Quantum kernel methods are a candidate approach for machine learning on near-term quantum hardware, but two practical problems limit their deployment: kernel values concentrate exponentially as the qubit count grows, and real-world datasets are often severely class-imbalanced. We present the first systematic empirical study of how these two effects interact. We define two diagnostic metrics, the Discrimination Ratio (DR) and the Quantum Imbalance Vulnerability Score (QIVS), which measure whether quantum kernels retain minority-class separability under concentration. Experiments span ten log-spaced imbalance ratios, five random seeds, five qubit counts (4 to 12), and five real-world datasets, and yield three results. First, DR stays above 1.0 at every qubit count tested (4 to 12), so the discriminative signal survives concentration. Second, at extreme imbalance (IR below 0.003), quantum kernel SVMs retain positive discriminative signal, crossing above DR=1.0 by IR≈0.0028, while the classical oversampling methods SMOTE and ADASYN produce zero minority-class recall throughout the same regime, a practical advantage for quantum kernels at the imbalance extreme. Third, QIVS follows a broadly monotonic decreasing trend, falling from 13.25 to 5.54 as the imbalance ratio increases, with a single minor fluctuation smaller than the cross-seed variability we measure elsewhere in the sweep. This trend makes QIVS a reliable diagnostic for practitioners choosing quantum kernels on imbalanced tasks.
Souvik Ghosh, Vijaysherly V.· 2026 International Conferenc...· 0 citations
Variational Quantum Classifiers (VQCs) have emerged as prime candidates for machine learning on NISQ systems. It stands to reason that the same depolarizing noise that drives quantum states toward the maximally mixed state would also alleviate the overconfidence of VQCs. This paper tests that hypothesis through an empirical study on three datasets at six noise levels, validated across ten random seeds and supported by a formal analysis of how the depolarizing channel contracts the measured logits. The primary finding is that depolarizing noise does not reduce overconfidence in VQCs that remain in the learnable regime: expected calibration error (ECE) never decreases with noise on any dataset, and on the lowest-variance dataset it increases significantly (Wilcoxon signed-rank p < 0.005 over ten seeds). We derive why the optimizer compensates for the channel and confirm the mechanism through a confidence-trajectory experiment. A secondary finding is that a less expressive ansatz can appear well-calibrated only because it collapses to a degenerate solution, demonstrating that ECE must always be reported alongside accuracy. Any effect of noise on accuracy is small and seed-dependent, and it is decoupled from calibration. Post-hoc temperature scaling reduces VQC ECE by 64 to 77 percent across all datasets and is the recommended calibration method for NISQ-era classifiers.
Souvik Ghosh, Amrita Kundu, Mithaguru et al.· 2026 International Conferenc...· 0 citations
Large Language Models (LLMs) trained using Chain-of-Thought (CoT) supervision have achieved state-of-the-art performance on complex reasoning tasks. However, the generation of long reasoning chains introduces substantial computational overhead during inference, limiting their deployment in low-latency and resource-constrained environments. This paper proposes AdaptiReason, a novel framework that dynamically compresses intermediate reasoning steps based on task difficulty and model confidence without requiring retraining of the underlying base model. AdaptiReason employs a lightweight difficulty estimator to determine the appropriate reasoning depth for each input, followed by a learned token-pruning policy that eliminates redundant or low-information reasoning steps. Experimental evaluation on the MATH, GSM8K, and ARC-Challenge benchmarks demonstrates that AdaptiReason reduces the average number of generated tokens by 3.7× while preserving 98.2% of the baseline reasoning accuracy. Furthermore, the proposed framework is model-agnostic and can be seamlessly integrated with instruction-tuned LLMs without requiring access to model parameters, relying solely on output logits for adaptive reasoning compression. The results demonstrate that AdaptiReason significantly improves inference efficiency while maintaining high reasoning performance, making it suitable for real-time and resource-constrained LLM applications.
V. A, Mithaguru, Amrita Kundu et al.· 2026 4th International Confe...· 0 citations
Federated Learning (FL) offers a privacy-preserving framework for training energy forecasting models across geographically distributed nodes without centralising raw metering data. However, the accuracy of classical FL algorithms degrades substantially when client data is heterogeneous, a condition that is the norm rather than the exception in distributed energy systems. This paper presents a novel systematic benchmark study quantifying accuracy degradation across four distinct heterogeneity axes: quantity skew (Dirichlet $\alpha$), covariate shift, temporal non-stationarity, and client availability. Here, five FL algorithms (FedAvg, FedProx, SCAFFOLD, Local-Only, and Centralized independent and identically distributed (IID) oracle) are evaluated with long short-term memory (LSTM) and Neural Basis Expansion Analysis for Time Series (N-BEATS) architectures across public datasets (NREL Solar, NREL Wind, Pecan Street, UK Smart Meter), generating over 190 experimental conditions. This degradation gap analysis reveals that standard FedAvg suffers up to 83.4% Mean Absolute Error (MAE) degradation relative to the centralised IID baseline under heavy quantity skew on solar data, with temporal non-stationarity emerging as the dominant single-axis contributor.
Souvik Ghosh, Amrita Kundu, Mithaguru et al.· 2026 Third International Con...· 0 citations