Week Ending 5.4.2025
RESEARCH WATCH: 5.4.2025
Fine-Tuning Without Forgetting: Adaptation of YOLOv8 Preserves COCO Performance
This groundbreaking study addresses a critical challenge in computer vision: how to adapt pre-trained object detectors to specialized domains without losing general capabilities. The researchers systematically evaluated different fine-tuning strategies for YOLOv8 models on a fruit detection dataset, discovering that deeper fine-tuning (unfreezing down to layer 10) dramatically improved specialized performance without compromising original COCO benchmark capabilities. This work provides compelling evidence that catastrophic forgetting—a common concern in transfer learning—can be avoided even with substantial adaptation. These findings have significant implications for deploying object detection in specialized fields like agriculture, medical imaging, and industrial inspection where both general and domain-specific performance matter.
Authors: Vishal Gandhi, Sagar Gandhi
Link: https://arxiv.org/abs/2505.01016v1
Date: 2025-05-02
Summary:
The success of large pre-trained object detectors hinges on their adaptability to diverse downstream tasks. While fine-tuning is the standard adaptation method, specializing these models for challenging fine-grained domains necessitates careful consideration of feature granularity. The critical question remains: how deeply should the pre-trained backbone be fine-tuned to optimize for the specialized task without incurring catastrophic forgetting of the original general capabilities? Addressing this, we present a systematic empirical study evaluating the impact of fine-tuning depth. We adapt a standard YOLOv8n model to a custom, fine-grained fruit detection dataset by progressively unfreezing backbone layers (freeze points at layers 22, 15, and 10) and training. Performance was rigorously evaluated on both the target fruit dataset and, using a dual-head evaluation architecture, on the original COCO validation set. Our results demonstrate unequivocally that deeper fine-tuning (unfreezing down to layer 10) yields substantial performance gains (e.g., +10\% absolute mAP50) on the fine-grained fruit task compared to only training the head. Strikingly, this significant adaptation and specialization resulted in negligible performance degradation (<0.1\% absolute mAP difference) on the COCO benchmark across all tested freeze levels. We conclude that adapting mid-to-late backbone features is highly effective for fine-grained specialization. Critically, our results demonstrate this adaptation can be achieved without the commonly expected penalty of catastrophic forgetting, presenting a compelling case for exploring deeper fine-tuning strategies, particularly when targeting complex domains or when maximizing specialized performance is paramount.
--------------------------------------------------------------------------------------------------------
WebThinker: Empowering Large Reasoning Models with Deep Research Capability
WebThinker addresses a fundamental limitation of large reasoning models: their inability to access current web information for comprehensive research tasks. While models like OpenAI-o1 and DeepSeek-R1 excel at reasoning, they rely solely on static internal knowledge. This novel framework integrates a Deep Web Explorer that enables models to autonomously search, navigate websites, and extract relevant information when encountering knowledge gaps. Using an innovative Think-Search-and-Draft strategy and reinforcement learning optimization, WebThinker significantly outperforms existing methods across complex reasoning benchmarks. This technology promises to transform how AI conducts deep research, enhancing capabilities in scientific writing, comprehensive analysis, and knowledge-intensive tasks that require synthesizing diverse, up-to-date information sources.
Authors: Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yutao Zhu, Yongkang Wu, Ji-Rong Wen, Zhicheng Dou
Link: https://arxiv.org/abs/2504.21776v1
Date: 2025-04-30
Summary:
Large reasoning models (LRMs), such as OpenAI-o1 and DeepSeek-R1, demonstrate impressive long-horizon reasoning capabilities. However, their reliance on static internal knowledge limits their performance on complex, knowledge-intensive tasks and hinders their ability to produce comprehensive research reports requiring synthesis of diverse web information. To address this, we propose \textbf{WebThinker}, a deep research agent that empowers LRMs to autonomously search the web, navigate web pages, and draft research reports during the reasoning process. WebThinker integrates a \textbf{Deep Web Explorer} module, enabling LRMs to dynamically search, navigate, and extract information from the web when encountering knowledge gaps. It also employs an \textbf{Autonomous Think-Search-and-Draft strategy}, allowing the model to seamlessly interleave reasoning, information gathering, and report writing in real time. To further enhance research tool utilization, we introduce an \textbf{RL-based training strategy} via iterative online Direct Preference Optimization (DPO). Extensive experiments on complex reasoning benchmarks (GPQA, GAIA, WebWalkerQA, HLE) and scientific report generation tasks (Glaive) demonstrate that WebThinker significantly outperforms existing methods and strong proprietary systems. Our approach enhances LRM reliability and applicability in complex scenarios, paving the way for more capable and versatile deep research systems. The code is available at https://github.com/RUC-NLPIR/WebThinker.
--------------------------------------------------------------------------------------------------------
Using LLMs in Generating Design Rationale for Software Architecture Decisions
This research explores how Large Language Models can address a persistent challenge in software development: documenting the reasoning behind architectural decisions. Design Rationale (DR) documentation is essential for software maintenance but often neglected due to developer time constraints. The study evaluated five LLMs generating DR for 100 architecture-related problems from Stack Overflow and GitHub using three prompting strategies. Results show promising precision (0.267-0.278), recall (0.627-0.715), and F1-scores (0.351-0.389), with approximately 65-69% of LLM-generated arguments not mentioned by humans still being helpful. This technology could revolutionize software documentation practices, preserving institutional knowledge and facilitating better architecture decisions through automated, comprehensive reasoning capture.
Authors: Xiyu Zhou, Ruiyin Li, Peng Liang, Beiqi Zhang, Mojtaba Shahin, Zengyang Li, Chen Yang
Link: https://arxiv.org/abs/2504.20781v1
Date: 2025-04-29
Summary:
Design Rationale (DR) for software architecture decisions refers to the reasoning underlying architectural choices, which provides valuable insights into the different phases of the architecting process throughout software development. However, in practice, DR is often inadequately documented due to a lack of motivation and effort from developers. With the recent advancements in Large Language Models (LLMs), their capabilities in text comprehension, reasoning, and generation may enable the generation and recovery of DR for architecture decisions. In this study, we evaluated the performance of LLMs in generating DR for architecture decisions. First, we collected 50 Stack Overflow (SO) posts, 25 GitHub issues, and 25 GitHub discussions related to architecture decisions to construct a dataset of 100 architecture-related problems. Then, we selected five LLMs to generate DR for the architecture decisions with three prompting strategies, including zero-shot, chain of thought (CoT), and LLM-based agents. With the DR provided by human experts as ground truth, the Precision of LLM-generated DR with the three prompting strategies ranges from 0.267 to 0.278, Recall from 0.627 to 0.715, and F1-score from 0.351 to 0.389. Additionally, 64.45% to 69.42% of the arguments of DR not mentioned by human experts are also helpful, 4.12% to 4.87% of the arguments have uncertain correctness, and 1.59% to 3.24% of the arguments are potentially misleading. Based on the results, we further discussed the pros and cons of the three prompting strategies and the strengths and limitations of the DR generated by LLMs.
--------------------------------------------------------------------------------------------------------
Voice Cloning: Comprehensive Survey
This survey provides a standardized framework for understanding the rapidly evolving field of voice cloning technology. As digital voice replication capabilities advance, the paper establishes consistent terminology and explores various approaches from foundational speaker adaptation techniques to cutting-edge few-shot, zero-shot, and multilingual text-to-speech methods. The authors comprehensively review evaluation metrics and available datasets, creating a valuable resource for researchers and developers. By consolidating current voice cloning algorithms, the survey promotes both research advancement and important work on detection technologies to prevent misuse. This foundation is crucial as voice cloning applications expand across industries from entertainment and accessibility to personalized AI assistants and secure voice authentication systems.
Authors: Hussam Azzuni, Abdulmotaleb El Saddik
Link: https://arxiv.org/abs/2505.00579v1
Date: 2025-05-01
Summary:
Voice Cloning has rapidly advanced in today's digital world, with many researchers and corporations working to improve these algorithms for various applications. This article aims to establish a standardized terminology for voice cloning and explore its different variations. It will cover speaker adaptation as the fundamental concept and then delve deeper into topics such as few-shot, zero-shot, and multilingual TTS within that context. Finally, we will explore the evaluation metrics commonly used in voice cloning research and related datasets. This survey compiles the available voice cloning algorithms to encourage research toward its generation and detection to limit its misuse.
--------------------------------------------------------------------------------------------------------
A Systematic Literature Review of Parameter-Efficient Fine-Tuning for Large Code Models
This systematic review addresses a critical challenge in AI-powered software engineering: adapting large code models to specific tasks without prohibitive computational costs. The researchers analyzed 27 peer-reviewed papers on Parameter-Efficient Fine-Tuning (PEFT) techniques, which update only a small subset of parameters rather than entire models. The resulting comprehensive taxonomy categorizes PEFT applications by task type, distinguishing between generative scenarios like code summarization and non-generative tasks such as code clone detection. This work provides valuable guidance for researchers and practitioners seeking to implement sustainable AI-powered development workflows, particularly in resource-constrained environments. These techniques could democratize access to powerful code models while maintaining performance across diverse software engineering applications.
Authors: Md Zahidul Haque, Saima Afrin, Antonio Mastropaolo
Link: https://arxiv.org/abs/2504.21569v1
Date: 2025-04-29
Summary:
The rise of Artificial Intelligence (AI)-and particularly Large Language Models (LLMs) for code-has reshaped Software Engineering (SE) by enabling the automation of tasks such as code generation, bug detection, and repair. However, these models require significant computational resources for training and fine-tuning, posing challenges for real-world adoption in resource-constrained environments. To address this, the research community has increasingly turned to Parameter-Efficient Fine-Tuning (PEFT)-a class of techniques that enables the adaptation of large models by updating only a small subset of parameters, rather than the entire model. In this Systematic Literature Review (SLR), we examine the growing application of PEFT techniques-across a wide range of software engineering tasks. We analyze how these methods are used to optimize various deep learning (DL) architectures, focusing on their impact on both performance and efficiency. Our study synthesizes findings from 27 peer-reviewed papers, identifying patterns in configuration strategies and adaptation trade-offs. The outcome of this review is a comprehensive taxonomy that categorizes PEFT usage by task type, distinguishing between generative (e.g., Code Summarization) and non-generative (e.g., Code Clone Detection) scenarios. Our findings aim to inform future research and guide the practical deployment of PEFT in sustainable, AI-powered software development. Our artifacts are publicly available at https://github.com/alvi75/SLR-PEFT
--------------------------------------------------------------------------------------------------------
Automated Unit Test Case Generation: A Systematic Literature Review
This systematic review explores the critical field of automated unit test generation, addressing the high cost and resource demands of software testing. The researchers identified information gaps regarding improvements to evolutionary approaches like Genetic Algorithms and Particle Swarm Optimization, as well as current challenges facing automated testing. The paper consolidates knowledge about hybrid algorithm combinations and integration with mutation testing and neural networks, while examining test criteria used in these approaches. By addressing limitations in readability and mocking capabilities, this research provides direction for more efficient automated testing solutions. These advancements could significantly reduce testing costs while improving software quality and reliability across industries increasingly dependent on robust software systems.
Authors: Jason Wang, Basem Suleiman, Muhammad Johan Alibasa
Link: https://arxiv.org/abs/2504.20357v1
Date: 2025-04-29
Summary:
Software is omnipresent within all factors of society. It is thus important to ensure that software are well tested to mitigate bad user experiences as well as the potential for severe financial and human losses. Software testing is however expensive and absorbs valuable time and resources. As a result, the field of automated software testing has grown of interest to researchers in past decades. In our review of present and past research papers, we have identified an information gap in the areas of improvement for the Genetic Algorithm and Particle Swarm Optimisation. A gap in knowledge in the current challenges that face automated testing has also been identified. We therefore present this systematic literature review in an effort to consolidate existing knowledge in regards to the evolutionary approaches as well as their improvements and resulting limitations. These improvements include hybrid algorithm combinations as well as interoperability with mutation testing and neural networks. We will also explore the main test criterion that are used in these algorithms alongside the challenges currently faced in the field related to readability, mocking and more.
--------------------------------------------------------------------------------------------------------
This research addresses the critical challenge of predicting power restoration times during natural disasters—a capability growing increasingly important as climate variability intensifies. The researchers developed a Longitudinal Tabular Transformer (LTT) model that leverages historical outage data and sequential updates to provide more accurate restoration time predictions. Tested on 34,000 storm-related outages across three major utilities, the model improved Customer Satisfaction Impact metrics by over 19% compared to existing methods. The approach incorporates customer-informed metrics and interpretability techniques to enhance transparency. This technology could significantly improve utility disaster response, enabling better resource allocation during outages and helping vulnerable populations make critical decisions during extended power disruptions when accurate timing information is essential.
Authors: Bogireddy Sai Prasanna Teja, Valliappan Muthukaruppan, Carls Benjamin
Link: https://arxiv.org/abs/2505.00225v1
Date: 2025-05-01
Summary:
As climate variability increases, the ability of utility providers to deliver precise Estimated Times of Restoration (ETR) during natural disasters has become increasingly critical. Accurate and timely ETRs are essential for enabling customer preparedness during extended power outages, where informed decision-making can be crucial, particularly in severe weather conditions. Nonetheless, prevailing utility practices predominantly depend on manual assessments or traditional statistical methods, which often fail to achieve the level of precision required for reliable and actionable predictions. To address these limitations, we propose a Longitudinal Tabular Transformer (LTT) model that leverages historical outage event data along with sequential updates of these events to improve the accuracy of ETR predictions. The model's performance was evaluated over 34,000 storm-related outage events from three major utility companies, collectively serving over 3 million customers over a 2-year period. Results demonstrate that the LTT model improves the Customer Satisfaction Impact (CSI) metric by an average of 19.08% (p > 0.001) compared to existing methods. Additionally, we introduce customer-informed regression metrics that align model evaluation with real-world satisfaction, ensuring the outcomes resonate with customer expectations. Furthermore, we employ interpretability techniques to analyze the temporal significance of incorporating sequential updates in modeling outage events and to identify the contributions of predictive features to a given ETR. This comprehensive approach not only improves predictive accuracy but also enhances transparency, fostering greater trust in the model's capabilities.
--------------------------------------------------------------------------------------------------------
This innovative research introduces ChestX-Reasoner, a medical AI system that transforms radiology diagnosis by incorporating structured reasoning processes that mirror clinical workflows. Unlike most medical AI approaches, this multimodal large language model leverages process supervision extracted directly from clinical reports, creating step-by-step reasoning chains. The two-stage training framework combines supervised fine-tuning with reinforcement learning guided by process rewards. Evaluated on the comprehensive RadRBench-CXR benchmark, ChestX-Reasoner significantly outperforms existing medical and general-domain models in both diagnostic accuracy and reasoning ability. This approach represents a major advancement for explainable AI in healthcare, with potential applications in clinical decision support, radiology education, and improving consistency in medical imaging interpretation across healthcare settings.
Authors: Ziqing Fan, Cheng Liang, Chaoyi Wu, Ya Zhang, Yanfeng Wang, Weidi Xie
Link: https://arxiv.org/abs/2504.20930v1
Date: 2025-04-29
Summary:
Recent advances in reasoning-enhanced large language models (LLMs) and multimodal LLMs (MLLMs) have significantly improved performance in complex tasks, yet medical AI models often overlook the structured reasoning processes inherent in clinical practice. In this work, we present ChestX-Reasoner, a radiology diagnosis MLLM designed to leverage process supervision mined directly from clinical reports, reflecting the step-by-step reasoning followed by radiologists. We construct a large dataset by extracting and refining reasoning chains from routine radiology reports. Our two-stage training framework combines supervised fine-tuning and reinforcement learning guided by process rewards to better align model reasoning with clinical standards. We introduce RadRBench-CXR, a comprehensive benchmark featuring 59K visual question answering samples with 301K clinically validated reasoning steps, and propose RadRScore, a metric evaluating reasoning factuality, completeness, and effectiveness. ChestX-Reasoner outperforms existing medical and general-domain MLLMs in both diagnostic accuracy and reasoning ability, achieving 16%, 5.9%, and 18% improvements in reasoning ability compared to the best medical MLLM, the best general MLLM, and its base model, respectively, as well as 3.3%, 24%, and 27% improvements in outcome accuracy. All resources are open-sourced to facilitate further research in medical reasoning MLLMs.
--------------------------------------------------------------------------------------------------------
This research addresses a fundamental challenge in biomedical AI: the insufficient quantity and quality of annotated scientific corpora for training specialized large language models. The innovative m-KAILIN framework employs a multi-agent architecture guided by the Medical Subject Headings (MeSH) hierarchy to autonomously extract, synthesize, and evaluate biomedical textual data from scientific literature. By generating high-quality question-answer pairs with minimal manual involvement, the system creates training data that maintains comprehensive coverage while ensuring consistency with biomedical ontologies. Experimental results demonstrate that models trained on these distilled datasets outperform leading life sciences LLMs and proprietary models, enabling even Llama3-70B to surpass GPT-4 with MedPrompt. This approach could transform how biomedical AI models are trained, accelerating advances in medical research and clinical applications.
Authors: Meng Xiao, Xunxin Cai, Chengrui Wang, Yuanchun Zhou
Link: https://arxiv.org/abs/2504.19565v1
Date: 2025-04-28
Summary:
The rapid progress of large language models (LLMs) in biomedical research has underscored the limitations of existing open-source annotated scientific corpora, which are often insufficient in quantity and quality. Addressing the challenge posed by the complex hierarchy of biomedical knowledge, we propose a knowledge-driven, multi-agent framework for scientific corpus distillation tailored for LLM training in the biomedical domain. Central to our approach is a collaborative multi-agent architecture, where specialized agents, each guided by the Medical Subject Headings (MeSH) hierarchy, work in concert to autonomously extract, synthesize, and self-evaluate high-quality textual data from vast scientific literature. These agents collectively generate and refine domain-specific question-answer pairs, ensuring comprehensive coverage and consistency with biomedical ontologies while minimizing manual involvement. Extensive experimental results show that language models trained on our multi-agent distilled datasets achieve notable improvements in biomedical question-answering tasks, outperforming both strong life sciences LLM baselines and advanced proprietary models. Notably, our AI-Ready dataset enables Llama3-70B to surpass GPT-4 with MedPrompt and Med-PaLM-2, despite their larger scale. Detailed ablation studies and case analyses further validate the effectiveness and synergy of each agent within the framework, highlighting the potential of multi-agent collaboration in biomedical LLM training.
--------------------------------------------------------------------------------------------------------
MINT: Multi-Vector Search Index Tuning
This research addresses a critical challenge in vector databases: optimizing search performance for multi-vector scenarios where each item has multiple feature vectors. The researchers introduce MINT, a framework that intelligently selects indexes to minimize latency while meeting storage and recall constraints. In multi-vector databases—increasingly important for multi-modal and multi-feature applications—the choice of indexes significantly impacts performance. The proposed algorithms achieve impressive latency improvements, demonstrating 2.1X to 8.3X speedup compared to baseline approaches. This technology has far-reaching implications for applications like image-text search, recommendation systems, and scientific data analysis, where efficiently retrieving items based on multiple vector representations can dramatically improve user experience and enable new capabilities in AI-powered systems.
Authors: Jiongli Zhu, Yue Wang, Bailu Ding, Philip A. Bernstein, Vivek Narasayya, Surajit Chaudhuri
Link: https://arxiv.org/abs/2504.20018v1
Date: 2025-04-28
Summary:
Vector search plays a crucial role in many real-world applications. In addition to single-vector search, multi-vector search becomes important for multi-modal and multi-feature scenarios today. In a multi-vector database, each row is an item, each column represents a feature of items, and each cell is a high-dimensional vector. In multi-vector databases, the choice of indexes can have a significant impact on performance. Although index tuning for relational databases has been extensively studied, index tuning for multi-vector search remains unclear and challenging. In this paper, we define multi-vector search index tuning and propose a framework to solve it. Specifically, given a multi-vector search workload, we develop algorithms to find indexes that minimize latency and meet storage and recall constraints. Compared to the baseline, our latency achieves 2.1X to 8.3X speedup.
--------------------------------------------------------------------------------------------------------
MICE for CATs: Model-Internal Confidence Estimation for Calibrating Agents with Tools
This research introduces Model-Internal Confidence Estimators (MICE), a novel approach for improving tool-using AI agents' decision-making capabilities. By examining the internal state of language models across intermediate layers, MICE produces better-calibrated confidence scores that help determine when to use external tools. Tested on Llama3 models with the simulated trial and error tool-calling dataset, MICE outperforms existing methods on calibration metrics and significantly improves tool-calling utility. The system demonstrates impressive sample efficiency and can generalize to unseen APIs without additional training. This technology addresses a critical need for safer, more reliable AI agents that can appropriately weigh risk versus reward when interacting with external systems, with applications ranging from personal assistants to industrial automation and scientific research tools.
Authors: Nishant Subramani, Jason Eisner, Justin Svegliato, Benjamin Van Durme, Yu Su, Sam Thomson
Link: https://arxiv.org/abs/2504.20168v1
Date: 2025-04-28
Summary:
Tool-using agents that act in the world need to be both useful and safe. Well-calibrated model confidences can be used to weigh the risk versus reward of potential actions, but prior work shows that many models are poorly calibrated. Inspired by interpretability literature exploring the internals of models, we propose a novel class of model-internal confidence estimators (MICE) to better assess confidence when calling tools. MICE first decodes from each intermediate layer of the language model using logitLens and then computes similarity scores between each layer's generation and the final output. These features are fed into a learned probabilistic classifier to assess confidence in the decoded output. On the simulated trial and error (STE) tool-calling dataset using Llama3 models, we find that MICE beats or matches the baselines on smoothed expected calibration error. Using MICE confidences to determine whether to call a tool significantly improves over strong baselines on a new metric, expected tool-calling utility. Further experiments show that MICE is sample-efficient, can generalize zero-shot to unseen APIs, and results in higher tool-calling utility in scenarios with varying risk levels. Our code is open source, available at https://github.com/microsoft/mice_for_cats.
--------------------------------------------------------------------------------------------------------
This research examines how structured educational interventions affect undergraduate computer science students' use of generative AI tools. The researchers implemented "AI-Lab" modules emphasizing guided scaffolding and mindful engagement across four courses at Purdue University, collecting data from 831 matched pre-post surveys and focus groups over three semesters. While overall usage frequency remained stable, students showed significant increases in comfort and openness with AI tools across conceptual, debugging, and homework problems. The intervention successfully bridged the gap between naive usage and reflective integration, helping students maintain awareness of their skill development while leveraging AI benefits. These findings provide evidence-based recommendations for computing educators seeking to responsibly integrate generative AI into curricula without undermining essential competencies, addressing a critical challenge in contemporary computer science education.
Authors: Ethan Dickey, Andres Bejarano, Rhianna Kuperus, Bárbara Fagundes
Link: https://arxiv.org/abs/2505.00100v1
Date: 2025-04-30
Summary:
Generative AI (GenAI) is rapidly entering computer science education, yet its effects on student learning, skill development, and perceptions remain underexplored. Concerns about overreliance coexist with a gap in research on structured scaffolding to guide tool use in formal courses. This study examines the impact of a dedicated "AI-Lab" intervention -- emphasizing guided scaffolding and mindful engagement -- on undergraduate students in Data Structures and Algorithms, Competitive Programming, and first-year engineering courses at Purdue University. Over three semesters, we integrated AI-Lab modules into four mandatory and elective courses, yielding 831 matched pre- and post-intervention survey responses, alongside focus group discussions. Employing a mixed-methods approach, we analyzed quantitative shifts in usage patterns and attitudes as well as qualitative narratives of student experiences. While the overall frequency of GenAI usage for homework or programming projects remained largely stable, we observed large effect sizes in comfort and openness across conceptual, debugging, and homework problems. Notably, usage patterns for debugging also shifted statistically significantly, reflecting students' more mindful and deliberate approach. Focus group discussions corroborated these results, suggesting that the intervention "bridged the gap" between naive GenAI usage and more nuanced, reflective integration of AI tools into coursework, ultimately heightening students' awareness of their own skill development. These findings suggest that structured, scaffolded interventions can enable students to harness GenAI's benefits without undermining essential competencies. We offer evidence-based recommendations for educators seeking to integrate GenAI responsibly into computing curricula and identify avenues for future research on GenAI-supported pedagogy.
--------------------------------------------------------------------------------------------------------
Fiber to the Room: Key Technologies, Challenges, and Prospects
This comprehensive analysis explores Fiber to the Room (FTTR) technology, the next evolution in optical access networks designed to deliver room-level coverage with superior bandwidth and latency characteristics. The paper examines three critical technical aspects: centralized scheduling through MAC-PHY layer convergence, integrated management via extended OMCI protocols, and innovative energy-saving mechanisms. The researchers also explore cutting-edge AI integration and passive sensing capabilities that enable intelligent scheduling and environment-aware optimization. This technology represents a significant advancement in residential and commercial connectivity, with applications ranging from smart homes and immersive entertainment to industrial automation and healthcare. FTTR's architecture provides the foundation for future ultra-reliable, high-performance networks that will support increasingly demanding applications in the connected world.
Authors: Jinhan Cai, Xiaolong Zhang, Xiang Wang, Tianhai Chang, Gangxiang Shen
Link: https://arxiv.org/abs/2504.20433v1
Date: 2025-04-29
Summary:
Fiber to the Room (FTTR) is a next-generation access network designed to deliver high bandwidth, low latency, and room-level optical coverage. This paper presents a comprehensive analysis of the FTTR system architecture and protocol stack, focusing on three key technical aspects: centralized scheduling and control, integrated management and maintenance, and green energy-saving mechanisms. A simplified FTTR architecture based on the convergence of the medium access control (MAC) and physical (PHY) layers is introduced to enhance coordination and scheduling efficiency. An extended remote management scheme, based on the optical network unit management and control interface (OMCI), is described to enable unified control across main fiber units (MFUs) and sub-fiber units (SFUs). Furthermore, a service-aware energy-saving framework is discussed for dynamic power optimization. The paper also explores the integration of artificial intelligence (AI) and passive sensing into FTTR systems to support intelligent scheduling, energy management, and environment-aware optimization. These insights provide technical guidance for the scalable deployment and future evolution of FTTR networks.
--------------------------------------------------------------------------------------------------------
This systematic review examines the emerging field of Automated Meta-analysis (AMA), which addresses the growing challenge of synthesizing exponentially increasing scientific literature. After screening 978 papers and analyzing 54 studies across multiple domains, the researchers found that current AMA efforts predominantly focus on data processing (57%) rather than advanced synthesis stages. Notably, only one study attempted full-process automation, revealing a critical gap in the field. Despite recent advances in large language models, their integration into statistical modeling and higher-order synthesis remains underdeveloped. The review highlights distinct implementation patterns across medical (67%) and non-medical applications (33%), identifying opportunities to bridge automation gaps across all meta-analysis stages to realize AMA's potential for scalable, domain-agnostic evidence synthesis in an era of information overload.
Authors: Lingbo Li, Anuradha Mathrani, Teo Susnjak
Link: https://arxiv.org/abs/2504.20113v1
Date: 2025-04-28
Summary:
Exponential growth in scientific literature has heightened the demand for efficient evidence-based synthesis, driving the rise of the field of Automated Meta-analysis (AMA) powered by natural language processing and machine learning. This PRISMA systematic review introduces a structured framework for assessing the current state of AMA, based on screening 978 papers from 2006 to 2024, and analyzing 54 studies across diverse domains. Findings reveal a predominant focus on automating data processing (57%), such as extraction and statistical modeling, while only 17% address advanced synthesis stages. Just one study (2%) explored preliminary full-process automation, highlighting a critical gap that limits AMA's capacity for comprehensive synthesis. Despite recent breakthroughs in large language models (LLMs) and advanced AI, their integration into statistical modeling and higher-order synthesis, such as heterogeneity assessment and bias evaluation, remains underdeveloped. This has constrained AMA's potential for fully autonomous meta-analysis. From our dataset spanning medical (67%) and non-medical (33%) applications, we found that AMA has exhibited distinct implementation patterns and varying degrees of effectiveness in actually improving efficiency, scalability, and reproducibility. While automation has enhanced specific meta-analytic tasks, achieving seamless, end-to-end automation remains an open challenge. As AI systems advance in reasoning and contextual understanding, addressing these gaps is now imperative. Future efforts must focus on bridging automation across all meta-analysis stages, refining interpretability, and ensuring methodological robustness to fully realize AMA's potential for scalable, domain-agnostic synthesis.
--------------------------------------------------------------------------------------------------------
T2VPhysBench: A First-Principles Benchmark for Physical Consistency in Text-to-Video Generation
This research introduces T2VPhysBench, a groundbreaking benchmark that systematically evaluates whether text-to-video AI systems obey fundamental physical laws. Despite impressive aesthetic quality and instruction-following capabilities, current models frequently produce videos that violate basic physical principles such as rigid-body collisions and gravitational dynamics. Through rigorous human evaluation, the researchers tested both open-source and commercial systems against twelve core physical laws, finding that all models scored below 0.60 on average in each category. Even detailed physics-specific prompts failed to correct these violations, and models readily generated physically impossible scenarios when instructed. This benchmark provides crucial insights for developing truly physics-aware video generation systems with applications in education, training simulations, creative media production, and scientific visualization where physical plausibility is essential.
Authors: Xuyang Guo, Jiayan Huo, Zhenmei Shi, Zhao Song, Jiahao Zhang, Jiale Zhao
Link: https://arxiv.org/abs/2505.00337v1
Date: 2025-05-01
Summary:
Text-to-video generative models have made significant strides in recent years, producing high-quality videos that excel in both aesthetic appeal and accurate instruction following, and have become central to digital art creation and user engagement online. Yet, despite these advancements, their ability to respect fundamental physical laws remains largely untested: many outputs still violate basic constraints such as rigid-body collisions, energy conservation, and gravitational dynamics, resulting in unrealistic or even misleading content. Existing physical-evaluation benchmarks typically rely on automatic, pixel-level metrics applied to simplistic, life-scenario prompts, and thus overlook both human judgment and first-principles physics. To fill this gap, we introduce \textbf{T2VPhysBench}, a first-principled benchmark that systematically evaluates whether state-of-the-art text-to-video systems, both open-source and commercial, obey twelve core physical laws including Newtonian mechanics, conservation principles, and phenomenological effects. Our benchmark employs a rigorous human evaluation protocol and includes three targeted studies: (1) an overall compliance assessment showing that all models score below 0.60 on average in each law category; (2) a prompt-hint ablation revealing that even detailed, law-specific hints fail to remedy physics violations; and (3) a counterfactual robustness test demonstrating that models often generate videos that explicitly break physical rules when so instructed. The results expose persistent limitations in current architectures and offer concrete insights for guiding future research toward truly physics-aware video generation.
--------------------------------------------------------------------------------------------------------
On the Limitations of Steering in Language Model Alignment
This paper examines the effectiveness and limitations of steering vectors as a mechanism for aligning language model behavior during inference. Using transformer hook interventions and antonym-based function vectors, the researchers evaluate how prompt structure and contextual complexity affect steering success. Their findings indicate that while steering vectors show promise for specific alignment tasks like value alignment, they may not provide a robust foundation for general-purpose alignment, particularly in complex scenarios. The study establishes a methodological framework for investigating the steering capabilities of reasoning models, contributing important insights to the field of AI alignment. This research has significant implications for developing more reliable, value-aligned AI systems that can maintain intended behavior across diverse deployment contexts.
Authors: Chebrolu Niranjan, Kokil Jaidka, Gerard Christopher Yeo
Link: https://arxiv.org/abs/2505.01162v1
Date: 2025-05-02
Summary:
Steering vectors are a promising approach to aligning language model behavior at inference time. In this paper, we propose a framework to assess the limitations of steering vectors as alignment mechanisms. Using a framework of transformer hook interventions and antonym-based function vectors, we evaluate the role of prompt structure and context complexity in steering effectiveness. Our findings indicate that steering vectors are promising for specific alignment tasks, such as value alignment, but may not provide a robust foundation for general-purpose alignment in LLMs, particularly in complex scenarios. We establish a methodological foundation for future investigations into steering capabilities of reasoning models.
--------------------------------------------------------------------------------------------------------
This research addresses the critical security challenges of image data in resource-constrained IoT and edge networks. The proposed Feature-Aware Chaotic Image Encryption scheme combines three innovative components: Feature-Aware Pixel Segmentation that reorganizes pixels based on edge intensity, Chaotic Chain Permutation using logistic maps with dynamically updated keys, and Chaotic Chain Confusion employing seed matrices for bitwise operations. Security evaluations demonstrate near-zero pixel correlation, entropy values approaching the theoretical maximum of 8, and strong resistance against differential cryptographic attacks. The scheme's efficiency makes it suitable for real-time deployment in resource-limited environments. This technology has significant applications in secure visual data transmission for smart surveillance, medical imaging, autonomous vehicles, and privacy-preserving distributed learning systems where protecting visual information is paramount.
Authors: Muhammad Shahbaz Khan, Ahmed Al-Dubai, Jawad Ahmad, Nikolaos Pitropakis, Baraq Ghaleb
Link: https://arxiv.org/abs/2505.00593v1
Date: 2025-05-01
Summary:
The security of image data in the Internet of Things (IoT) and edge networks is crucial due to the increasing deployment of intelligent systems for real-time decision-making. Traditional encryption algorithms such as AES and RSA are computationally expensive for resource-constrained IoT devices and ineffective for large-volume image data, leading to inefficiencies in privacy-preserving distributed learning applications. To address these concerns, this paper proposes a novel Feature-Aware Chaotic Image Encryption scheme that integrates Feature-Aware Pixel Segmentation (FAPS) with Chaotic Chain Permutation and Confusion mechanisms to enhance security while maintaining efficiency. The proposed scheme consists of three stages: (1) FAPS, which extracts and reorganizes pixels based on high and low edge intensity features for correlation disruption; (2) Chaotic Chain Permutation, which employs a logistic chaotic map with SHA-256-based dynamically updated keys for block-wise permutation; and (3) Chaotic chain Confusion, which utilises dynamically generated chaotic seed matrices for bitwise XOR operations. Extensive security and performance evaluations demonstrate that the proposed scheme significantly reduces pixel correlation -- almost zero, achieves high entropy values close to 8, and resists differential cryptographic attacks. The optimum design of the proposed scheme makes it suitable for real-time deployment in resource-constrained environments.
--------------------------------------------------------------------------------------------------------
Digital Twin-Empowered Cooperative Autonomous Car-sharing Services: Proof-of-Concept
This innovative research demonstrates a digital twin-powered system for optimizing autonomous car-sharing services in real-world conditions. By integrating data from roadside units and connected autonomous vehicles within a campus environment digital twin, the system uses Age of Information metrics to maintain data freshness and dynamically adapt to changing traffic conditions. Proof-of-concept results show a 22% improvement in delivery efficiency compared to conventional shortest-path methods, while simulation results demonstrate a 12% overall efficiency improvement and 23% reduction in peak average AoI. This technology has significant implications for smart mobility solutions, particularly in complex urban environments where traditional routing approaches fall short. The successful implementation confirms the practical feasibility of cooperative driving systems that could transform urban transportation through more efficient, responsive vehicle routing.
Authors: Kazuma Nonomura, Kui Wang, Zongdian Li, Tao Yu, Kei Sakaguchi
Link: https://arxiv.org/abs/2504.20542v1
Date: 2025-04-29
Summary:
This paper presents a digital twin-empowered real-time optimal delivery system specifically validated through a proof-of-concept (PoC) demonstration of a real-world autonomous car-sharing service. This study integrates real-time data from roadside units (RSUs) and connected and autonomous vehicles (CAVs) within a digital twin of a campus environment to address the dynamic challenges of urban traffic. The proposed system leverages the Age of Information (AoI) metric to optimize vehicle routing by maintaining data freshness and dynamically adapting to real-time traffic conditions. Experimental results from the PoC demonstrate a 22% improvement in delivery efficiency compared to conventional shortest-path methods that do not consider information freshness. Furthermore, digital twin-based simulation results demonstrate that this proposed system improves overall delivery efficiency by 12% and effectively reduces the peak average AoI by 23% compared to the conventional method, where each vehicle selects the shortest route without considering information freshness. This study confirms the practical feasibility of cooperative driving systems, highlighting their potential to enhance smart mobility solutions through scalable digital twin deployments in complex urban environments.
--------------------------------------------------------------------------------------------------------
IK Seed Generator for Dual-Arm Human-like Physicality Robot with Mobile Base
This research tackles a fundamental challenge in humanoid robotics: solving inverse kinematics (IK) for human-sized service robots with mechanical limitations. The researchers developed an innovative method for generating optimal initial guesses for numerical IK solvers, using scaled Jacobian matrices to calculate manipulability while accounting for joint limits. By optimizing these initial conditions through genetic algorithms and leveraging reachability maps, the system significantly increases the probability of successfully solving complex IK problems. The method was validated in three typical household scenarios, demonstrating its practical application. This technology could dramatically improve the functionality of human-sized service robots in everyday environments, enabling more natural movement and manipulation capabilities for applications in elderly care, hospitality, healthcare, and household assistance where human-like physicality provides significant advantages.
Authors: Jun Takamatsu, Atsushi Kanehira, Kazuhiro Sasabuchi, Naoki Wake, Katsushi Ikeuchi
Link: https://arxiv.org/abs/2505.00871v1
Date: 2025-05-01
Summary:
Robots are strongly expected as a means of replacing human tasks. If a robot has a human-like physicality, the possibility of replacing human tasks increases. In the case of household service robots, it is desirable for them to be on a human-like size so that they do not become excessively large in order to coexist with humans in their operating environment. However, robots with size limitations tend to have difficulty solving inverse kinematics (IK) due to mechanical limitations, such as joint angle limitations. Conversely, if the difficulty coming from this limitation could be mitigated, one can expect that the use of such robots becomes more valuable. In numerical IK solver, which is commonly used for robots with higher degrees-of-freedom (DOF), the solvability of IK depends on the initial guess given to the solver. Thus, this paper proposes a method for generating a good initial guess for a numerical IK solver given the target hand configuration. For the purpose, we define the goodness of an initial guess using the scaled Jacobian matrix, which can calculate the manipulability index considering the joint limits. These two factors are related to the difficulty of solving IK. We generate the initial guess by optimizing the goodness using the genetic algorithm (GA). To enumerate much possible IK solutions, we use the reachability map that represents the reachable area of the robot hand in the arm-base coordinate system. We conduct quantitative evaluation and prove that using an initial guess that is judged to be better using the goodness value increases the probability that IK is solved. Finally, as an application of the proposed method, we show that by generating good initial guesses for IK a robot actually achieves three typical scenarios.
--------------------------------------------------------------------------------------------------------
This research addresses a fundamental challenge in traffic prediction: balancing long-term trends with short-term fluctuations. The researchers developed a hybrid deep learning framework that processes these complementary aspects of traffic dynamics in parallel, enhanced by Bahdanau attention mechanisms that selectively focus on critical time steps. Experimental results demonstrate significant improvements in prediction accuracy across multiple time horizons compared to baseline models, with the attention mechanism particularly enhancing short-term forecast precision. This approach offers substantial benefits for intelligent transportation systems, enabling more effective congestion mitigation, dynamic traffic management, and urban mobility planning. By improving the robustness and precision of traffic predictions, particularly for transient phenomena, this technology could transform how cities manage traffic flow and respond to changing conditions.
Authors: Adway Das, Agnimitra Sengupta, S. Ilgin Guler
Link: https://arxiv.org/abs/2504.19967v1
Date: 2025-04-28
Summary:
Traffic flow prediction is a critical component of intelligent transportation systems, yet accurately forecasting traffic remains challenging due to the interaction between long-term trends and short-term fluctuations. Standard deep learning models often struggle with these challenges because their architectures inherently smooth over fine-grained fluctuations while focusing on general trends. This limitation arises from low-pass filtering effects, gate biases favoring stability, and memory update mechanisms that prioritize long-term information retention. To address these shortcomings, this study introduces a hybrid deep learning framework that integrates both long-term trend and short-term fluctuation information using two input features processed in parallel, designed to capture complementary aspects of traffic flow dynamics. Further, our approach leverages attention mechanisms, specifically Bahdanau attention, to selectively focus on critical time steps within traffic data, enhancing the model's ability to predict congestion and other transient phenomena. Experimental results demonstrate that features learned from both branches are complementary, significantly improving the goodness-of-fit statistics across multiple prediction horizons compared to a baseline model. Notably, the attention mechanism enhances short-term forecast accuracy by directly targeting immediate fluctuations, though challenges remain in fully integrating long-term trends. This framework can contribute to more effective congestion mitigation and urban mobility planning by advancing the robustness and precision of traffic prediction models.
--------------------------------------------------------------------------------------------------------