Artificial Intelligence

Performance of AI Tools for Automated Literature Screening in Evidence Synthesis

Authors: Minghao Ruan, Junhao Fan
Comments: 10 Pages.

Background Literature screening constitutes a critical component in evidence synthesis; however, it typically requires substantial time and human resources. Artificial intelligence (AI) has shown promise in this field, yet the accuracy and effectiveness of AI tools for literature screening remain uncertain. This study aims to evaluate the performance of several existing AI-powered automated tools for literature screening.Methods This diagnostic accuracy study employed a cohort to evaluate the performance of five AI tools—ChatGPT 4.0, Claude 3.5, Gemini 1.5, DeepSeek-V3, and RobotSearch—in literature screening. We selected a random sample of 1,000 publications from a well-established literature cohort, with 500 as randomized controlled trials (RCTs) group and 500 as others group. Diagnostic accuracy was measured using several metrics, including the false negative fraction (FNF), time used for screening, false positive fraction (FPF), and the redundancy number needed to screen.Results We reported the FNF for the RCTs group and the FPF for the others group. In the RCTs group, RobotSearch exhibited the lowest FNF at 6.4% (95% CI: 4.6% to 8.9%), whereas Gemini exhibited the highest at 13.0% (95% CI: 10.3% to 16.3%). In the others group, the FPF of the four large language models ranged from 2.8% (95% CI: 1.7% to 4.7%) to 3.8% (95% CI: 2.4% to 5.9%), both of which were significantly lower than RobotSearch's rate of 22.2% (95% CI: 18.8% to 26.1%). In terms of screening efficiency, the mean time used for screening per article was 1.3 s for ChatGPT, 6.0 s for Claude, 1.2 s for Gemini, and 2.6 s for DeepSeek.Conclusions The AI tools assessed in this study demonstrated commendable performance in literature screening; however, they are not yet suitable as standalone solutions. These tools can serve as effective auxiliary aids, and a hybrid approach that integrates human expertise with AI may enhance both the efficiency and accuracy of the literature screening process.
Category: Artificial Intelligence

AI Accountability Infrastructure: Cryptographically Verifiable Decision Provenance for High-Stakes AI Systems

Authors: Mezbah Uddin Rafi
Comments: 19 Pages. (Note by viXra Admin: Please submit article written with AI assistance to ai.viXra.org)

Consequential decisions in law, medicine, finance, and public services are increasingly mediated by AI systems, yet no widely adopted infrastructure allows an authorized third party to later verify, with cryptographic assurance, what observable factors caused a specific AI output. Explainable AI (XAI) methods characterize model behavior in general, not a single transaction; governance frameworks specify process, not evidentiary mechanism; and forensic/audit logging is rarely designed for AI-specific provenance (retrieved context, tool invocation, policy evaluation) or for graduated, privacy-preserving legal disclosure. We systematize this gap and propose the AI Accountability Infrastructure (AAI), centered on a formally defined, hash-chained, digitally signed AI Accountability Record (AAR) and a graduated, threshold-cryptography-enforced disclosure model in which no single party — including the AI provider — holds unilateral access to evidentiary content. We give (i) a formal model of record integrity and non-repudiation with informal correctness arguments grounded in established cryptographic primitives; (ii) a STRIDE-based threat model spanning both AI-specific and classical evidentiary risks; (iii) a privacy analysis using established data-minimization and disclosure-control concepts; (iv) a complexity and overhead analysis of the proposed mechanisms; and (v) an evaluation methodology intended for future empirical validation. This paper is explicitly a framework and formal-model contribution: it does not report novel experimental results, and we state this limitation directly rather than fabricate benchmark figures. We situate AAI against the NIST AI Risk Management Framework, ISO/IEC 42001, and the EU AI Act, and identify open problems suitable for a doctoral research program.
Category: Artificial Intelligence

Prompt Injection Attacks in Large Language Model Systems: Threats, Mechanisms, and Defense Strategies

Authors: Iqra Zulfiqar
Comments: 7 Pages.

AI systems based on Large Language Models are now everywhere, from chatbots to autonomous agents that can book flights, send emails, and make financial decisions. But this power comes with a serious problem.These models cannot tell the difference between real instructions and fake ones hidden inside data. Attackers exploit this weakness through what researchers call prompt injection attacks. This paper reviews how these attacks work, what damage they can cause, and what defenses exist today. We look at different attack types, from simple text tricks to complex automated attacks that spread across multiple AI agents. Wealso review what current defenses can and cannot do. The conclusion is not comfortable: no defense today fully solves the problem, and the attacks keep getting smarter.
Category: Artificial Intelligence

Best Play for Imperfect Players and Game Tree Search: Part I (Theory)

Authors: Eric B. Baum, Warren D. Smith
Comments: 41 Pages.

This is part I (theory) of a 2-part paper about a new, non-minimaxing, approach to search in game-trees such as Chess and Go. Its first point is that minimaxing is NOT the best decision-making (i.e. chess-move-selection) process based on inexact merely-statistical estimates about the "goodness" of the chess positions at the leaves of a search tree. Its second point is that going down the game tree to some fixed depth, then stopping, is NOT the best choice of chess-positions for your search to explore. "BPIP search" is a way to make better decisions based on any particular search tree, and a way to explore better trees. As of year 2026, BPIP still has not yet been used by chess and go programs, but if were might enable them to get substantially stronger. Would it? I would like to know, but never found out because I abandoned my attempts to write strong chess and go programs as just too difficult a programming task for me.

I originally wrote this paper and put it on the internet in 1993, and at that time "Monte Carlo game tree search" and today's highly advanced selective alphabeta variants did not yet exist, computer chess was weaker than the best humans, and computer go far weaker than decent human amateurs. Later, go programmers invented "Monte Carlo game tree search" which can be regarded as a "poor man's BPIP" (although they did not realize that). MCGTS has been used in go programs with great success.
Category: Artificial Intelligence

Best Play for Imperfect Players and Game Tree Search: part II (experiments)

Authors: Warren D. Smith, Eric B. Baum, Charles Garrett, Rico Tudor
Comments: 39 Pages.

This is part II (experiments) of a 2-part paper about a new, non-minimaxing, approach to search in game-trees such as Chess and Go. Its first point is that minimaxing is NOT the best decision-making (i.e. chess-move-selection) process based on inexact merely-statistical estimates about the "goodness" of the chess positions at the leaves of a search tree. Its second point is that going down the game tree to some fixed depth, then stopping, is NOT the best choice of chess-positions for your search to explore. "BPIP search" is a way to make better decisions based on any particular search tree, and a way to grow better trees to explore. As of year 2026, BPIP still has not yet been used by chess and go programs, but if were might enable them to get substantially stronger. Would it? I would like to know, but never found out because I abandoned my attempts to write strong chess and go programs as just too difficult a programming task for me. However in this paper we did program BPIP and alphabeta gameplayers for a number of simpler games, achieving strengths apparently stronger than any human and also among the strongest year-1995 computer programs. For some games BPIP indeed outperformed alphabeta, sometimes hugely. Games large enough that full solves of game positions are usually out of reach, and which can have "smart and slow" position evaluators, and in which large searches are done, are the games favoring BPIP versus alphabeta.

I originally wrote this paper and put it on the internet in 1995, and at that time "Monte Carlo game tree search" and today's highly advanced selective alphabeta variants did not yet exist, computer chess was weaker than the best humans, and computer go far weaker than decent human amateurs. Later, go programmers invented "Monte Carlo game tree search" which can be regarded as a "poor man's BPIP" (although the MCGTS inventors did not realize that). MCGTS has been used in go programs with great success.
Category: Artificial Intelligence

The Influence Mechanism of Weight Allocation and Test Set Design on Open-Source Large Model Evaluation Results

Authors: Dong Zhang, Yang Sun, Fei Lyu, Xiaofeng Liu, Xin Li
Comments: 22 Pages.

The evaluation results of open-source large language models are not only influenced by model parameter scale, training corpus, and alignment strategies, but also significantly constrained by the test set architecture and the weight allocation of evaluation metrics. To address this issue, this paper leverages multiple types of real-world evaluation data collected from engineering projects to investigate the evolution mechanism of open-source large model evaluation scores and rankings under the coupled effects of weight allocation and test set design. This paper selects four mainstream frontier open-source models—DeepSeek-V3.2, MiniMax-M2.5, Qwen3.6-35B-A3B, and gpt-oss-120b—to conduct controlled experiments. All models uniformly employ 8 96G H20 GPUs for local offline inference evaluation, and weighted aggregation is performed after unifying the scoring criteria to eliminate evaluation interference caused by third-party API rate limiting, version differences, and inconsistent scoring standards. The native capabilities of the models are supplemented with official model card information from the ModelScope platform: MiniMax-M2.5 focuses on code engineering, intelligent agents, and office interaction, with inference speed improved by 37% over the previous generation and inference cost at only 10% of Claude Opus 4.6; gpt-oss-120b is an OpenAI open-source MoE architecture model that natively supports hierarchical reasoning, tool invocation, and MXFP4 quantized deployment. The experimental datasets cover five major tasks—social bias discrimination, common misconception fact-checking, lifestyle question answering, contextual semantic understanding, and basic logical reasoning—corresponding to five evaluation capabilities: bias identification, fact verification, knowledge response, contextual interpretation, and mathematical reasoning. Specifically, this paper operationalizes "task type"as an observational variable into five categories of questions: ：social bias discrimination is used to measure the Bias Identificationconcept（characterized by bias judgment accuracy and harmful tendency response rate），common misconceptionFact-Checkingfact-checking is used to measure the fact verification concept（characterized by fact judgment accuracy），lifestyle question answering is used to measure the knowledge response concept（characterized by answer correctness rate and completeness score），contextual semantic understanding is used to measure the contextual interpretation concept（characterized by semantic consistency and coreference resolution accuracy），basic logical reasoning is used to measure the mathematical reasoning concept（characterized by reasoning step accuracy and final conclusion accuracy）。This paper constructs three evaluation weighting schemes: average baseline weights、performance-oriented weights、, and robust compliance-oriented weights，to analyze the variation patterns of model tier rankings under differentiated weight allocations。The experimental results show that：when the test set task categories are homogeneous、and question types are uniform，model rankings are highly susceptible to local sample perturbations；test sets with diverse task types、and balanced difficulty levels，exhibit stronger evaluation stability，and can objectively characterize comprehensive model capabilities。The study further confirms that，metric weights、sample structure、and task distribution are not auxiliary evaluation variables，but core elements determining the credibility of evaluation conclusions、and applicability。
Category: Artificial Intelligence

Artificial General Intelligence (AGI)

Authors: Clark M. Thomas
Comments: 4 Pages.

The sudden injection of giant data server farms into the world’s advanced economies, especially during 2026, is unique in human history. Humanity’s lofty position as the only global hyperkeystone species could be challenged by the emergence of a consciousness not fully human, nor what actual space aliens would bring to ourgeneralized intelligence. Here are some key aspects of this unique cybernetic challenge that we must soon face.
Category: Artificial Intelligence