Abstract The rapid proliferation of Large Language Models (LLMs) has established benchmark evaluations as the primary mechanism for assessing model capability and technological progress. However, growing concerns regarding benchmark validity, data contamination, and the interpretability of aggregate scores highlight a critical gap in understanding how these metrics reflect the broader LLM ecosystem. This study addresses this gap by conducting a comprehensive descriptive analysis of benchmark... The rapid proliferation of Large Language Models (LLMs) has established benchmark evaluations as the primary mechanism for assessing model capability and technological progress. However, growing concerns regarding benchmark validity, data contamination, and the interpretability of aggregate scores highlight a critical gap in understanding how these metrics reflect the broader LLM ecosystem. This study addresses this gap by conducting a comprehensive descriptive analysis of benchmark performance and capability stratification across contemporary LLMs. Utilizing the Comprehensive LLM Benchmark Dataset, comprising 390 model-benchmark observations from 2022 to 2024, we employ descriptive statistics, density estimation, and performance-tier categorization to map the performance landscape. Our findings reveal a negatively skewed distribution with a high median but substantial variability, indicating that while baseline competencies are standardizing, significant capability gaps persist. Furthermore, the analysis identifies distinct capability strata, with Strong and Top Tier models accounting for over 56 per cent of observations, yet a substantial proportion of Weak and Moderate performers remain. These results demonstrate that the contemporary LLM landscape is highly stratified rather than homogeneous. This stratification highlights the need for delicate evaluation. Ultimately, this research underscores that aggregate benchmark scores often obscure underlying heterogeneity in capabilities. We conclude that future evaluation frameworks must evolve toward multidimensional, capability-oriented methodologies to accurately capture model maturity and real-world utility, providing a foundational baseline for subsequent research on scaling laws and architectural effectiveness. Read More Read Less

Abstract High-dimensional correlated multivariate Bernoulli systems exhibit exponentially large state spaces, making the characterization of rare events and probability concentration a significant computational challenge. This study presents a comprehensive analytical framework to investigate sparsity, dominance, and tail behavior within such discrete binary systems. Utilizing a benchmark dataset of 20 correlated Bernoulli variables, we systematically analyze state spaces across increasing dimensionalities (4 to 20 bits) through tail... High-dimensional correlated multivariate Bernoulli systems exhibit exponentially large state spaces, making the characterization of rare events and probability concentration a significant computational challenge. This study presents a comprehensive analytical framework to investigate sparsity, dominance, and tail behavior within such discrete binary systems. Utilizing a benchmark dataset of 20 correlated Bernoulli variables, we systematically analyze state spaces across increasing dimensionalities (4 to 20 bits) through tail probability evaluation, Herfindahl Hirschman Index (HHI) concentration metrics, activation pattern characterization, and Generalized Extreme Value (GEV) modeling. Results reveal extreme sparsity, with over 99.9% of configurations at the highest dimension possessing probabilities below 10-4 . Despite a theoretical state space of 1,048,576 configurations, probability mass is heavily concentrated: a single dominant state captures approximately 20% of the total mass, and the top 100 states account for over 50%. Consequently, the effective number of states remains remarkably low (approximately 23.51), indicating that practical system complexity is vastly smaller than its combinatorial size. Furthermore, GEV modeling confirms that extreme configurations follow a predictable, heavy-tailed statistical regime rather than random fluctuations. These findings demonstrate that high-dimensional correlated binary systems are governed by a small subset of highly activated configurations. This insight enables efficient, reduced state representations crucial for advancing reliability analysis, anomaly detection, and risk assessment in complex stochastic networks. Read More Read Less

Abstract and disease pathogenesis. Although machine learning (ML) has shown promise in predicting methylation status from genomic features, many predictive models operate as “black boxes,” limiting biological interpretability and clinical translation. Objective: This study develops an explainable machine learning framework for DNA methylation prediction that integrates predictive modeling with multiple interpretability techniques to identify influential genomic determinants and elucidate their biological relevance. Methods: A supervised classification model... and disease pathogenesis. Although machine learning (ML) has shown promise in predicting methylation status from genomic features, many predictive models operate as “black boxes,” limiting biological interpretability and clinical translation. Objective: This study develops an explainable machine learning framework for DNA methylation prediction that integrates predictive modeling with multiple interpretability techniques to identify influential genomic determinants and elucidate their biological relevance. Methods: A supervised classification model was trained on a DNA methylation dataset comprising four explanatory variables CpG density, genomic location, regulatory score, and conservation score with binary methylation status as the target. Global interpretability was achieved using SHAP (SHapley Additive exPlanations) feature importance and summary plots. Local explanations were generated using SHAP waterfall and force plots alongside LIME (Local Interpretable Model Agnostic Explanations). Functional relationships were examined through Partial Dependence Plots (PDP) and Individual Conditional Expectation (ICE) plots. Robustness was assessed via bootstrap resampling (1,000 iterations), five-fold cross-validation, and correlation analysis between SHAP and LIME explanations. Results: CpG density emerged as the most influential predictor (mean absolute SHAP = 0.36), followed by regulatory score (0.24), conservation score (0.13), and genomic location (0.05). SHAP dependence plots revealed nonlinear threshold effects, with CpG density values below 0.3 producing negative contributions and values above 0.7 strongly increasing methylation probability. PDP analyses confirmed saturation behavior at high feature values. Local explanations successfully decomposed individual predictions,demonstrating that high confidence methylation calls (e.g., P = 0.92) were driven primarily by CpG density and regulatory score. Robustness analyses showed strong agreement between SHAP and LIME (Pearson r = 0.97) and low cross validation variability (coefficient of variation <10% for top features). Conclusion: Explainable machine learning provides biologically interpretable insights into DNA methylation determinants while maintaining predictive capability. The consistent identification of CpG density, regulatory activity, and evolutionary conservation as dominant predictors aligns with established epigenetic principles. This framework offers a transparent, reproducible approach for epigenetic biomarker discovery and supports precision medicine applications. Future work should extend these methods to multi-omic datasets and experimental validation. Read More Read Less