Artificial Intelligence

2602 Submissions

[5] viXra:2602.0135 [pdf] submitted on 2026-02-23 19:40:12

On the Limits of Prompt Repetition: A Multi-Asset Evaluation of LLM Inference Heuristics for Financial Time-Series Prediction

Authors: Avinash Chaurasiya
Comments: 20 Pages. (Note by viXra Admin: Please submit article written with AI assistance to ai.viXra.org)

Prompt repetition has recently been proposed as a simple inference-time modificationcapable of improving the performance of non-reasoning large language models(LLMs). By duplicating the input prompt, the technique aims to improve attentionutilization without incurring additional computational cost. While empirical gainshave been reported on deterministic language benchmarks, it remains unclearwhether such improvements generalize to stochastic prediction domains whereuncertainty originates from external information rather than prompt structure.In this work we conduct a systematic, multi-asset evaluation of prompt repeti-tion in financial time-series forecasting, spanning four representative instruments:GOOGL, MSFT, NVDA, and GLD. We compare a logistic-regression baselineagainst LLM predictions under both standard prompting and prompt repetition,assessing directional accuracy, Brier score, bootstrap confidence intervals, McNemarsignificance tests, and calibration reliability diagrams. Across all assets and allmetrics we find no statistically meaningful improvement attributable to promptrepetition. We further provide an information-theoretic proof showing that anytransformation preserving input entropy cannot increase predictive mutual infor-mation in noise-dominated environments. Our findings establish a clear boundarycondition for prompt-engineering techniques and underscore the necessity of domain-aware evaluation before deploying LLM inference strategies beyond natural languageprocessing.
Category: Artificial Intelligence

[4] viXra:2602.0117 [pdf] submitted on 2026-02-21 20:09:44

Advancing U.S. Public Health Surveillance: Leveraging Generative AI and Agentic Systems for the National HIV Behavioral Surveillance Program

Authors: Satyadhar Joshi
Comments: 14 Pages. (Note by viXra Admin: Please submit article written with AI assistance to ai.viXra.org)

This comprehensive technical paper responds to the Centers for Disease Control and Prevention's Federal Register notice (Docket No. CDC-2025-0753) concerning the revision of the National HIV Behavioral Surveillance System (NHBS). We propose an integrated framework leveraging Generative AI (GenAI) and agentic systems to enhance the NHBS data collection methodology across 21 Metropolitan Statistical Areas (MSAs). Our approach addresses all five evaluation criteria specified by the Office of Management and Budget: (1) necessity and practical utility, (2) accuracy of burden estimates, (3) enhancement of data quality, utility, and clarity, (4) minimization of respondent burden through technology, and (5) assessment of information collection costs. Drawing on recent research in AI-assisted surveying, we demonstrate how Large Language Models (LLMs), adaptive interviewing systems, and human-AI hybrid frameworks can transform NHBS from a periodic cross-sectional survey into a dynamic, real-time surveillance tool while reducing the estimated 3,398-hour annual burden. We provide detailed implementation recommendations for the proposed three-year cycle, addressing ethical considerations, validation requirements, and quality assurance protocols for deployment in public health settings. This expanded framework includes comprehensive technical specifications, cost-benefit analyses, and risk mitigation strategies to support evidence-based decision-making for CDC leadership.
Category: Artificial Intelligence

[3] viXra:2602.0055 [pdf] submitted on 2026-02-08 17:33:59

Efficient PII Removal with Zero-Shot NER and Dual Encoders

Authors: Rutvik Acharya, Nitin Agarwal
Comments: 5 Pages.

Personally Identifiable Information (PII) removal is a critical task in data privacy and security, requiring the identification and redaction of sensitive entities such as names, addresses, and social security numbers from unstructured text. Traditional Named Entity Recognition (NER) models used for PII removal are limited to predefined entity types, necessitating retraining for each new PII category. This paper presents zero-shot NER architectures that enable the efficient removal of any type of PII without extensive retraining.We leverage two advanced architectures for zero-shot NER in the context of PII removal: bi-encoder and poly-encoder models. The bi-encoder architecture separates the encoding of input text and PII entity types into distinct transformer models, allowing for efficient and scalable processing. PII entity type encodings can be pre-computed and reused across different input texts, reducing computational overhead. The poly-encoder architecture enhances the bi-encoder approach by incorporating a post-fusion step to model interactions between input text and PII entity representations explicitly, addressing the lack of inter-entity understanding in standalone bi-encoder models.To evaluate the effectiveness of these architectures for PII removal, we conduct experiments using a diverse, high-quality dataset containing various types of PII. We compare the performance of our proposed models with existing zero-shot NER approaches, such as GLiNER, in terms of precision, recall, and F1 score. The results demonstrate that our bi-encoder model outperforms GLiNER in identifying and removing PII entities, setting a new benchmark for zero-shot NER in the context of data privacy and security.These architectures offer several advantages for PII removal, including the ability to recognize an unlimited number of PII entities simultaneously, faster inference with preprocessed PII entity embeddings, and better generalization to unseen PII categories. These advancements enable the development of efficient and scalable PII removal systems capable of handling diverse and evolving PII requirements, ensuring compliance with data privacy regulations and protecting sensitive information.In this paper, we present an adaptive approach to PII detection that dynamically selects between GLINER and Presidio models based on contextual analysis. Our methodology first analyzes input text for regional markers, script patterns, and format variations to determine the most suitable model for PII detection. GLINER is prioritized for Western contexts and standardized formats, while Presidio handles region-specific and non-standard patterns. This context-aware selection is complemented by a robust validation framework that includes both primary and secondary validation layers, confidence scoring, and enhanced processing for ambiguous cases. Experimental results demonstrate an 12%-14% improvement in overall accuracy compared to single-model approaches, with particularly strong performance in handling diverse regional formats and multi-script environments, while maintaining acceptable processing overhead.
Category: Artificial Intelligence

[2] viXra:2602.0030 [pdf] submitted on 2026-02-05 11:05:58

Neural Networks with {-1,0,1} Weights Playing Atari Space Invaders (1) Trained by Evolution Strategy

Authors: Hidehiko Okada
Comments: 8 Pages.

In prior work, discrete-weight neural networks trained via evolutionary algorithms have been investigated, demonstrating the feasibility of binary-weight models on reinforcement learning tasks including Atari Space Invaders. In this study, we extend this line of research by evaluating ternary-weight neural networks with weights in {-1,0,1} and comparing their performance with binary-weight counterparts {-1,1}. Using Evolution Strategy to train multilayer perceptron controllers for the Atari Space Invaders task, the author analyzes the effects of weight representation and evolutionary hyperparameters. Experimental results show that ternary-weight networks achieved higher average performance than binary-weight networks with identical architectures, although the difference was not statistically significant. Additionally, a larger population size combined with fewer generations was found to be more effective than smaller populations with longer training durations, consistent with prior findings. These results suggest that population size plays a critical role in compensating for the limited global search capability of ES.
Category: Artificial Intelligence

[1] viXra:2602.0020 [pdf] submitted on 2026-02-03 20:52:10

An Approach to Identify Machine Learning and Deep Learning Methods and Data to Understand Common Time Variant Models That Can Forecast Both Crypto Currencies and Equities

Authors: Sanath Shenoy
Comments: 27 Pages. (Note by viXra Admin: Please submit article written with AI assistance to ai.viXra.org)

Cryptocurrencies and Equity markets have been the most attractive investment in the modernworld. While these are very attractive, they have been subject to a lot of volatility in their behaviour due to multiple reasons, such as macroeconomic conditions and regulation in various economic activities within countries and the world as a whole. There has been minimal research in this area to use Machine learning and deep learning approach to help predict the price of cryptocurrency and equities and other behaviours. There has also been very minimal research to understand the relationship on whether the change in the value of equities has an effect on the value of cryptocurrencies and vice versa. These effects are due to correlational causal or any other kind of relationship between the values of both the investment asset classes. This research aims to identify models based on Machine learning and deep learning that can predict the price or value of cryptocurrencies and equities.
Category: Artificial Intelligence