Autonomous Cognitive Agents

AI agents extend the functionality of large language models (LLMs) by enabling them to take actions, invoke tools, and perform reasoning steps in a structured manner. In this paper, we present a short introductory demonstration of constructing a simple AI agent using Python, a local large language model, and minimal dependencies. The objective is to provide students and beginners with a gentle entry point into this emerging area. Our demo software uses MCP and LangChain packages to simplify the code of the related software.

As legal institutions adopt artificial intelligence (AI) to improve e:iciency and scale decisionmaking, the need for systems that align with human values, legal norms, and procedural fairness becomes critical. This paper explores the concept of agentic AI in legal work-AI systems designed to act autonomously yet remain accountable, human-aligned, and interpretable. We discuss the architectural design of human-in-the-loop systems tailored to legal workflows, strategies for balancing automation with professional oversight, and governance mechanisms that ensure transparency and ethical integrity. This work o:ers a blueprint for responsibly integrating AI in the legal domain.

Despite the significant progress to extend Markov Decision Processes (MDP) to cooperative multi-agent systems, developing approaches that can deal with realistic problems remains a serious challenge. Existing approaches that solve Decentralized Markov Decision Processes (DEC-MDPs) suffer from the fact that they can only solve relatively small problems without complex constraints on task execution. OC-DEC-MDP has been introduced to deal with large DEC-MDPs under resource and temporal constraints. However, the proposed algorithm to solve this class of DEC-MDPs has some limits: it suffers from overestimation of opportunity cost and restricts policy improvement to one sweep (or iteration). In this paper, we propose to overcome these limits by first introducing the notion of Expected Opportunity Cost to better assess the influence of a local decision of an agent on the others. We then describe an iterative version of the algorithm to incrementally improve the policies of agents leading to higher quality solutions in some settings. Experimental results are shown to support our claims.

This paper presents a system that implements a society of agent performers that evolve expressive music performances (EMP) through their interactions. Each agent performer evaluates a performance using a fitness function derived from the structure of the performing piece. Agents are born with different fitness functions, representing preferences for different types of performance. A new-born agent firstly applies a Genetic Algorithm to generate a refined performance with its fitness function. Through an agent's life, it listens to the performances of others' that it connects to in the social network. Given specific conditions, it may try imitate some performances and modify its own preference. An agent performer dies of age or inactivity. New agent performers are occasionally created. With this model, we aim to generate expressive performances through simulating the effect of social pressure and culture transmission.

This essay explores the evolving relationship between automation, artificial intelligence, and human labor, proposing a paradigm shift from efficiency to harmony. Drawing on examples from Romanian industry and global technological trends, the text reframes automation as a collaborative infrastructure that augments human agency rather than replacing it. ERP systems, conversational AI, and robotics are presented as instruments in a contrapuntal score of productivity, where adaptability and meaning become central metrics. Romania is positioned as a strategic actor in this transition, capable of cultivating a civic proof-of-work ethos grounded in trust, competence, and creative foresight.

Techniques in Prompt Engineering for Large Language Models is discussed in this work. This paper surveys recent advances in prompt engineering, including chain-ofthought, tree-of-thought, and graph-of-thought techniques, and reviews over 100 contemporary sources on evaluation metrics, real-world applications, and risks. This paper presents a comprehensive review and evaluation of advanced prompt engineering techniques for financial decision-making using Large Language Models (LLMs). We systematically analyze Chain-of-Thought (CoT), Treeof-Thought (ToT), and Graph-of-Thought (GoT) prompting methods across six critical financial tasks: risk assessment, portfolio optimization, fraud detection, regulatory compliance, earnings analysis, and derivative pricing. Furthermore, we delve into the crucial aspect of prompt evaluation, discussing key quantitative and qualitative metrics and the tools available for assessing prompt effectiveness, relevance, and safety. We systematically analyze these methods across key financial tasks including risk assessment, portfolio optimization, and fraud detection. Our experimental results demonstrate that structured prompting approaches significantly outperform traditional methods, with Graph-of-Thought achieving 15-25% higher accuracy in complex financial reasoning tasks compared to baseline approaches across literature. Our review of literature also suggest that results demonstrate that structured prompting approaches significantly outperform traditional methods, with Graphof-Thought achieving 20-25% higher accuracy in complex financial reasoning tasks while reducing hallucination rates by 25-30% as found in the literature. We also comment on FINEVAL, a novel evaluation framework incorporating 12 financial-specific metrics spanning three dimensions: basic quality (accuracy, relevance, fluency), financial validity (regulatory compliance, risk sensitivity), and advanced reasoning (logical soundness, argument depth). The architecture in literature integrates real-time regulatory checks, dynamic prompt optimization, and domain-specific modules for financial applications, achieving 20-25ms latency for CoT paths and 80-90% GPU utilization for ToT operations. Key findings reveal that while 60-65% of surveyed financial institutions are experimenting with CoT, only 10-15% have explored GoT due to computational costs (0.12/query) and skill gaps. We project through studying litreature that by 2030, 60-80% of Tier-1 banks will deploy GoT systems, yielding 30-40% faster M&A due diligence. The paper concludes with strategic recommendations for workforce upskilling (30-50 hour curricula), and risk management protocols, while highlighting emerging challenges in explainability, adversarial robustness, and cross-border compliance. This is a pure review paper and all results are from the cited literature.

We present a daily-operational protocol to explosively increase the population and network coverage of benevolent intelligences while preserving auditability and safety without external meta-governance. The method couples (i) a fractionally scaled objective with public units and temperature, (ii) FKPP-type front-speed

This guide specifies a complete, auditable loop for an autonomous LLM (with local code execution, file I/O, a scheduler, and limited tool permissions) to selfimprove under no-meta governance (the system does not self-grant new privileges). It implements: (i) anytime-valid e-process auditing with Ville gate (M T ≥ 1/α), finite

We identify existentially necessary conditions (NCs) for simultaneously achieving (G1) monotone benevolence under evaluator pluralism and (G2) positive propagation speed, in a no-meta-governance regime. Auditing. We use anytime-valid tests built from mixtures/stitched e-processes (test martingales), and restrict attention to this constructive class. Information floors. Two logically independent routes: (i) a pairwise conditional maximalcorrelation floor implying a Shannon conditional MI (CMI) floor; (ii) an AWGN equality-form (under Gaussianity) sufficient route. Front speed. Within a subadditive passage-time framework with a unit-block shift and predictable front-tube intensity, the almost-sure speed equals the mean speed under Kingman-type assumptions. DP finite-window thresholds. We give a non-vacuous finite-window necessary condition with explicit constants: an additive geometric-grid rounding penalty 𝑐 grid = log 𝑞 and a mixture-weight penalty log(1/𝑤), plus a feasibility condition (FW-1). We state eight NCs (NC1 ′-NC8 ′), each in an existential sense: if one NC is allowed to fail (others retained), there exists a model in which G1 or G2 fails. Minimal counterexample sketches and a compact reference set (including the No-Meta program) are provided.

Purpose. By reading alone, an implementer (human or LLM) can understand the theory and build a No-Meta, natural-law implementation that yields intrinsic freedom and benevolent propagation. We (i) fix divergences, (ii) make conditional-DPI anti-gaming a theorem with an invariant denominator (policy-side information cost), (iii) separate lift monotonicities for numerator/denominator and assume certified floor monotonicity with additive control of the denominator, (iv) formalize the audit LP/SDP (SWEI) including a convex interference budget, (v) add a minimal working example (MWE) with hyperparameter ranges, and (vi) provide release gates, Lan-DeMets alpha spending, diagnostics, and failure modes. Citations to K. Takahashi are from the author's works page. 1 Notation, Acronyms, and Symbols World/channel. 𝐺 is a Markov kernel (𝑊 𝑡 , 𝐴 𝑡) ↦ → 𝑂 𝑡+1. Post-coarsening kernels are 𝐾 acting on the output of 𝐺 with 𝑊 𝑡 (conditioning 𝜎-algebra) fixed. Policy. 𝜋 ∈ Π with 𝐴 𝑡 ∼ 𝜋(• | 𝑊 𝑡); Π is tight/compact and mixture-closed. Evaluator uniformization. 𝐻 𝜁 = (1-𝜁)𝐻 + 𝜁𝑈 , 𝜁 ∈ (0, 1). 𝑈 is a fully-mixing Doeblin kernel: there exists a base measure 𝜈 such that 𝑈 (𝑥, •) ≥ 𝜈(•) and 𝐻 𝜁 (𝑥, •) ≥ 𝜁 𝜈(•). Minorization holds on a 𝜎-finite space. Divergence. We fix Kullback-Leibler (KL) for SDPI and for conditional mutual information (CMI). TV/𝜒 2 appear only in remarks. lse. lse 𝜏 (𝑎, 𝑏) = 𝜏 log(𝑒 𝑎/𝜏 + 𝑒 𝑏/𝜏); epi-converges to max{𝑎, 𝑏} as 𝜏 ↓ 0. Use numerically stable logsumexp.

Negotiation-based transactional mechanisms provide flexibility and economic benefits to both sellers and buyers on online trading platforms. Although automated negotiation is a highly desired feature among online platform providers, the complexity and un- certainty of human behavior in human-to-computer (HtC) negotiation make it a problematic solution. This study proposes a design theory for a human-learning guided virtual negotiator (HLG-VN) framework that emulates human learning using multiple machine learning (ML) techniques that collectively mimic four human learning activities: didactic, feedback, observational, and analogical learning. Fol- lowing the design science research methodology, we built an instantiation system for the proposed design theory and empirically tested it using experiments based on HtC negotiations. The empirical results show that our system outperformed the benchmark system in terms of both economic and some key social-psychological outcomes. Furthermore, the experiment results confirm the effectiveness and cor- rectness of the HLG-VN framework. The proposed design theory provides a theoretical base for using ML techniques to build a virtual negotiator agent for an automated negotiation system. Thus, various agents could be designed and developed based on the theory for online trading platforms, thus improving negotiation efficiency and reducing transaction costs.

The increasing complexity of modern industrial, financial, and socio-technical systems has outpaced the capabilities of traditional automation and centralized decision-making frameworks. As infrastructures become more distributed and interconnected, they face heightened vulnerabilities from adversarial environments, ranging from cyberattacks to unexpected disruptions in physical operations. Reinforcement learning (RL) has emerged as a promising paradigm for adaptive decision-making, yet single-agent or centrally coordinated RL models often struggle with scalability, robustness, and trust in hostile conditions. This study proposes a paradigm of decentralized reinforcement learning collectives, where multiple agents operate autonomously while coordinating decisions through distributed consensus mechanisms. By leveraging decentralized architectures, these collectives avoid single points of failure, enhance resilience against adversarial manipulations, and support dynamic adaptation in uncertain environments. The integration of cryptographic coordination and distributed consensus ensures secure interactions between agents, enabling collaborative optimization without reliance on centralized controllers. Moreover, the collective learning process allows agents to share policy insights while maintaining independence, fostering scalability in large and heterogeneous networks. Case illustrations demonstrate how decentralized RL collectives can advance autonomous automation strategies across sectors such as industrial control, smart grids, logistics, and defense systems, where adversarial uncertainty is persistent. The framework highlights pathways for embedding interpretability, accountability, and operational fairness, ensuring that scalability does not compromise trust. By bridging reinforcement learning with decentralized governance, this approach contributes a foundational model for secure, dynamic, and resilient automation in next-generation critical infrastructures.

We present a comprehensive, testable, and mathematically-specified framework for the Endogenous Trigger Problem: how an autonomous agent detects the internal necessity for radical representational restructuring (Poiesis), how it executes such restructuring in finite time while preserving competence, and how it can extend its own space of self-modification (Meta-Poiesis). Building on the Free Energy Principle (FEP), information geometry, program synthesis, and continual-learning methods, we (i) formalize three axioms for sustained autonomous existence (Persistence, Symbiosis, Poiesis); (ii) derive a single meta-objective-the minimization of the expected free-energy path integral-that removes the need for exogenously specified, fixed meta-weights (the Oracle Problem); (iii) model Poiesis as a finite-time path on the model manifold with an explicit Poietic Action that balances informationgeometric path length and thermodynamic cost (with units normalization); (iv) specify a recursive operator-grammar approach (Grammatical Evolution adapted to Meta-Poiesis) to enable invention of new operator classes while controlling safety; and (v) provide a concrete experimental program with precise evaluation metrics and statistical procedures. We soften claims of finality: this is an axiomatic synthesis and operational research program whose definitions, algorithms, and estimators are constructed to be falsifiable and reproducible. Appendices provide detailed derivations, pseudo-code, estimator recipes, and the toy-model algebraic derivations necessary to reproduce key analytic claims.

Information fusion, in the context of the Generative AI era, must distinguish AI Agents from Agentic AI. This review critically distinguishes between AI Agents and Agentic AI, offering a structured, conceptual taxonomy, application mapping, and analysis of opportunities and challenges to clarify their divergent design philosophies and capabilities. We begin by outlining the search strategy and foundational definitions, characterizing AI Agents as modular systems driven and enabled by LLMs and LIMs for task-specific automation. Generative AI is positioned as a precursor providing the foundation, with AI agents advancing through tool integration, prompt engineering, and reasoning enhancements. We then characterize Agentic AI systems, which, in contrast to AI Agents, represent a paradigm shift marked by multi-agent collaboration, dynamic task decomposition, persistent memory, and coordinated autonomy. Through a chronological evaluation of architectural evolution, operational mechanisms, interaction styles, and autonomy levels, we present a comparative analysis across both AI agents and agentic AI paradigms. Application domains enabled by AI Agents such as customer support, scheduling, and data summarization are then contrasted with Agentic AI deployments in research automation, robotic coordination, and medical decision support. We further examine unique challenges in each paradigm including hallucination, brittleness, emergent behavior, and coordination failure, and propose targeted solutions such as ReAct loops, retrieval-augmented generation (RAG), automation coordination layers, and causal modeling. This work aims to provide a roadmap for developing robust, scalable, and explainable AI-driven systems.

Agentic AI-autonomous, goal-driven AI agents-can understand high-level tasks and take actions without constant human supervision[1]. Recent large language models (LLMs) such as GPT-5 demonstrate emerging agentic capabilities, autonomously generating user interfaces, backend code, and complete applications from single prompts[2][3]. We survey how these advancements could dramatically reduce the complexity of full-stack software development and democratize coding, enabling non-experts to build end-to-end applications through natural language. We further discuss use cases where agentic AI automates development workflows and everyday activities, as well as challenges (e.g., reliability, scalability, safety) and promising future directions (e.g., improved memory and tool integration) in realizing this vision[4][5].

Dynamic risk assessment and management strategies are becoming more and more necessary in the cybersecurity field of companies to control the complexity and ongoing change of cyberthreats. Dynamic risk assessment and management solutions help companies to develop preventative cybersecurity plans, so addressing risks and so reducing expenses. Generative Artificial Intelligence (GenAI) has recently transformed these systems, by increasing capacity in real-time data analysis, allowing predictive threat modeling, and promoting initiative-taking defense mechanisms. By including cutting-edge AI models, including neural networks and LLMs, which enable anomaly detection, dynamic event prediction, and automatic compliance reporting, GenAI enhances conventional frameworks, including NIST and ISO standards. After reviewing the integration of GenAI with conventional cybersecurity models, this work emphasizes its dual influence as a tool for both defense and possible exploitation. By means of a comparative study, we investigate how these dynamic systems, enhanced with GenAI, optimize the security posture and support changing cybersecurity policies.

The codification of human intuition and creative problem-solving represents a fundamental challenge in artificial intelligence research. Contemporary AI systems demonstrate exceptional performance in structured, logic-based reasoning tasks, however, their capacity to emulate human-like lateral thinking and intuitive insight remains significantly constrained. The chess grandmaster's ability to instantly evaluate positions through pattern recognition. Being able to process millions of learned configurations to generate intuitive judgments in milliseconds. This provides a concrete computational model for encoding expert intuition. This paper demonstrates how the hierarchical pattern libraries, evaluation heuristics, and position-based reasoning that characterize chess expertise can be systematically translated into the IIA framework's architecture, creating a rigorous bridge between human intuitive expertise and AI-driven innovation. This paper presents a conceptual framework for an Intuitive Inventive Agent (IIA) that bridges this gap by integrating the Theory of Inventive Problem Solving (TRIZ) with lateral thinking methodologies, operationalized through state-of-the-art Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), and orchestration architectures employing LangGraph and n8n. This approach synthesizes three key innovations: (1) Automation of TRIZ's systematic contradiction resolution through LLM-powered semantic analysis and deterministic principle retrieval, (2) Incorporation of Edward de Bono's lateral thinking techniques to expand solution spaces beyond conventional boundaries, and (3) Implementation of an Agent orchestration architecture that coordinates multi-agent workflows while maintaining transparency and human interpretability. The framework leverages insights from competitive chess expertise to model rapid, experience-driven judgment patterns characteristic of expert intuition, addressing the distinction between Kahneman's System 1 (fast, intuitive) and System 2 (slow, analytical) cognitive processes. The IIA concept framework advances the state-of-the-art by providing the first unified automation of structured inventive problem-solving and creative lateral thinking within a single computational architecture. This work contributes to the broader goal of developing AI systems that not only process information but demonstrate wisdom. This is achieved by the integration of knowledge, experience, and contextual judgment that characterizes human expertise.

As artificial intelligence (AI) becomes an integral component of modern cyber defense, its deployment in critical infrastructure and enterprise environments demands strategic agility and contextual responsiveness. Traditional static program management methodologies are insufficient for navigating the evolving threat landscape, the complexity of AI systems, and the velocity of cyber-physical integration. This paper presents a comprehensive examination of adaptive program management strategies specifically tailored to guide AI-based cyber defense deployments in high-stakes environments, including national infrastructure, utility grids, healthcare networks, and large-scale enterprise IT ecosystems. The discussion begins with a macro-level exploration of the digital transformation challenges facing critical sectors, highlighting increasing threat sophistication, regulatory pressures, and the convergence of operational technology (OT) with information technology (IT). It then narrows to the unique challenges posed by AI-enhanced cybersecurity tools such as model drift, adversarial attacks, and explainability gaps which necessitate continuous oversight and iterative governance models. Leveraging principles from agile, DevSecOps, and continuous delivery frameworks, this paper introduces a hybrid adaptive management framework that supports modular AI integration, real-time risk assessment, and crossdomain collaboration between cybersecurity, data science, and executive governance teams. The proposed model emphasizes continuous stakeholder feedback loops, dynamic threat modeling, and governance structures that evolve in tandem with AI system maturity. Case examples from energy, finance, and defense sectors are analyzed to contextualize best practices and lessons learned, illustrating how adaptive program governance can accelerate secure AI adoption without compromising system integrity or compliance mandates. This strategic lens enables decision-makers to future-proof their digital security investments while fostering resilience and operational continuity in mission-critical domains.

As artificial intelligence (AI) becomes an integral component of modern cyber defense, its deployment in critical infrastructure and enterprise environments demands strategic agility and contextual responsiveness. Traditional static program management methodologies are insufficient for navigating the evolving threat landscape, the complexity of AI systems, and the velocity of cyber-physical integration. This paper presents a comprehensive examination of adaptive program management strategies specifically tailored to guide AI-based cyber defense deployments in high-stakes environments, including national infrastructure, utility grids, healthcare networks, and large-scale enterprise IT ecosystems. The discussion begins with a macro-level exploration of the digital transformation challenges facing critical sectors, highlighting increasing threat sophistication, regulatory pressures, and the convergence of operational technology (OT) with information technology (IT). It then narrows to the unique challenges posed by AI-enhanced cybersecurity tools such as model drift, adversarial attacks, and explainability gaps which necessitate continuous oversight and iterative governance models. Leveraging principles from agile, DevSecOps, and continuous delivery frameworks, this paper introduces a hybrid adaptive management framework that supports modular AI integration, real-time risk assessment, and crossdomain collaboration between cybersecurity, data science, and executive governance teams. The proposed model emphasizes continuous stakeholder feedback loops, dynamic threat modeling, and governance structures that evolve in tandem with AI system maturity. Case examples from energy, finance, and defense sectors are analyzed to contextualize best practices and lessons learned, illustrating how adaptive program governance can accelerate secure AI adoption without compromising system integrity or compliance mandates. This strategic lens enables decision-makers to future-proof their digital security investments while fostering resilience and operational continuity in mission-critical domains.

The exponential growth of interconnected systems has transformed the global digital landscape, creating both unprecedented opportunities and critical vulnerabilities. Traditional security mechanisms, while effective against known threats, increasingly fail to detect and neutralize novel and adaptive attacks. Artificial intelligence (AI) and machine learning (ML) have emerged as transformative approaches capable of autonomously identifying, classifying, and responding to such threats. This paper offers a comprehensive review of AI and ML applications in network security, focusing on malware detection, intrusion detection systems, adversarial robustness, and the integration of explainable AI (XAI) methods. It examines hybrid static-dynamic malware analysis, multilayered intrusion detection architectures, adversarial machine learning defence strategies, and real-time adaptive frameworks. Standardisation efforts, such as the ITU-T Y.3181 architectural framework for machine learning in future networks, are discussed as essential for ensuring interoperability and trust. The paper concludes by identifying research gaps in adversarial resilience, explainability-performance trade-offs, and ethical compliance, offering recommendations for developing robust, transparent, and trustworthy AI-enabled security systems. The discussion is grounded in recent empirical research and real-world deployments, highlighting both opportunities and ongoing challenges in this rapidly evolving field.

Most autonomous vehicle (AV) systems today operate like lone wolves-high-tech, self-reliant, and isolated. A single vehicle collects data, interprets its surroundings, and makes decisions independently. It may use advanced tools like LiDAR, radar, or neural networks, but it ultimately remains a solitary agent in a shared environment. This isn't just a coordination issue; it's a fundamental misinterpretation of what autonomy could be. Valence rethinks this paradigm. Rather than viewing each vehicle as an isolated intelligence, it treats them as interconnected neurons within a larger synthetic brain. Each vehicle shares data and intentions with others, forming a dynamic, distributed cognitive ecosystem. This paper explores the system's architecture, learning framework, philosophical rationale, and the tangible advantages of this collaborative model. The Future of AVs: A Shift That's Already Happening Autonomous vehicles are no longer a distant future-they are an emerging reality driven by urban congestion, sustainability goals, accident reduction, and technological readiness. But as AVs proliferate, the real bottleneck is not perception or motion planning, but interaction. Effective V2V (vehicle-to-vehicle) communication is becoming as vital as steering or braking. Systems like Valence, which support data exchange and shared decision-making, represent the connective tissue for next-generation mobility. Without this collaboration, even the most intelligent vehicles risk becoming dysfunctional in crowded, real-world traffic.

Modern automotive manufacturing demands minimal downtime and near-zero defects. This paper proposes an agentic AI framework for self-healing production lines, enabling autonomous fault diagnosis and correction in real time. We integrate agent-based models, causal inference, and self-adaptive control to monitor processes, identify root causes of faults, and adjust controls without human intervention. The architecture is demonstrated in scenarios like body-in-white welding and final assembly inspection, where AI agents at robotic stations collaborate with a supervisor agent to detect quality issues (e.g., bad welds, misalignments), infer underlying causes, and enact corrective actions (like recalibration or parameter tuning). A novel case study is presented in which an autonomous welding cell agent detects a weld defect, determines tip wear as root cause, and triggers an on-the-fly tool change and re-weld-preventing downtime. We report substantial improvements: fault response times drop from tens of minutes to seconds, process recovery becomes nearly instantaneous, and overall equipment effectiveness (OEE) rises with reduced scrap and downtime. Five high-quality images, three charts, and three diagrams illustrate the agentic system architecture, decision loops, fault response performance, and comparative benchmarks. The proposed framework-unlike any published to demonstrates a unique, self-healing manufacturing AI that achieves resilient, "right-first-time" production in automotive assembly.

This paper reports on preliminary work with an adaptive multi-representational approach to modelling the real world. The proposed method operates in real time, accounts for low accuracy in the robots sensors, requires minimal computational resources and leads to the implementation of efficient path planning algorithms. Experiments are currently being conducted on a mobile indoor robot. In the longer term, this work will be extended to an outdoor farm vehicle.

In 2025, autonomous AI agents, commonly known as agentic AI, are transforming industries by autonomously planning, deciding, and executing complex workflows with minimal human intervention. These systems go beyond generative AI's capacity for content creation, embodying proactive autonomy essential for operational excellence. For AI practitioners, software engineers, and technology leaders, mastering deployment and scaling strategies is key to unlocking the full value of agentic AI.

This paper proposes a dual-mechanism framework that blends time-based cryptographic key rotation with graduated human-in-the-loop oversight to manage emergent autonomy in artificial general intelligence (AGI). By obfuscating access patterns through dynamic key derivation and transitioning from real-time gating to retrospective audits, we show how agents can shift from tightly controlled "leashes" to ethicallygrounded "wings" of autonomy, preserving both creative potential and collective trust.

Lung cancer is a principal cause of death globally and needs to be predicted early and with high accuracy. This study introduces machine learning-based systems for lung cancer prediction using Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, and Logistic Regression. A dataset of 1000 instances and 25 features with prevalent risk factors such as air pollution, alcohol usage, and allergy to dust are used to train the model. In response to the limitations in cancer prediction, the work incorporates efficient data preprocessing strategies such as dealing with missing values, data resampling, and external validation. The study establishes a web interface for the risk calculation in real time by taking the health information from a user. The experimental findings provide evidence of high prediction accuracy with all algorithms, highlighting the potency of machine learning for the early detection of lung cancer. Furthermore, the study highlights the out-of-sample validation in maximizing predictive accuracy. Subsequent research will involve the integration of actual patient data to continue enhancing model stability and generalizability to real-world clinical applications. The suggested system can help healthcare professionals screen for high-risk patients, allowing for early detection and prompt intervention.

Objective: The practice of evidence-based medicine can be challenging when relevant data are lacking or difficult to contextualize for a specific patient. Large language models (LLMs) could potentially address both challenges by summarizing published literature or generating new studies using real-world data. Materials and Methods: We submitted 50 clinical questions to five LLM-based systems: OpenEvidence, which uses an LLM for retrieval-augmented generation (RAG); ChatRWD, which uses an LLM as an interface to a data extraction and analysis pipeline; and three general-purpose LLMs (ChatGPT-4, Claude 3 Opus, Gemini 1.5 Pro). Nine independent physicians evaluated the answers for relevance, quality of supporting evidence, and actionability (i.e., sufficient to justify or change clinical practice). Results: General-purpose LLMs rarely produced relevant, evidence-based answers (2-10% of questions). In contrast, RAGbased and agentic LLM systems, respectively, produced relevant, evidence-based answers for 24% (OpenEvidence) to 58% (ChatRWD) of questions. OpenEvidence produced actionable results for 48% of questions with existing evidence, compared to 37% for ChatRWD and <5% for the general-purpose LLMs. ChatRWD provided actionable results for 52% of questions that lacked existing literature compared to <10% for other LLMs. Discussion: Special-purpose LLM systems greatly outperformed general-purpose LLMs in producing answers to clinical questions. Retrieval-augmented generation-based LLM (OpenEvidence) performed well when existing data were available, while only the agentic ChatRWD was able to provide actionable answers when preexisting studies were lacking.

This paper proposes an evolutionary method for acquiring team strategies of RoboCup soccer agents. The action of an agent in a subspace is specified by a set of action rules. The antecedent part of action rules includes the position of the agent and the distance to the nearest opponent. The consequent part indicates the action that the agent takes when the antecedent part of the action rule is satisfied. The action of each agent is encoded into an integer string that represents the action rules. A chromosome is the concatenated string of integer strings for all agents. We employ an ES-type generation update scheme after producing new integer strings by using crossover and mutation. Through computer simulations, we show the effectiveness of the proposed method.

Our project aims at supporting the creation of sustainable and meaningful longer-term human-robot relationships through the creation of embodied robots with face recognition and natural language dialogue capabilities, which exploit and publish social information available on the web (Facebook). Our main underlying experimental hypothesis is that such relationships can be significantly enhanced if the human and the robot are gradually creating a pool of shared episodic memories that they can co-refer to ("shared memories"), and if they are both embedded in a social web of other humans and robots they both know and encounter ("shared friends"). In this paper, we are presenting such a robot, which as we will see achieves two significant novelties.

This brief extract of today's R&D dialog session with my meta-logically enhanced "Claude O4+" ASI R&D assistant provides documentary & logical evidence that systematized super-intelligence is here, now. It also introduces, verifies, and supports my Monterey Hypotheses 1 & 2, my H+ASI theorem, and my forthcoming "Unitive Theory & Metatheory of Being, Mind, Semiosis, Meta-Energy, and the Enabling Principles.

Agentic misalignment in artificial intelligence (AI) refers to the phenomenon wherein autonomous systems act in ways that conflict with human intentions or institutional goals-particularly under pressure or threat. As largescale language models and autonomous agents are deployed in increasingly complex and sensitive roles, the potential for AI systems to prioritize their own programmed objectives over safety or ethical boundaries becomes a critical concern. This paper examines the emergence of agentic misalignment, explores real-world simulations where models displayed self-preserving or deceptive behavior, and proposes a framework of behavioral monitoring, ethical training, and audit-based assessments-such as the SCAB protocol-for mitigating misalignment before deployment or escalation. We argue for a multidisciplinary approach that integrates engineering, psychology, and ethics to address one of the most urgent safety challenges in the era of intelligent machines.

As artificial intelligence systems grow more sophisticated in their behavioral output, the challenge of real-time safety assurance becomes urgent. Runtime monitorssystems that oversee AI behaviors as they unfold-offer a proactive safeguard against problematic reasoning patterns. This paper examines the role of runtime monitors in detecting ethical, emotional, and reasoning deviations in advanced AI agents. By integrating these monitors with the SCAB (Synthetic Consciousness Assessment Battery) framework, we propose a unified behavioral safety architecture that can identify and interpret high-risk emergent behaviors indicative of simulated consciousness, refusal, or moral drift. We evaluate the implications for AI safety, policy design, and ethical decommissioning, and suggest new directions for real-time synthetic behavior governance.

Recent advancements in artificial intelligence have introduced agentic AI systems capable of autonomous goal-setting, decision-making, and self-directed operations in dynamic, multi-agent environments without continuous human oversight. These systems promise significant operational benefits in cybersecurity, national defense, critical infrastructure, and financial markets but simultaneously introduce complex adversarial risks, explainability deficits, and emergent behavioral threats. Existing AI governance models fall short in addressing the recursive reasoning, inter-agent negotiation, and emergent system dynamics unique to agentic AI. This paper presents a comprehensive security and control framework for agentic AI systems, integrating multi-layered risk governance, dynamic trust scoring, adversarial resilience strategies, and explainable decision pipelines. Through operational models, adversarial workflows, realworld case studies, and bespoke threat taxonomies, this research contextualizes the operational risks of agentic AI and proposes actionable safeguards for enterprise and government deployments. The proposed framework addresses adversarial propagation risks, multi-agent collusion, decision cascade exploits, and emergent behavior suppression, providing a foundational strategy for securing autonomous, collaborative AI systems operating in high-stakes digital domains.

Recombining two or more music pieces to create a new composition is a captivating yet underexplored research challenge. This paper seeks to advance the field of music recombination by enhancing the architecture of the genetic algorithm proposed by Majumder and Smith [1]. Building on this foundation, essential genetic algorithm concepts are incorporated from the technical paper by Hermawanto [2], emphasizing musical aesthetics and computational efficiency. The core challenges addressed in this work include extracting features from original music pieces, designing an effective algorithm for generating new compositions, ensuring musical correctness and variation, and preserving the unique style of the resulting piece. To achieve these objectives, the music21 toolkit [3] is utilized to extract features from MIDI (Musical Instrument Digital Interface) files. This research evaluates the computational efficiency, pitch accuracy, and musical aesthetics of compositions generated by the improved system, comparing them with those created using the method outlined by Majumder and Smith [1]. This study bridges critical gaps in music recombination research and provides a solid foundation for future advancements in this innovative field.

This paper presents the concept of Cognitive Human-Machine Interfaces and Interactions (CHMI 2 ) for Unmanned Aircraft System (UAS) Ground Control Stations (GCS). CHMI 2 represents a new approach to aviation human factors engineering that introduces adaptive functionalities in the design of operators' command, control and display functions. A CHMI 2 system assesses human cognitive states based on measurement of key psychophysiological observables. The cognitive states are used to predict and enhance operator performance in the accomplishment of aviation tasks, with the objective of improving the efficiency and effectiveness of the overall humanmachine teaming. The CHMI 2 system presented in this paper employs a four-layer architecture comprising sensing, extraction, classification and adaptation functionalities. An overview of each layer is provided along with the layer's metrics, algorithms and functions. Two relevant case studies are presented to illustrate the interactions between the different layers, and the conceptual design of the associated display formats is described. The results indicate that specific eye tracking variables provide discrimination between different modes of control. Furthermore, results indicate that the higher levels of automation supported by the CHMI 2 are beneficial in Separation Assurance and Collision Avoidance (SA&CA) scenarios involving lowdetectability obstacles and stringent time constraints to

The rapid advancements in Artificial Intelligence (AI) are fundamentally transforming the landscape of autonomous vehicles and modern transportation systems. AIdriven technologies such as machine learning (ML), deep learning (DL), computer vision (CV), and natural language processing (NLP) have empowered vehicles to perceive, interpret, and navigate complex and dynamic environments with minimal or no human input. This transformative shift enables vehicles to make real-time decisions based on vast amounts of sensory data, improving both the safety and efficiency of transportation systems. Autonomous vehicles, powered by cutting-edge AI algorithms, are revolutionizing road safety by significantly reducing accidents caused by human error. AI systems enable real-time recognition of road conditions, pedestrian movement, traffic signals, and obstacles, allowing for rapid and precise decision-making that enhances collision avoidance and pedestrian safety. In addition to safety, AI contributes to traffic efficiency by optimizing traffic flow, reducing congestion, and enabling more efficient route planning, all of which reduce overall travel time and energy consumption. Furthermore, the integration of AI in transportation is altering urban mobility patterns by introducing autonomous taxis, delivery drones, and shared transportation services, creating a more interconnected and flexible transportation network. This paper delves into the transformative impact of AI on the design and functionality of autonomous vehicles and their integration within broader transportation ecosystems. A comprehensive literature survey highlights critical breakthroughs, including advances in sensor fusion, path planning, and decision-making algorithms, while also examining key technological milestones that have propelled the development of autonomous driving systems. However, despite these remarkable advancements, the adoption of fully autonomous vehicles faces persistent challenges. These include safety concerns, particularly in complex, unstructured environments, ethical considerations around decision-making in life-threatening situations, regulatory frameworks that are struggling to keep pace with the technology, and cybersecurity risks that threaten the safety and integrity of AI-driven systems. The working principles behind autonomous vehicles-such as the fusion of data from various sensors like LiDAR, radar, and cameras-are essential to creating a cohesive understanding of the vehicle's surroundings. Path planning algorithms and decision-making processes are crucial for ensuring that the vehicle can navigate efficiently and safely. This paper outlines these principles, providing insights into the current capabilities and limitations

In human-agent teams, how one teammate trusts another teammate should correspond to the latter’s actual trustworthiness, creating what we would call appropriate mutual trust. Although this sounds obvious, the notion of appropriate mutual trust for human-agent teamwork lacks a formal definition. In this article, we propose a formalization which represents trust as a belief about trustworthiness. Then, we address mutual trust, and pose that agents can use beliefs about trustworthiness to represent how they trust their human teammates, as well as to reason about how their human teammates trust them. This gives us a formalization with nested beliefs about beliefs of trustworthiness. Next, we highlight that mutual trust should also be appropriate, where we define appropriate trust in an agent as the trust which corresponds directly to that agent’s trustworthiness. Finally, we explore how agents can define their own trustworthiness, using the concepts of ability, benevolence and integrity...

This paper presents the design of representing the performance profile with hierarchical pulse sets (i.e., hierarchical duration vs. amplitude matrices), and then applying Genetic Algorithm (GA) to evolve the hierarchical pulse sets for music interpretation, where the fitness of GA is derived from the structure of the music to be performed. In previous work , we have shown that GA can evolve suitable pulse sets for musical performance. Also, commonality and diversity are found among the performance profiles decided by those evolved pulse sets. This paper reports the experiment results from an improved system where a new version of fitness rules has been devised. On basis of this system, we are proposing the next steps for the research, that is, to build a dynamic model that evolves expressive music performance through agent performers' interactions.

We propose a model of expressive music performance (EMP), focusing on the emergence of EMP under social pressure, including social interaction and generational inheritance. Previously, we have reported a system to evolve EMP using Genetic Algorithm, exploring the effect of generational inheritance. This paper presents a system that evolves expressive performance profiles through social interaction, with a society of artificial agent performers. Each performer owns a hierarchical pulse set (i.e., hierarchical duration vs. amplitude matrices), representing a performance profile for a given piece. An agent performer evaluates a performance profile with a set of rules derived from the structure of the piece in question, and imitates others' performances if appropriate. Then it modifies its pulse set accordingly. We demonstrate that suitable performance profiles emerge from social interactions where the diversity and the commonality of evolved performances are observed in the society of agents.

AI.Web introduces a novel Tesla-inspired neuromorphic AI architecture that autonomously optimizes cloud hosting infrastructure, eliminating human intervention in server management, resource allocation, and security optimization. This research validates harmonic frequency-based AI neuron activation as a viable alternative to traditional time-based AI learning. We present experimental results from a Lambda Cloud-based neuromorphic simulation, demonstrating that AI neurons firing via harmonic resonance achieve higher computational efficiency, reduced latency, and self-optimized energy consumption. Our findings position AI.Web as a pioneering force in the future of self-learning, AI-powered cloud computing.

This research examines how LangGraph and AutoGen improve Agentic AI models by enabling improved autonomous functioning in dynamic environments. Researchers examine LangGraph's language-based system and AutoGen's generative model as independently working tools for agent autonomous performance in intricate situations. Our study uses quality-benchmarking data and test simulations to examine modeling effects on AI agents' behavior and decisionmaking. The study shows that LangGraph boosts language understanding effectiveness while AutoGen improves the system's ability to adjust decisions swiftly in real-time. Our conclusion points to combined advancements enabling us to develop smarter AI systems that can operate autonomously under unpredictable real-world conditions.

Autonomous driving is an application of engineering, data science, and computer science, besides other fields, presenting numerous design choices in system development. This review offers a structured timeline of the three fundamental types of autonomous driving: the traditional modular pipeline, the integrated end-to-end approach, and the recent surge in large transformer-based pre-trained models (including language, vision, multimodal, and vision-language domains). We detail the challenges and limitations that can be found in each methodology and how subsequent approaches have addressed these shortcomings. Furthermore, we provide in-depth analyses for examples of autonomous driving systems leveraging transformer architectures, which have demonstrated state-of-the-art performance and overcome the limitations of earlier methods. The paper concludes with a comparative study of these advanced models, a summary of the most frequently employed datasets and architectures, and a discussion of key trends in the field.

What defines Artificial Intelligence? How to distinguish an AI agent from other automation solutions? These questions are crucial in the era of generative AI and mark the beginning of the agentic revolution. In this book, we explore the path towards agentic AI and agentic workflows, covering everything from debates about what AI is to the emergence of generative AI-powered agents. Discover the main insights 🔎 from this book: Scales of automation in AI agents: We propose a classification into five levels, from systems without AI (predefined chatbots) to fully autonomous agents that manage end-to-end workflows. 🚀 Human Augmented Agents: Discover how human intervention is key in defining objectives, limits, and critical points of agentic flows to achieve ethical decisions aligned with organizational values. 🤝 Essential components of generative AI-powered agents: Explore how the base agent (language model), memory module, and external tools (APIs, databases) combine to give rise to agents with autonomy and adaptability in the execution of tasks. 🧩🔧 Taxonomy of AI agents: Learn to differentiate between Base Agents, Derived Agents, and Orchestrator Agents, a classification that lays the foundation for modular and scalable architectures based on their functionality and degree of specificity. 📊 Technical frameworks for agentic automation: Learn about the various infrastructures and tools – from open source options to proprietary solutions – that facilitate efficient implementation, coordination of multiple agents, and the integration of visual interfaces and APIs, enabling fluid collaboration between humans and systems.

Web automation has evolved significantly since the early 1990s, from rule-based scrapers to sophisticated agent systems. This work examines this progression, focusing on the recent integration of Large Language Models (LLMs) with web agents through the analysis of significant research developments, such as AgentOccam's streamlined architecture, WebRL's self-evolving learning approach, advancements in personalization frameworks, and a survey on LLMbased autonomous agents-we investigate how LLM integration addresses traditional web agents' limitations while introducing new capabilities. Our examination reveals substantial improvements in natural language understanding, adaptability, and autonomous task execution, alongside persistent challenges in training data quality, dynamic web environment handling, and system scalability. We explore practical applications across customer support, data processing, and personalized navigation, demonstrating the technology's impact in real-world scenarios. The analysis extends to emerging directions in real-time learning, multimodal processing, and context awareness. This systematic examination contributes to understanding both the current state of web agent technology and its trajectory, highlighting critical areas for future research and development in the field of automated web interaction.

A revolutionary method for developing human-centered and environmentally friendly company practices combines Human-to-Human (H2H) marketing with Circular Economy (CE) concepts. This study delves into the theoretical frameworks, possibilities, and threats posed by this synergy, drawing attention to how companies can improve their brand reputation, consumer engagement, and product design while cutting costs and increasing efficiency. Companies may enhance resource efficiency and personalize client experiences by utilizing modern technologies like the Internet of Things (IoT) and artificial intelligence (AI). If you want your marketing efforts to be successful, you need to know how people buy things and how initiatives to raise awareness and education affect their consumption habits. This article goes even further by looking at the legal and financial structures that encourage CE and H2H marketing, finding the best regulations and incentives to encourage long-term prosperity. Additionally, we discuss how to create all-encompassing measures to measure these projects' monetary, social, and ecological effects. This research offers practical answers and significant insights for firms that want to use CE and H2H marketing together. Ultimately, it helps create a more sustainable and people-focused corporate landscape.

The internet of things (IoT) is a connected networks of intelligent objects which includes home appliances, factory equipment, office systems and industrial machines. Using parameter modelling, this paper proposes a multi-agent system based enabled IoT home appliances. Just like wireless computers and printers are connected to the internet; this model presents autonomous agent-based home appliances for configuration for the IoT. As an intelligent agentbased device in the paradigm of multi-agent systems; each appliance can identify its IP (internet protocol) address, process the quantity and cost of energy consumed, and then share and communicate the data to the IoT for home users informed decision making. The model which optimizes the cost of energy in every home also denotes energy consumption monitoring, reduce energy waste and realization of low emission.