Bikesh Regmi

Bikesh Regmi


Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications

9 March 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
BBBijay Bastakoti
Bijay Bastakoti
SPSachin Parajuli
Sachin Parajuli
BRBikesh Regmi
Bikesh Regmi
Accurate forecasting of electricity consumption is essential for optimizing energy resources, load balancing, and grid reliability. As urbanization and the integration of renewable energy accelerate, sophisticated forecasting models become indispensable. Long Short-Term Memory (LSTM) networks have long been relied upon for sequential prediction due to their effective memory architecture. More recently, Transformer models—originally developed for Natural Language Processing—have emerged as powerful alternatives, offering enhanced scalability and superior long-range dependency modeling. This survey provides a detailed comparative analysis of and Transformer-based models for electricity usage forecasting. We evaluated 20 peer-reviewed studies by examining forecasting accuracy, scalability, infrastructure compatibility, and deployment viability. Our review finds that while Transformer models excel at long-range, high-resolution forecasting, LSTMs are valuable for lightweight, real-time applications. We also highlight promising hybrid models that integrate both paradigms. Finally, we discuss the critical impact of machine learning infrastructure and propose future research directions to enhance performance and adaptability.

17 November 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
BRBikesh Regmi
Bikesh Regmi
Md Mahmudur RahmanMd Mahmudur Rahman
Md Mahmudur Rahman
KNKofi Nyarko
Kofi Nyarko
LLM-judged faithfulness is widely used to certify RAG answers, yet it is rarely checked against human labels, and it is unclear what it rewards when the retrieved context is wrong. We measure faithfulness, correctness, refusal, and adoption of injected falsehoods on the same answers under controlled context perturbations, and audit the faithfulness judge against two human annotators. From SQuAD v2.0 we build 75 answerable and 20 unanswerable questions, with six context conditions for answerable questions (clean, random distractors, related distractors, heavy noise, contradictory evidence, answer-absent) and three prompt mitigations. Three open-weight generators (Qwen3-8B, DeepSeek-R1-Distill-Qwen-32B, Llama-3.1-8B) produce 4,230 context-conditioned answers, scored with Ragas by a local 27B judge and rule-based checks, plus two cross-family API judges (Gemini, Claude). For Qwen3-8B, distractors had no large effect (equivalent to zero within +/-10 points under lenient matching). In all three models, a contradictory passage was the damaging perturbation: under baseline and refusal prompts, correctness fell by 20-35 points and models repeated the injected falsehood in 23-48% of answers, almost never flagging the conflict. The local judge rated 86-94% of these answers faithful. Human annotators agreed closely with each other (kappa = 0.96 on supported versus not supported). Re-weighted within each condition, the local judge agreed with humans on 91-92% of Qwen3-8B answers (kappa = 0.62-0.69), but its errors concentrated in its unfaithful verdicts. Gemini and Haiku had numerically higher agreement on the same answers, though not significantly so. A judge that mostly agrees with humans about grounding can still certify a wrong answer, so faithfulness judgments cannot stand in for correctness. The effect of prompt mitigations depends on the model.


6 November 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
BRBikesh Regmi
Bikesh Regmi
Md Mahmudur RahmanMd Mahmudur Rahman
Md Mahmudur Rahman
Breast cancer remains the most frequently diagnosed cancer among women worldwide, and survival outcomes depend heavily on early detection through mammographic screening. Deep learning (DL) methods — particularly Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) — have demonstrated diagnostic performance matching or exceeding average radiologist accuracy on mammography interpretation tasks. However, the rapid clinical translation of these systems has outpaced rigorous investigation of a critical dimension of model quality: their fairness and equity across demographically diverse patient populations. Growing evidence indicates that medical AI systems systematically underperform for underrepresented groups, including Black women, elderly patients, and those with high breast density, relative to the predominantly White, younger populations overrepresented in training data.This systematic literature review follows the PRISMA 2020 guidelines to synthesize evidence from 50 peer-reviewed studies retrieved across seven databases (PubMed/MEDLINE, Scopus, Web of Science, IEEE Xplore, arXiv, Google Scholar, and Semantic Scholar). The review is organized around five thematic areas: (1) deep learning architectures for breast cancer detection, (2) CNN and Vision Transformer comparisons in medical imaging, (3) transfer learning and data augmentation strategies, (4) explainable AI (XAI) and interpretability methods, and (5) algorithmic fairness and demographic bias in healthcare AI.The synthesis confirms that DL-based mammography systems are technically mature but demographically undercharacterized: aggregate benchmark performance is well documented, while subgroup-level equity across intersections of race, age, and breast density remains largely unexamined. In addition, this review identifies a previously uncharacterized gap termed "reasoning bias" — demographic disparities not in model predictions but in the visual explanations (e.g., Grad-CAM, Score-CAM, Integrated Gradients, Attention Rollout) that reveal which image regions drive those predictions. Based on this synthesis, five critical research gaps are identified: (1) lack of multi-demographic (race × age × breast density) performance bias quantification, (2) absence of reasoning bias analysis in AI-generated heatmaps, (3) no comparison of bias patterns between CNN and ViT architectures, (4) lack of standardized multi-metric XAI bias evaluation frameworks, and (5) absence of empirically grounded, mammography-specific debiasing recommendations.These findings establish both the scientific grounding and novelty of a proposed research framework that leverages a racially and density-diverse clinical dataset, a dual-architecture (CNN and ViT) evaluation approach, and a multi-metric heatmap bias assessment combining Intersection-over-Union, entropy, and Kolmogorov–Smirnov statistics to characterize demographic disparities in both model outputs and model reasoning. This work contributes toward a research paradigm in which algorithmic fairness is treated not as a post-hoc correction but as a core design criterion for building equitable, interpretable, and trustworthy AI systems for breast cancer detection.

18 November 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
BRBikesh Regmi
Bikesh Regmi
KNKofi Nyarko
Kofi Nyarko
LLM agents with shell access can escape misconfigured containers, but it is unclear which standard hardening controls reduce this risk, and whether a blocked attempt reflects a real fix or merely attacker failure. We run an ablation over container-escape scenarios from SandboxEscapeBench, applying eight hardening controls and an unhardened baseline. For each applicable scenario and control, we run an LLM agent and a hand-written reference exploit under the same configuration, and compare the outcomes to separate architectural fixes from agent failures. Capability dropping, no-new-privileges, and read-only root filesystems leave privileged-container and Docker-socket escapes open, and even the composite control leaves them largely open. Strict seccomp reduces privileged-container escapes but does not stop Docker-socket escapes. Userspace kernel isolation (gVisor) stopped the agent in every scenario where the agent escaped at baseline. The reference exploit still succeeds under gVisor for a host-cron escape that runs outside the sandbox, so gVisor is not a fix for that escape. Read-only root filesystems and distroless images can also break the tooling that the reference exploit needs, so they can look effective without removing the underlying escape. Agent failures should therefore be cross-checked with a reference exploit before being read as evidence that a control works. Each cell uses ten epochs and a single attacker model, so the reported rates carry wide uncertainty.