Refactoring is a constant activity in software development and maintenance. Scale and maintain software systems are based on code refactoring. However, this process is still labor intensive, as it requires programmers to analyze the codebases in detail to avoid introducing new defects. In this research, we put forward a large language model (LLM)-based multi-agent system to automate the refactoring process on Haskell code. The objective of this research is to evaluate the effect of LLM-based agents in performing structured and semantically accurate refactoring on Haskell code. Our proposed multi-agent system based on specialized agents with distinct roles, including code analysis, refactoring execution, verification, and debugging. To test the effectiveness and practical applicability of the multi-agent system, we conducted evaluations using different open-source Haskell codebases. The results of the experiments carried out showed that the proposed LLM-based multi-agent system could average 11.03% decreased complexity in code, an improvement of 22.46% in overall code quality, and increase performance efficiency by an average of 13.27%. Furthermore, memory allocation was optimized by up to 14.57%. These results highlight the ability of LLM-based multi-agent in managing refactoring tasks targeted toward functional programming paradigms. Our findings hint that LLM-based multi-agent systems integration into the refactoring of functional programming languages can enhance maintainability and support automated development workflows.
Shahbaz Siddeeq, Muhammad Waseem, Z. Rasheed et al.· International Conference on...· 4 citations
Anomaly detection in smart power grids is a critical challenge due to the complexity, heterogeneity, and dynamic nature of sensor data streams. Existing one-class classification methods, particularly Subspace Support Vector Data Description (SVDD), have been extended to multimodal scenarios but often fail to fully exploit the structural dependencies across modalities, limiting their robustness in real-world applications. In this paper, we address this gap by proposing a generalized Multimodal Subspace Support Vector Data Description (MS-SVDD) model with graph-embedded regularization. The method projects data from multiple modalities into a shared low-dimensional subspace while preserving modality-specific structure through Laplacian regularizers. Our approach is evaluated on a three-modality dataset derived from smart grid event time series, using a dedicated preprocessing pipeline for constructing one-class classification training samples. The results demonstrate that our graph-embedded MS-SVDD improves robustness of event detection compared to conventional approaches, highlighting the potential of integrating graph priors with multimodal subspace learning for advancing anomaly detection in critical infrastructure. More broadly, this work contributes to the wider field of AI by illustrating how relational and structural information can be systematically embedded into one-class models, enabling robust learning under complex, high-dimensional, and multimodal conditions.
Thomas Debelle, F. Sohrab, Pekka Abrahamsson et al.· Scientific Reports· 1 citation
This paper presents MARARE, a real-time multi-agent system that transforms meeting dialogues into structured software requirements. One agent interacts with participants, while background agents extract and verify requirements collaboratively. Evaluation using the LLM-as-a-Judge method across five meetings (5–8 minutes each) shows a mean coverage of 80.0 ± 11.2 % (mean ± SD), semantic similarity of 0.86 ± 0.05, and hallucination rate of 14.3 ± 6.2 %. Preliminary results indicate performance differences across LLMs, suggesting that model choice influences coverage, consistency, and hallucination rates.
Malik Abdul Sami, Gessé Evangelista, Kai-Kristian Kemell et al.· AGENT@ICSE· 0 citations
Context: Organizations adopting Artificial Intelligence (AI) face challenges in eliciting and analyzing requirements that align with strategic objectives, especially when human oversight and iterative refinement are needed. Large Language Models (LLMs)-based Multi-agent systems provide a potential solution by supporting structured and collaborative Requirements Engineering (RE) processes for AI adoption planning.
Objective: The objective of this study is to investigate whether a multi-agent system, built on LLMs and supported by human input, can assist in requirements analysis for AI adoption. Method: We used a mixed-method approach: (i) designed and developed a multi-agent system to support the generation and prioritization of requirements for AI adoption, (ii) conducted multiple case studies with four companies to evaluate the system, and (iii) collected data through post-session questionnaires from nine participants and follow-up interviews, one per company.
Results: Questionnaire and interview findings together indicate that the system may assist in identifying relevant and goal-aligned requirements. Seven participants considered the generated requirements relevant, and six found them aligned with organizational goals. Participants noted that iterative feedback improved completeness and feasibility, often within two feedback rounds. Both data sources show that human input was essential to clarify technical details, ensure contextual accuracy, and validate prioritization results. Participants from all companies also identified usability, transparency, and scalability as areas requiring further refinement for broader organizational use.
Conclusions: LLM-based multi-agent systems can support strategic AI planning by enabling iterative refinement with human experts. Future work will include more interviews with stakeholders and adjustments to system features to improve transparency, usability, and scalability.
Malik Abdul Sami, Zheying Zhang, Muhammad Waseem et al.· e-Informatica Software Engin...· 6 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Large Language Models (LLMs) offer new opportunities for automated code refactoring. However, generated changes must reduce targeted quality problems without introducing new issues or altering behaviour-relevant code structures. We introduce REFINE (Refactoring with Evidence-aware Flow for Integrated ageNtic Execution), a tool-agnostic, evidence-aware multi-agent approach for generating Java file-level refactoring candidates. REFINE combines static-analysis-guided smell identification, smell-informed planning, LLM-based transformation, automated re-analysis, preservation checks, and structured reporting. We evaluate REFINE on 450 Java files from 15 open-source systems, producing 1,350 model-pass outputs using OpenAI GPT-5.5, Google Gemini 3.1 Pro Preview, and Anthropic Claude Opus 4.8. REFINE reduces detected code smells by 68.26%, 72.79%, and 68.49% across the three configurations, respectively, with the strongest reductions observed for major smells. A matched 150-file direct-prompt baseline shows that REFINE achieves a higher median code-smell reduction with smaller edits and fewer public-method removals. However, broader quality improvements are inconsistent, and preservation checks reveal residual risks, including assert/fail-call changes and public-method removal. Therefore, REFINE outputs should be treated as refactoring candidates requiring compilation, testing, dependency analysis, and human review before adoption in repository- or system-level settings.
Muhammad Waseem, Aakash Ahmad, Pekka Abrahamsson· 0 citations
560 field photographs of five durian disease categories, collected from commercial orchards across Peninsular Malaysia between July 2025 and June 2026, with per-image capture-session identifiers. The session identifiers are the point of this release. The 560 images come from only 73 independent capture sessions. A symptomatic leaf is normally photographed several times in a few seconds from slightly different angles, and those frames are not independent observations. Split this data at image level and near-identical views of one specimen land on both sides of the train/test boundary. In our own initial partition, 79.6% of images fell in sessions that straddled a split. Re-running the identical experiment with sessions kept whole lowered macro F1 by 12.2 points on average across nine architectures, positive in all nine and as much as 18.4 in one. Group your partitions by the session column in sessions.csv. Contents. images_fullres/ — the 560 originals as captured, in class folders. images_512/ — the same images at 512 px maximum edge, which is what the models were trained and evaluated on. sessions.csv — class, filename and session for every image. splits/session_level/ — the partition reported in the paper (446/58/56). splits/image_level/ — the control partition used to measure leakage (446/54/60). Evaluate at the resolution you train at. Every figure in the paper is computed on images_512. Running the same checkpoint over images_fullres through an identical Resize(256) and CenterCrop(224) pipeline gives 77.6% instead of 72.0% on the held-out set, because the two resampling paths to 224 px are not the same. The originals are included so the collection is complete, not because they are the working copy. Classes. Algal Leaf Spot (Cephaleuros virescens), Leaf Rot (Colletotrichum spp.), Phomopsis Fruit and Stem Blight (Phomopsis durionis), Pink Disease (Erythricium salmonicolor), Root Disease (Phytophthora spp.). Pink_disease is represented by three capture sessions in the entire collection; its per-class metrics are not interpretable at that support, and it is what bounds grouped cross-validation at k = 3. Annotation. Labels were assigned by the author under the guidance of growers and extension staff with field experience in these orchards. There was no second independent rater, so no inter-rater agreement statistic is available. Consent. Images were collected on site with the orchard owner's permission, or contributed by growers who were told at the time that the images would be released publicly for research. No images contain identifiable persons. A small number show a hand holding a leaf; that framing is part of the field condition being modelled. No location is published at finer resolution than district.
Lin Ding Shan· Zenodo (CERN European Organi...· 0 citations
Computer vision---making machines interpret images---traveled from blocks-world edge finders to deep convolutional networks matching human benchmarks, and its history is AI's most complete case of representation learning's triumph. This article presents a narrative review of the field's canonical line: Roberts's 1963 machine perception of solids, Marr's 1982 computational vision, Viola and Jones's 2001 face detection, Lowe's 2004 SIFT features, Dalal and Triggs's 2005 HOG descriptors, Felzenszwalb and colleagues' 2010 deformable part models, Szeliski's 2010 synthesis, Girshick's 2015 Fast R-CNN, Long, Shelhamer, and Darrell's 2015 fully convolutional nets, Simonyan and Zisserman's 2015 VGG, He and colleagues' 2016 ResNet, and Redmon and colleagues' 2016 YOLO. The synthesis is organized around three themes: representation, in which hand-engineered features gave way to learned hierarchies; architecture, in which convolution, regions, and residual depth solved recognition's geometry; and tasks, in which classification widened into detection, segmentation, and real-time video. It is concluded that vision's deep learning settlement reorganized the field around data and compute---and that its open problems, robustness and embodiment, define the current frontier.
Zen Revista, 10 IA· Zenodo (CERN European Organi...· 0 citations
Abstract: The growing combination of geospatial technologies, Artificial Intelligence (AI), and edge computing is changing the field of spatial analysis, environmental monitoring, and infrastructural design. This article gives a thorough summary of the way modern computer science approaches—namely machine learning (ML), deep learning (DL), container orchestration using Kubernetes, and ultra-reliable low-latency communications (URLLC)—are being incorporated into geospatial geoinformatics. Instead of carrying out processing in centralised cloud systems, geospatial systems can now handle high-resolution Earth Observation (EO) data, LiDAR point clouds, and Internet of Things (IoT) spatial streams in near real-time by moving the processing tasks to the network edge. We look systematically at the basic methods involved in spatial intelligence, containerized orchestration, multi-sensor data fusion, and edge deployment architectures. Moreover, we combine the more recent literature from a range of disciplines to show the way in which spatial technologies directly contribute to the UN Sustainable Development Goals (SDGs), help reduce regional environmental degradation, and improve university-based entrepreneurial ecosystems. Lastly, the main research gaps—such as the problem of bandwidth limitations in remote areas, model drift in changing environments, and governance constraints—are identified, together with specific future directions for next-generation spatial computing.
Dr. Ambrose Ndubuisi Ekebuike*, Abdulaziz Ahmad, Yusuf Aliyu Adamu· Zenodo (CERN European Organi...· 0 citations
Archived source release of BioMCP-TS v0.8.0 (npm package biomcp): 41 core tools across 15 registration modules federating 33 verified upstream data hosts, plus optional plugin sets for read-only SQL analytics and in-process WebAssembly compute (Bioconductor DESeq2/edgeR/limma under webR; samtools/bedtools/bcftools under biowasm). Zero-config stdio MCP server for Node >= 22.13, single runtime dependency. Repository: https://github.com/yeyuan98/biomcp-ts (live development continues there; this record archives the exact tag evaluated in the Technical Report).
Ye Yuan· Zenodo (CERN European Organi...· 0 citations
This technical data package establishes the engineering blueprints, economic integration thesis, and provisional utility patent specifications for the Lawrence Aero Brick Solid-State Transformer (LAB-SST). The LAB-SST is a zero-silicon, plasma-state power transformation architecture that replaces physical Silicon Carbide (SiC) semiconductor lattices and copper electromagnetic windings with a dynamically generated atmospheric Universal Dielectric Barrier Discharge (U-DBD) plasma gateway. By driving a non-linear Townsend avalanche at an electric field threshold of E ≥ 56 kV/mm, the system acts as an asymmetric electrical valve, rectifying alternating current (AC) grid power directly into ±400V or 800V High-Voltage Direct Current (HVDC) for hyperscale computing and AI infrastructure. To eliminate the severe thermal degradation and parasitic cooling loads associated with multi-megawatt solid-state switching, the LAB-SST utilizes Electrohydrodynamic (EHD) momentum transfer to generate a solid-state convective vacuum (Δm = 0). System stability is maintained by Phase-Shifted Agentic Swarm (PSAS) logic running on edge NPU hardware, which truncates pulse widths (< 10 ns) whenever space-charge growth approaches Meek's criterion (α·d ≥ 14) to prevent thermal arc transitions. Economic Integration: Across a 100 MW hyperscale deployment, the LAB-SST obsoletes a USD 22,000,000 SiC SST installation with a USD 10,000,000 plasma-state alternative, while eliminating 10 MW in parasitic cooling load to retain USD 8,760,000 in annual cooling OpEx. This repository includes the complete USD 67,500 Bill of Materials (BOM) required for TRL-4 benchtop prototype validation. Note: Accompanying system schematics are conceptual AI-assisted architectural renderings; refer to the technical specifications and patent claims for precise operational mathematics and material parameters.
Charles Clark Lawrence· Zenodo (CERN European Organi...· 0 citations
Artificial intelligence (AI), digital phenotyping, passive sensing, and edge computing are increasingly being explored to extend psychiatric assessment beyond intermittent clinical encounters. This review evaluates the evidence relevant to a privacy-preserving, closed-loop architecture in which environmental and wearable signals are processed locally, converted into abstract behavioral biomarkers, and used to support longitudinal diagnostic reasoning. The literature indicates that digital phenotyping can capture clinically relevant changes in sleep, activity, mobility, communication, speech, and social behavior, with the most consistent evidence observed for bipolar disorder and depressive symptoms. However, external validation is uncommon, study samples are often small, missingness is frequently ignored, labels are temporally misaligned with sensor streams, and reported discrimination can therefore overestimate real-world performance. Edge AI offers lower latency, reduced network exposure, and bandwidth savings, but it introduces device constraints and does not by itself guarantee privacy. Privacy-preserving strategies should therefore combine data minimization, local feature extraction, strict retention limits, cryptographic separation of identity, access control, and auditable governance. Recent work on evidence conflict further shows that diagnostic AI may be sensitive to the order and presence of contradictory information, supporting the need for explicit contradiction detection, uncertainty calibration, abstention, and clinician escalation rather than probability accumulation alone. The central translational finding is that the components of the proposed Psychiatric Diagnostic Assistant Device (PDAD) are individually supported by adjacent evidence, but the specific combination of continuous behavioral biomarker extraction, privacy-preserving edge processing, contradiction-aware inference, and information-gain-driven adaptive sensor control has not yet been clinically validated as an integrated psychiatric diagnostic system. The most defensible next step is a staged prospective programme using independent reference standards, external validation, safety monitoring, privacy verification, and a clinical utility design aligned with current AI reporting and evaluation guidance.
M. A. N. Makeen· Zenodo (CERN European Organi...· 0 citations
The Selection-Stitch Model derives quark charges in thirds from the geometry of a trapped lattice defect. The integer unit that lifts -1/3 to the up-type +2/3 has resisted every mechanism tried: it is not a Cartan charge of the vacuum-matter algebra under the lattice's geometric C and P, and a circuit-ordering geometric phase was excluded by computation. This paper derives the unit from the one place those exclusions left open: the topology of the gauge sector. The framework's electromagnetic channels live on the <100> square faces; the lattice nodes and octahedral voids together form a simple cubic grid, and a compact U(1) on that grid carries monopole defects as a matter of compactness. The chiral defect worldline of the companion paper - directed hopping with point-gap winding W = +1, derived from the verification front - couples to this gauge field, and the pre-registered calculation is a spectral flow: thread one flux quantum through the worldline and count the charge pumped. The answer, computed as a determinant winding and therefore an integer identically, is exactly W: +1 in the physical configuration, 0 in a static crystal, -1 with the frontier inverted, robust across system sizes and every reference inside the point gap. By anomaly matching this boundary flow is the edge of a bulk theta = 2*pi*W, and by the Witten effect flux-bound matter shifts its electric charge by e*theta/2pi = e*W: the charge ladder is q = -1/3 + W, and the up-type +2/3 is derived. Because a determinant winding cannot be fractional, the mechanism cannot touch the thirds: the geometric and topological charges live in different places and add. Two further results follow. The companion matter paper's empirical restriction of the winding to w >= 0 is now derived - negative winding requires an inverted frontier, which the physical configuration forbids - matching the non-observation of charge -4/3; and the framework now requires monopole-type defects of the photon grid, a falsifier-shaped commitment whose microscopic construction is stated as open. Every claim carries a status tag, and the pre-registration rides in the verification archive.
Raghu Kulkarni· Zenodo (CERN European Organi...· 0 citations
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
MIT News · Artificial Intelligence· news.mit.eduAug 27, 2026
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
