and confirmed that the proposed academic domain-specialized SLM outperforms general-purpose LLMs in terms of the accuracy, stability and importance of the extracted information. Future work includes constructing a knowledge graph by integrating the proposed SLM with general-purpose LLMs.
When a patient is diagnosed with rectal cancer, staging is performed to assess whether the cancer has metastasized to other parts of the body using diagnostic imaging such as CT scans. Recently, there has been a growing interest in applying machine learning for diagnostic support. In our previous studies, we proposed methods to augment limited datasets using generative models; however, concerns were raised that these models might generate unintended images. In this study, we aim to improve the accuracy of lymph node classification. We introduce Denoising Diffusion Probabilistic Models (DDPM) in addition to the Denoising Diffusion Implicit Models (DDIM) employed in prior work. Furthermore, we propose two filtering methods to eliminate unintended images. Additionally, we conducted a comprehensive parameter analysis of the diffusion schedules for both DDIM and DDPM to optimize the generation process.
This study investigates the usability of consumer-generated online comments to elucidate BEV drivers' concerns and perceptions regarding these vital facilities. A multi-label text classification was implemented and the walking accessibility of nearby amenities where BEV drivers might dwell while waiting was assessed for more than 500 fast-charging stations in California, USA. BERT achieved the lowest Hamming loss, closely followed by RoBERTa. The most prominent themes in the reviews were the availability of charging points (including waiting times before a charger became available), followed by the charging speed. In terms of amenities, BEV drivers frequently referenced food and dining options as well as retail outlets, toilet facilities (or the lack thereof) and cleanliness-related concerns such as bins, odors and trash. Although parking fees and difficulties in locating charging stations were mentioned in fewer than 10% of the reviews, several charging stations exhibited these issues as the most salient themes, being referenced in up to 80% of a charging station's reviews analyzed in the study. The majority of fast-charging stations were found to have nearby amenities belonging to the category 'Drinks and snacks', 'Restaurant', 'Fast food' and/or 'Groceries'. This paper is an extended version of the paper "Fast Charging Stations in the USA: A Multi-Label Text Classification of Online Reviews".
Knowledge distillation (KD) plays a crucial role in reducing computational costs, accelerating inference, and improving model generalization. However, when there is a significant capacity gap between teacher and student networks, KD often struggles to transfer knowledge effectively, and robustness under distribution shifts remains a major challenge. To address these issues, we propose learning-to-learn knowledge distillation (L2L-KD), a dynamic temperature-controlled KD framework that progressively increases learning difficulty, mimicking how human learners advance from basic to complex concepts. To further enhance robustness and generalization, we introduce a counterfactual data augmentation technique that leverages the Metropolis-Hastings algorithm to generate fluent and semantically coherent out-of-domain (OOD) samples. We evaluate L2L-KD across in-domain, OOD, and adversarial scenarios, and the results show that it consistently outperforms existing KD approaches while substantially improving robustness. Moreover, building upon this foundation, we extend the core learning philosophy to a new unsupervised cross-domain framework, demonstrating that the dynamic distillation principles of L2L-KD can naturally generalize to broader domain adaptation tasks.
Extracting information from scientific literature presents persistent challenges due to the scarcity and complexity of scientific entities and relations, the esoteric nature of domain-specific language, and the high cost of generating labeled training data. This study re-introduces ELSIE-Blob, a relation extraction preprocessing method designed to preserve contextual integrity by identifying concise, information-rich text segments ("Blobs") that contain key components such as polymer names, property indicators (e.g. T-m), and numerical values (e.g. 125(degrees)C). Originally developed for extracting glass-transition temperatures (T-g), we extend and evaluate ELSIE-Blob for extracting polymer melting points (T-m) from materials science literature. After preprocessing, the natural language processing task involves extracting the polymer-property pair contained within the text, using state-of-the-art Large Language Models. Beyond reducing the volume of text for input, ELSIE-Blob addresses a core limitation of Large Language Models: The difficulty of reasoning over long-range dependencies in unstructured text, often caused by attention dilution. We show that accessible Large Language Models fail to extract polymer-property relations without ELSIE-Blob pre-processing. By segmenting input into domain-relevant, information-dense units via weak supervision, our method enables accurate relation extraction without any labeled data. Experiments on 50 full-text articles using multiple Large Language Models and prompt configurations demonstrate the value of domain-aware preprocessing in improving scientific information extraction.
The rapid rise of large language models (LLMs) has driven their widespread adoption, especially as the core component of open-source applications, which we refer to as LLM-Powered Apps (LPAs). Despite the rapid growth of this ecosystem, little is known about how these applications are built in the open-source world, especially in terms of their architectural and design decisions, deployment strategies, and security practices, which remain poorly understood.In this paper, we conduct a comprehensive empirical study of 89 popular open-source LPAs on GitHub, with the goal of characterizing their design choices and identifying common security and safety concerns. We systematically collect a set of architectural and operational attributes, classify each LPA by its primary purpose and analyze how functionality influences architectural and security design. Our findings reveal dominant design patterns as well as recurring risks, such as inadequate access control, lack of telemetry transparency, and design assumptions that break down in complex runtime environment. We also conducted an in-depth study of 376 GitHub issues from two popular LPAs and two LIFs developed in open-source communities. We summarize the root causes of the GitHub issues by creating a taxonomy of three themes based on the software development life-cycle. By surfacing these trends and vulnerabilities, our study provides a foundational understanding of how LPAs are currently built and deployed in the open-source ecosystem as well as the recurring issues and pitfalls for developing and maintaining LPAs. The results offer practical insights for developers, researchers, and platform maintainers seeking to build more robust and secure LLM-integrated software systems. Furthermore, we propose a set of guideline for the secure use of API keys, encapsulated within a pre-commit hook and a GitHub Action workflow for easy integration into new and existing GitHub projects using API keys.
Safety-critical sectors such as power transmission and autonomous mobility are not replacing their deterministic controllers; they are adding artificial intelligence (AI) layers to existing systems to gain additional, sustainable business value. This approach increases the socio-technical complexity, creating pressure for a structured means to compare benefits, oversight effort and risk over time. In this descriptive study, we introduce Epistemic Gain G, a scalar derived from the Delta - eta - zeta model that links foresight gains to human oversight and system friction. Within this framework, G > 0 is treated as a necessary condition for epistemically sustainable scaling. We then formulate a conjectured governance-level Law of Diminishing Returns that holds up to a Scaling Failure Threshold, beyond which marginal upgrades begin to destroy value. Drawing on recent empirical studies, we further sketch the Delta - eta - zeta model and show how G can be displayed in software development lifecycle dashboards. This paper extends the earlier IEEE AI & times;B 2025 conference paper in three main directions: (i) provide the theoretical foundation of the M-Vector and formalize it as the explicit epistemic state M(t) underpinning the Delta - eta - zeta model; (ii) introducing semantic instability phi and epistemic drift xi as properties inspired by causal representation learning, used here as an AI-safety and governance lens; and (iii) identifying canonical regions of M-space for deterministic, vital, symbolic, sub-symbolic and generative AI systems. The overarching aim is to formalize the theoretical basis of G and its time-variant machine form, enabling it to serve as a governance indicator for when to scale, optimize, or pause AI deployments.
With the release of ChatGPT, Pandora's box of generative AI has been opened, and the new technology is here to stay. This also impacts academic education, where it bears certain challenges while also providing new opportunities. Students and educators alike will have to adapt to the newly available technology and find ways to use it in a meaningful and profitable way. One way to make good use of generative AI in the classroom is to build personalized learning environments for students that adapt to their individual progress and enhance the learning process with technologies that are tailored to the modern students. In the presented work, two of these technologies and the concept of spaced repetition are used to build a personalized learning module within the learning management system (LMS) "ILIAS". One of the tools used is an AI chatbot based on the ChatGPT API that is trained on the lecturer's actual course materials and enhanced with answers from ChatGPT itself. The other technology improves educational videos by providing learners with a button to pause the video at any time and receive additional explanations about the content currently discussed, while considering the context of the entire video so far. These two AI tools are then provided in combination with a spaced repetition algorithm, which creates a personal learning environment that is highly tailored to the individual learners' needs.
Agentic large language models (LLMs) have emerged as powerful tools for autonomously interacting with external environments and performing multi-step reasoning. While most existing approaches rely on in-context learning with multi-turn few-shot prompts, these methods often require long inputs and, consequently, incur high computational costs and latency. Agent fine-tuning offers a resource-aware alternative by enabling models to internalize procedural reasoning patterns and domain-specific knowledge through demonstrations and curated training data. However, its effectiveness in highly specialized technical microdomains remains underexplored. This work investigates agent fine-tuning with knowledge distillation for adapting LLMs to Hitachi's JP1 middleware, a complex microdomain centered on IT operations management. We fine-tune models using JP1-specific corpora extracted from manuals and textbooks, together with distilled reasoning trajectories (ReAct and CoT) generated by larger LLMs (GPT-4). At inference time, we incorporate retrieval-augmented generation (RAG) with an agentic prompt and introduce a context-answer extractor (CAE) to improve grounding and relevance. On JP1 certification examinations, our model that was continually pre-trained on JP1 manuals and further fine-tuned with ReAct trajectories achieves substantial improvements over the base model - 13% (Engineer), 12% (Professional), and 10% (Consultant) - while delivering up to 9.78 times higher cost efficiency than strong general-purpose LLMs such as GPT-4. These results demonstrate that agent fine-tuning combined with knowledge distillation is a highly effective and economically scalable strategy for building high-fidelity LLM agents tailored to specialized technical domains.
As Machine Learning (ML) becomes ever more ubiquitous, it is critical to increase the rigor of design and testing. In traditional software, this is done using the Requirements Engineering (RE) process, but RE looks different for ML because it is data-centric. There are different Non-Functional Requirements (NFRs) and standard testing techniques do not apply. While there are standards for verifying NFRs in traditional software, there is no standard measurement for ML NFRs, e.g. how is a model verified to meet an explainability NFR? In traditional software, NFRs are decomposed into Functional Requirements (FRs), but without clear measurements for ML NFRs, their decomposition into FRs is nearly impossible. However, recently, research has shown that provenance can help improve model transparency and reproducibility. This work builds on such literature and suggests provenance as a lower-level NFR to connect high-level NFRs, e.g. explainability and transparency, and FRs, thereby enabling concrete model verification based on requirement specifications. This work examines types of ML provenance and their use in decomposing model NFRs into verifiable FRs, thereby better aligning ML development with RE and increasing the rigor of ML testing. This paper aggregates current literature on provenance for ML and provides a method of measurement for otherwise unquantifiable NFRs. This work also analyzes trends in how provenance can decompose various ML NFRs and future directions for the field.
We introduce a Reinforcement Learning (RL)-based framework to optimize discrete natural language prompts for enhancing both the accuracy and clarity in sentence simplification. Using a lightweight PPO policy, our method learns to guide a frozen small-scale LLaMA-3.2 3B model toward effective simplification for supporting user-centric computational thinking tasks. Results show that our RL-optimized prompts significantly surpass manual baselines in semantic fidelity, logical coherence, and instructional quality. Moreover, the proposed RL-optimized prompting approach enables a much smaller LLM to achieve results that are comparable in clarity and instructional value to those produced by a much larger LLaMA-3.3 70B model.
Agentic large language models (LLMs) have become prominent for autonomously interacting with external environments and performing multi-step reasoning tasks. Most approaches leverage these capabilities via in-context learning with few-shot prompts, but this often results in lengthy inputs and higher computational costs. Agent fine-tuning offers an alternative by enabling LLMs to internalize procedural reasoning and domain-specific knowledge through training on relevant data and demonstration trajectories. While prior studies have focused on general domains, their effectiveness in specialized technical microdomains remains unclear. This paper explores agent fine-tuning for domain adaptation within Hitachi's JP1 middleware, a microdomain for specialized IT operations. We fine-tuned LLMs using JP1-specific datasets derived from domain manuals and distilled reasoning trajectories generated by LLMs themselves, enhancing decision making accuracy and search efficiency. During inference, we used an agentic prompt with retrieval-augmented generation and introduced a context-answer extractor to improve information relevance. On JP1 certification exam questions, our method achieved a 14
Dynamic data patterns and adversarial environments present significant challenges for machine learning models in financial service applications such as fraud detection and financial crime prevention. These systems must continuously adapt to evolving behaviors, requiring frequent model retraining with up-to-date data. However, this introduces a trade-off between incorporating recent trends and preserving valuable historical knowledge, all while managing the scalability of training datasets. Moreover, in high-throughput, time-sensitive deployment scenarios, the latency introduced by retraining can critically impair a model’s ability to respond promptly to emerging threats. Ensuring model agility under strict latency and governance constraints thus becomes a key challenge. In this study, we propose a temporal knowledge distillation-based label augmentation framework that leverages insights from prior model generations to accelerate the training of new models. By transferring distilled knowledge from older models to augment the labeling of current training data, our approach significantly reduces retraining time while enhancing model performance. Experimental evaluations demonstrate that the proposed method improves retraining efficiency while consistently enhancing predictive accuracy, training time, model agility, and robustness under dynamic, real-world financial conditions.
