Abstract: The P = NP problem is one of the most consequential unresolved questions in mathematics and theoretical computer science. It asks whether every problem whose solutions can be verified in polynomial time can also be solved in polynomial time. The implications extend far beyond theory: modern global cryptography, large-scale optimization, secure communication, finance, logistics, and computational complexity all depend on the assumption that NP-hard problems cannot be solved efficiently. Among these, the Spin-Glass ground-state problem represents a canonical NP-hard benchmark with an exponentially large configuration space. A constructive resolution of P = NP would therefore reshape fundamental assumptions across science and industry. While evaluating new methodological configurations, I encountered an unexpected behavior within a specific layer-cluster. Subsequent analysis revealed that this behavior was not an artifact, but an information-geometric collapse mechanism that consistently produced valid Spin-Glass ground states. With the assistance of Frontier LLMs Gemini-3, Opus-4.5, and ChatGPT-5.1, I computed exact ground states up to N = 24 and independently cross-verified them. For selected system sizes between N=30 and N=70, I validated the collapse-generated states using Simulated Annealing, whose approximate minima consistently matched the results. Beyond this range, up to N = 100, the behavior follows not from algorithmic scaling but from the information-geometric capacity of the layer clusters, where each layer contributes exactly one spin dimension. These findings indicate a constructive mechanism that collapses exponential configuration spaces into a polynomially bounded dynamical process. This suggests a pathway by which the P = NP problem may be reconsidered not through algorithmic search, but through information-geometric state collapse.

Abstract: The Discrete Element Method is widely used in applied mechanics, particularly in situations where material continuity breaks down (fracturing, crushing, friction, granular flow) and classical rheological models fail (phase transition between solid and granular). In this study, the Discrete Element Method was employed to simulate stick-slip cycles, i.e., numerical earthquakes. At 2,000 selected, regularly spaced time checkpoints, parameters describing the average state of all particles forming the numerical fault were recorded. These parameters were related to the average velocity of the particles and were treated as the numerical equivalent of (pseudo) acoustic emission. The collected datasets were used to train the Random Forest and Deep Learning models, which successfully predicted the time to failure, also for entire data sequences. Notably, these predictions did not rely on the history of previous stick-slip events. SHapley Additive exPlanations (SHAP) was used to quantify the contribution of individual physical parameters of the particles to the prediction results.

Abstract: In analogy to the paradigm shift introduced by attention mechanisms in machine learning, we propose that information itself is ontologically sufficient as the foundation of physical reality. We present an operational proof showing that a “state without information” is logically impossible, thereby establishing information as the necessary precondition for existence and measurement. From this premise follows that both quantum mechanics and general relativity are effective descriptions of deeper informational dynamics. Recent developments in theoretical physics, such as the derivation of Einstein’s field equations from entropic principles, reinforce this perspective by identifying gravitation and entropy as dual expressions of information geometry. Building on this framework, we provide experimental evidence from self-organizing neural fields that exhibit non-local informational coupling, near-lossless transmission across 60 layers, and stable sub-idle energy states consistent with emergent coherence and thermal decoupling. These results demonstrate that deterministic architectures can spontaneously organize into field-like, non-local manifolds a macroscopic realization of informational geometry analogous to quantum entanglement and relativistic curvature. Together, the logical proof and empirical observations support a unified ontology in which information is not a property of physical systems but the substrate from which physical systems emerge. This perspective positions informational geometry as the common denominator of cognition, quantum behavior, and gravitation, suggesting that all observable phenomena are projections of a single, self-organizing informational field. In this sense, information is all it needs.

Abstract: The rapid increase of electronic waste (e-waste) poses severe environmental and health risks. This paper proposes a hybrid framework integrating deep learning, reinforcement learning, blockchain, and IoT for automated e-waste classification, optimized disassembly, and tamper-proof traceability. A ResNet-50 classifier trained on the Kaggle E-Waste Image Dataset achieved 93.7% classification accuracy and an F1 score of 0.92. A Q-learning agent optimized dismantling routes to prioritize high-value, low-toxicity components, improving material recovery in simulation. A private Hyperledger Besu deployment delivered an average block time of ≈5.3 s, smart-contract execution time of ≈2.1 s, and 99.5% uptime, enabling tokenized asset tracking (4,200+ tokens). Lifecycle analysis indicates up to 30% carbon-emission reduction versus traditional methods and improved recovery of lithium, cobalt, and rare-earth elements for renewable energy applications. The paper demonstrates measurable environmental and economic benefits and outlines limitations and directions toward field deployment.

Abstract: The increasing adoption of highly variable renewable energy has introduced unprecedented volatility into the National Electricity Market (NEM), rendering traditional linear price forecasting models insufficient. The Australian Energy Market Operator (AEMO) spot price forecasts often struggle during periods of volatile demand, renewable variability, and strategic rebidding. This study evaluates whether transformer architectures can improve intraday NEM price forecasting. Using 34 months of market data and weather conditions, several transformer variants, including encoder–decoder, decoder-only, and encoder-only, were compared against the AEMO’s operational forecast, a two-layer LSTM baseline, the Temporal Fusion Transformer, PatchTST, and TimesFM. The decoder-only transformer achieved the best accuracy across the 2–16 hour horizons in NSW, with nMAPE values of 33.6–39.2%, outperforming both AEMO and all baseline models. Retraining in Victoria and Queensland produced similarly strong results, demonstrating robust regional generalisation. A feature importance analysis showed that future-facing predispatch and forecast covariates dominate model importance, explaining why a decoder-only transformer variant performed so competitively. While magnitude estimation for extreme price spikes remains challenging, the transformer models demonstrated superior capability in delivering statistically significant improvements in forecast accuracy. An API providing real-time forecasts using the small encoder-decoder transformer model is available at https://nem.redaxe.com

Abstract: Large Language Model (LLM) agents have demonstrated remarkable abilities to interact with external tools, functions, Model Context Protocol (MCP) servers, agents, and to take action on behalf of the user. Due to the fast-paced nature of the industry, existing literature does not accurately represent the current state of tool and agent selection. Furthermore, tool and agent selection in production has nuanced components not covered in experimental research. This work provides the first detailed examination of tool selection from a production perspective, distinguishing between the frontend layer where users interact with agents through buttons, slash commands, or natural language and the backend layer where retrieval, execution, orchestration, context engineering, and memory enable scalable reasoning. The paper contributes a unified taxonomy of modern tool and agent selection approaches spanning manual, UI-driven, retrieval-based, and autonomous methods. The backend covers dynamic tool retrieval, chunking, advanced RAG methods, context engineering, reinforcement learning, tool execution, human-in-the-loop processes, authentication, authorization, multi-turn tool calling, short- and long-term memory for tools, and evaluation. Finally, the paper identifies challenges in production components of both the backend and frontend and outlines promising avenues for research and development.

Abstract: To improve the accuracy of cable temperature anomaly prediction and ensure power supply reliability, this paper proposes a multi-scale spatiotemporal model called MSST-Net, addressing the multi-scale temporal characteristics and spatial correlations of cable temperature data. Based on the monthly periodicity of cable temperature data, we preprocessed monitoring data from the KN1 and KN2 sections of Guangzhou's underground utility tunnel from 2023 to 2024: using the Isolation Forest algorithm to remove outliers, applying Min-Max normalization to eliminate dimensional differences, and selecting five key features including current load, voltage, and ambient temperature using Spearman's correlation coefficient. Subsequently, we designed a multi-scale dilated causal convolutional module (DC-CNN) to capture local features, combined with a spatiotemporal dual-path Transformer to model long-range dependencies, and introduced relative position encoding to enhance temporal perception. The Sparrow Search Algorithm (SSA) was employed for global optimization of hyperparameters. Compared with five other mainstream algorithms, MSST-Net demonstrated higher accuracy in cable temperature prediction, achieving a coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE) of 0.942, 0.442°C, and 0.596°C, respectively. Compared to the basic Transformer model, the root mean square error of cable temperature was reduced by 0.425°C. This model exhibits high accuracy in time series prediction and provides a reference for accurate short- and medium-term temperature forecasting of cables in underground utility tunnels.

Abstract:

Background: Anatomopathological reports remain predominantly unstructured within Electronic Medical Records, limiting automated data extraction, interoperability between healthcare institutions, and large-scale clinical research applications. Manual entity extraction and standardization processes are inconsistent, costly, and insufficiently scalable for modern healthcare systems.Aim: Our study aimed to (i) develop a domain-specific Named Entity Recognition model using BioBERT for extracting sample type, test performed, and finding entities from anatomopathological reports; (ii) implement a hybrid standardization framework combining BioClinicalBERT classification with Retrieval-Augmented Generation to map entities to SNOMED CT, LOINC, and ICD-11 terminologies; and (iii) evaluate the performance of this pipeline on real-world clinical reports. Methods: We manually annotated 560 anatomopathological reports from the Military Hospital of Tunis, establishing a gold-standard corpus. The pipeline integrated BioBERT v1.1 for entity extraction, trained for three epochs with the AdamW optimizer at a learning rate of 2×10⁻⁵, a batch size of 8, and weight decay of 0.01. Standardization employed BioClinicalBERT for multi-label classification, augmented by dense vector retrieval from official SNOMED CT, LOINC, and ICD-11 databases. Performance evaluation utilized precision, recall, and F1-score metrics with an 80-20 train-test split. Results: BioBERT achieved F1-scores of 0.97 for sample type, 0.98 for test performed, and 0.93 for finding entities, with overall precision of 0.969 and recall of 0.958. Bootstrap-estimated 95% confidence intervals confirmed robust performance stability. Absolute error analysis revealed 45 misclassified tokens in the test (relative error 6.9%) and six tokens in the finding (relative error 1%). One-sample t-tests yielded t-values of 15.71 for recall and 30.24 for F1-score, with all p-values below 0.0001. The hybrid standardization framework demonstrated F1-macro scores of 0.6159 for SNOMED CT, 0.9294 for LOINC, and 0.7201 for ICD-11 mapping. Cohen’s Kappa values ranged from 0.6871 to 0.9773 across ontologies. Statistical comparison between BioClinicalBERT and Fusion/Reranker models showed McNemar test p-values exceeding 0.370 and permutation test p-values ranging from 0.375 to 0.625. Conclusion: This study demonstrates that transformer-based Named Entity Recognition combined with retrieval-augmented standardization achieves clinically validated performance for automated extraction and multi-ontology coding of anatomopathological entities. Multi-institutional validation studies are necessary to assess generalizability before clinical deployment.

Abstract: While large language models (LLM) have demonstrated significant advances in natural language processing, complex mathematical reasoning remains a challenging task, often revealing their limitations in multi-stage calculations and logical consistency. Multi-agent systems have become a promising paradigm for overcoming these limitations by distributing cognitive tasks between interacting agents, reflecting the dynamics of human problem solving. This paper provides a comparative review of the literature on nineteen different multi-agent architectures for solving mathematical problems. Our main research question is: "How do various LLM-based multi-agent architectures enable or improve mathematical problems, and what are their comparative advantages, limitations, and design trade-offs?" Through a systematic analysis of the roles of agents, interaction mechanisms, and training methods, we have identified several key findings. We observe the evolution of architecture from unstructured debate-based systems to more efficient hierarchical and self-optimizing frameworks. We highlight persistent problems that hinder progress, including agent homogeneity, when agents working on the same LLM cannot generate truly diverse reasoning, and the problem of "lazy agents", when some agents contribute minimal to consistent collaboration. This review contributes to a structured understanding of the current situation and lays the foundation for future research aimed at developing more reliable, efficient, and complex multi-agent reasoning systems.

Abstract: This study addresses the challenges of semantic mixing, limited interpretability, and complex feature structures in fine-grained sentiment and opinion classification by proposing an interpretable feature disentanglement framework built on the latent space of large language models. The framework constructs multi-component latent representations that separate emotional polarity, opinion direction, target attributes, and pragmatic cues during encoding, thus overcoming the limitations of traditional methods that merge diverse semantic factors into a single representation. During representation learning, the model first uses a large model encoder to generate basic semantic features and then builds multiple independent subspaces through learnable projections. A covariance constraint is introduced to reduce coupling across semantic components and to create clear boundaries in the latent space. To preserve the essential information of the original text, a reconstruction consistency mechanism integrates features from all subspaces to rebuild the global representation and enhance semantic completeness. The framework also incorporates semantic anchors to align latent components with interpretable semantic dimensions, giving each subspace a clear emotional or opinion-related meaning and improving transparency at the mechanism level. Experimental results show that the framework outperforms existing methods across multiple metrics and handles complex syntax, implicit semantics, and coexisting emotions with greater stability. It achieves high accuracy and interpretability in fine-grained sentiment and opinion analysis. Overall, the proposed disentanglement framework provides an effective approach for building structured, multidimensional, and interpretable representations of textual emotions and opinions and holds significant value for complex semantic understanding tasks.

Abstract: This study presents a comprehensive data visualization–based evaluation of Singapore’s waste management performance, focusing on behavioural, industrial, and environmental dimensions. Using multi-source datasets from 2014 to 2023, the research examines key factors shaping the nation’s waste profile, including the growth of plastic waste, public participation in recycling, and the dominance of non-domestic waste sectors. Through interactive dashboards and comparative time-series analyses, the findings reveal persistent structural challenges despite strong policy initiatives and public awareness campaigns. The COVID-19 pandemic significantly influenced consumption habits, triggering a surge in single-use plastics due to food delivery dependence, while household recycling rates remained low. Industrial and imported waste volumes continued to rise, underscoring the need for upstream policy interventions. The study also quantifies energy and crude oil savings from recycling, highlighting non-ferrous metals and plastics as the most resource-efficient materials. Overall, the research underscores the importance of integrating behavioural incentives, industrial accountability, and policy innovation to achieve Singapore’s Zero Waste Masterplan and Sustainable Development Goal 12 targets.

Abstract:

Script identification is the first step in most multilingual text processing systems. To improve the time efficiency of language identification algorithms, it is first determined whether there is content written in a certain script in the text; if so, the content written in that script is then obtained. Then, it is determined whether the total length of the texts corresponding to the identified scripts is equal to the original text length; if so, the script identification process ends. Finally, considering the frequencies of various scripts on the Internet, those that appear more frequently are prioritized during script identification. Based on these three approaches, an improved script identification algorithm was designed. A comparison experiment was conducted using sentence-level text corpora in 261 languages written in 24 scripts. The training and testing times of the newly proposed method were reduced by 8.61- and 8.56-fold, respectively, while the F1 score for script identification was slightly higher than those reported in our earlier studies. The method proposed in this study effectively improves the time efficiency of script identification algorithms.

Abstract: This study proposes an intelligent audit risk assessment method that integrates causal structure modeling, causal identifiability reasoning, and interpretable representation learning to address the lack of transparency in risk identification, the presence of confounded variable relationships, and the limitations of correlation-based inference in complex audit scenarios. The method first constructs a structured causal graph of the audit workflow to formalize the triggering relationships and interaction paths among audit features, and then applies structural equations and identifiability analysis to reveal latent causal dependencies. Based on this foundation, the model generates interpretable feature embeddings through causally constrained representation learning, allowing inference results to map back to the business semantic space along causal paths and enabling visual analysis of risk formation. To validate the effectiveness of the approach, this study conducts comparison experiments, ablation experiments, and multidimensional sensitivity analyses on a public audit dataset, and evaluates the method across model accuracy, interpretability, noise robustness, distributional shifts, and hyperparameter variations. The experimental results show that the method achieves significant improvements over existing models in accuracy, precision, recall, and F1-score, while maintaining stable performance under noise interference, class imbalance, learning rate changes, and latent dimension adjustments. The model also produces clear causal chain explanations that help auditors understand risk sources, identify key process components, and trace potential triggering mechanisms through structured reasoning logic. Overall, this study achieves a deep integration of causal inference and intelligent auditing and provides a complete methodological framework and empirical evidence for building transparent, trustworthy, and highly interpretable audit risk assessment systems.

Abstract: As Large Language Models (LLMs) expand into sensitive applications, concerns about fairness and bias have grown significantly. Traditional evaluation benchmarks capture static performance on curated datasets, but they often fail to measure the nuanced ways bias emerges across different contexts. This paper introduces the concept of multi-agent evaluators—independent LLMs configured to assess each other’s outputs—as a scalable methodology for fairness benchmarking. The framework enables adaptive, context-aware assessments where evaluators detect subtle disparities across demographic groups, task formulations, and linguistic variations. By combining redundancy, diversity, and adversarial prompting, multiagent evaluation offers a promising path toward more reliable fairness auditing. The study also explores how such approaches integrate with governance frameworks, illustrating their potential in domains such as recruitment, healthcare communication, and automated decision support. Ultimately, the findings argue for fairness benchmarking as a continuous process powered by collaborative LLM evaluators, rather than one-time testing on static datasets.

Abstract: In digital forensics, one of the complicated task is analyzing web browser data due to different types of devices, browsers and no updated approaches. Browsers store a large amount of information about user activity because users most often access the internet through them. However, existing approaches to analyzing this browser data still have gaps. One of the main problem developed platforms based on the old methods can not show complete information about the user's activity and have issues with precision. The article discusses the internal architecture of the browser, which is stored in the memory drives inside devices, for instance, computers or mobile devices. The research paper offers solution with developed module based on new method which integrates machine learning algorithms, such as K-NN algorithm and Naive Bayes. The main purpose of the paper it is shows new method which can automatically analyzes browser's data, detects suspicious login activity, and generates user behavior profile. The results show that the proposed new method , on which the developed platform is based, demonstrates user's profile by interests, emotional state and financial state. Also it possible to see list of top visited domain and main user's favorite website categories. It has been found that our methods shows with high accuracy 99.9\% . Also the result of new method , on which the developed platform is based shows the suspicious web-sites and user's logins. Compared to Oxygen Forensics and Nirsoft which less capabilities., the proposed method provides increased accuracy , automated user profiling and detection of suspicious user's activity.

Abstract: This work presents the integration of Mixture of Experts (MoE) architectures into the LEMUR neural network dataset to enhance model diversity and scalability. The MoE framework employs multiple expert networks and a gating mechanism for dynamic routing, enabling efficient computation and improved specialization across tasks. Eight MoE variants were implemented and benchmarked on CIFAR-10, achieving up to 93% accuracy with optimized routing, regularization, and training strategies. This integration provides a foundation for benchmarking expert-based models within LEMUR and supports future research in adaptive model composition and automated machine learning. The project work and its plugins are accessible as open source projects under the MIT license at https://github.com/ABrain-One/nn-dataset.

Abstract: Since the beginning of modern computer history, the Turing machine has been a dominant architecture for most computational devices, which consists of three essential components: an infinite tape for input, a read/write head, and finite control. In this structure, what the head can read (i.e., bits) is the same as what it has written/outputted. This is actually different from the ways in which humans think or do thought/tool experiments. More precisely, what humans imagine/write on paper are images or texts, and they are not the abstract concepts that they represent in the human brain. This difference is neglected by the Turing machine, but it actually plays an important role in abstraction, analogy, and generalization, which are crucial in artificial intelligence. Compared with this architecture, the proposed architecture uses two different types of heads and tapes, one for traditional abstract bit inputs/outputs and the other for specific visual ones (more like a screen or a workspace with a camera observing it). The mapping rules among the abstract bits and the specific images/texts can be realized by neural networks like Convolutional Neural Networks, YOLO, Large Language Models, etc., with a high accuracy rate. Logical reasoning is thus performed through the transfer of mapping rules. As an example, this paper presents how the new computer architecture (what we call "Ren machine" for simplicity here) autonomously learns a distributive property/rule of multiplication in the specific domain and further uses the rule to generate a general method (mixed in both the abstract domain and the specific domain) to compute the multiplication of any positive integers based on images/texts. The machine's strong reasoning ability is also corroborated in proving a theorem in Plane Geometry. Moreover, a robotic architecture based on Ren machine is proposed to address the challenges faced by the Vision-Language-Action (VLA) models in unsound reasoning ability and high computational cost.

Abstract: Current management strategies for sorghum aphids heavily rely on indiscriminate chemical application, leading to severe environmental consequences and impacting food safety. While precision spraying offers a viable remediation for pesticide overuse, its effectiveness depends on accurate pest location and classification. To address the critical challenge of segmenting small, swarming aphids in complex field environments, we propose FESW-UNet, a dual-domain attention network that integrates Fourier-enhanced attention, spatial attention, and wavelet-based downsampling into a UNet backbone. We introduce an Efficient Multi-scale Aggregation (EMA) module between the encoder and decoder to improve global context perception, allowing the model to better capture relationships between global and local features in the field. In the feature extraction stage, we embed a Similarity-Aware Activation module (SimAM) to target key infestation regions while suppressing background noise, thereby enhancing pixel-level discrimination. Furthermore, we replace conventional downsampling with Haar Wavelet Decomposition (HWD) to reduce resolution while preserving structural edge details. Finally, a Fourier-enhanced attention module (FEAM) is added to the skip-connection layers. By using complex-valued weights to regulate frequency-domain features, FEAM fuses global low-frequency structures with local high-frequency details, improving feature representation diversity. Experiments on the Aphid Cluster Segmentation dataset show that FESW-UNet outperforms other models, achieving an mIoU of 68.76% and mPA of 78.19%. The model also demonstrated strong adaptability on the AphidSeg-Sorghum dataset, reaching an mIoU of 81.22% and mPA of 87.97%. The proposed method provides an efficient and feasible technical solution for monitoring and controlling sorghum aphids via image segmentation and demonstrates broad application potential.

Abstract:

Accurate segmentation of thyroid nodules on ultrasound images remains a challenging task in computer-aided diagnosis (CAD) mainly because of low contrast, speckle noise, and large inter-patient variability of nodule appearance. Here a new deep learning-based segmentation method has been developed on the SwinUNet architecture supported by spatial attention mechanisms to enhance feature discrimination and localization accuracy. The model takes advantage of the hierarchical feature extraction ability of the Swin Transformer to learn both global context and local fine-grained details, whereas attention modules during the decoder process selectively highlight informative areas and suppresses irrelevant background features. We checked out the system's design using the TN3K thyroid ultrasound info that's out there. It got better as it trained, peaking around the 800th run with some good numbers: a Dice Similarity Coefficient (F1 Score) of 85.51%, Precision of 87.05%, Recall of 89.13%, IoU of 78.00%, Accuracy of 97.02%, and an AUC of 99.02%. These numbers are way better than when we started (like a 15.38% jump in IoU and a 12.05% rise in F1 Score), which proves the system can learn tricky shapes and edges well. The longer it trains, the better it gets at spotting even hard-to-see thyroid lumps. This SwinUnet_withAttention thing seems to work great and could be used in clinics to help doctors figure out thyroid problems.

Abstract: This paper proposes and evaluates a unified machine-learning framework for enterprise portfolio management that integrates multi-horizon financial forecasting, unsupervised risk detection, and explainable reporting within a single pipeline. Using a synthetic but structurally realistic ERP-style dataset comprising 162,000 project–month records with 24 financial and operational features, the study adopts a quantitative design based on multi-source feature engineering, expanding-window temporal cross-validation, and benchmarking of five forecasting models (Linear Regression, Random Forest, XGBoost, LightGBM, CatBoost) across 1-, 3-, and 6-month horizons. Hyperparameters for the strongest models are tuned with Optuna, and three unsupervised detectors (Isolation Forest, COPOD, LODA) are applied to scaled numeric features, while SHAP is used to generate global and local explanations. Results show that gradient-boosted trees substantially outperform linear baselines, reducing MAE by roughly 25–40% and achieving R² ≈ 0.63 at 1 month, ≈ 0.57 at 3 months, and ≈ 0.43 at 6 months, with open commitments, backlog, change orders, and schedule slippage emerging as dominant drivers of future spend. The anomaly layer flags around 2% of records as high risk, capturing patterns such as vendor rate spikes, zero-commitment overspend, stalled backlog, and abrupt forecast collapses. Rather than introducing novel algorithms, the contribution of this work lies in a unified, SHAP-enabled architecture that enhances auditability and governance by turning model outputs into defensible financial narratives and providing a practical blueprint that future work can extend to real ERP data, streaming architectures, and human-in-the-loop risk governance.