The future of human-AI integration in high-stakes environments-ranging from defense operations to emergency response-requires AI systems that function as intuitive teammates rather than isolated tools. Yet a persistent perspective gap remains: video understanding models rely on third-person, broadcast-style views, while multi-agent reinforcement learning (MARL) systems operate on egocentric inputs but often depend on centralized critics during training, reducing the need for decentralized policies to internalize teamcentric structure. As a result, agents may optimize behavior without learning policy-level representations of the tactical picture grounded in their own observations. To address this gap, we propose Egocentric Team AI, a research direction centered on learning implicit, distributed Common Operating Pictures from first-person views. Building on our prior work with the X-Ego-CS dataset and Cross-Ego Contrastive Learning (CECL), we outline how cross-egocentric representation alignment can be extended from passive video understanding to active multi-agent control. Specifically, we propose integrating cross-ego contrastive objectives into MARL within a multi-agent Doom-based environment as a testbed for decentralized, team-aware policy learning. By positioning crossegocentric alignment as an inductive bias for decentralized coordination, this work charts a path toward embodied systems capable of adaptive cooperation without explicit communication.
Agentic AI refers to autonomous systems powered by Large Language Models (LLMs) that can perceive environmental states, reason through complex decisions, and execute actions to achieve business objectives with minimal human intervention. It can assist in building a scalable framework for autonomous decision-making in business applications. This paper proposes a framework for designing and implementing autonomous agents that solve real-world business problems. The extensive pretraining of LLMs enables agents to navigate vast problem spaces more effectively, facilitating accelerated problem-solving. The methodology adopts a modular architecture that divides the complex business problems into sub-tasks, each of which can be handled by sub-agents. Furthermore, the integration of learning and adaptation mechanisms allows these systems to modify their responses to environmental feedback, thus improving performance over time. Our findings indicate that while Agentic AI outperforms manual baseline and traditional rule-based approaches in terms of operational speed and adaptability, challenges related to LLM stochastic behavior in edge cases persist, necessitating human-AI collaboration. Ultimately, this architecture enables the deployment of autonomous decision-making agents capable of navigating the ever-evolving business requirements of modern enterprises.
As machine intelligences become more sophisticated, individuals and societies will encounter myriad attributions of consciousness to machines. Some of these attributions may be correct; many will be illusory. This paper argues that computationalist functionalist theories of machine consciousness can and should seek to explain illusions of machine consciousness. A theory that accounts for illusions reduces the risk of producing false positives---theories that explain merely the appearance of consciousness rather than genuine consciousness. Further, given the consequential nature of consciousness attributions, whether veridical or illusory, such a theory helps society navigate issues of blame and recourse when mistaken attributions cause harm.
Earth-observing satellite constellations are increasingly expected to operate autonomously in dynamic, resource-constrained, and failure-prone environments. As constellation size and heterogeneity grow, centralized and static task scheduling paradigms struggle to provide the adaptability required to maintain mission continuity under disruptions, intermittent connectivity, and limited onboard storage. This paper presents a decentralized task rescheduling framework inspired by a service-industry scheduling analogy that models satellites as resource constrained agents operating under memory restrictions, spatial access limitations, and intermittent offloading opportunities. A multi-criteria redistribution algorithm prioritizes memory availability, proximity, capability, and workload to reassign tasks following agent loss. Through large-scale simulation, we demonstrate that the proposed framework preserves high-priority task throughput and fair workload distribution despite irreversible capacity loss. The resulting system provides a physically motivated abstraction for studying resilient, decentralized scheduling and establishes a bridge between intuitive heuristic policies and future learning-based autonomy for heterogeneous low Earth orbit satellite constellations.
The rapid integration of generative AI into education, industry, and everyday creative practices has introduced significant efficiencies, and yet it has also raised ethical concerns around authenticity, originality, and the erosion of human creativity. Many generative AI systems prioritize speed and content production over user development. Simply delivering synthesized outputs tends to limit opportunities for users to focus on the creative process and creative thinking. Additionally, users might unintentionally plagiarize others' ideas from the synthesized output from AI systems without realizing it. In this paper, we present InspireTrace, an early-stage generative AI prototype designed to support creativity through guided inspiration rather than direct solution, intentionally leaving synthesis and ideation to the user. By explicitly attributing source materials, InspireTrace aims to reduce plagiarism risks and foster a healthier, more trustworthy creativity ecosystem. We discuss the system's design rationale and its implications from creative practice in human-AI collaboration.
Learning algorithms assume meaningful input---state, context, relational structure. A continuous stream provides none of this. Before a system can learn from time, it must translate time into state. We argue that the central obstacle to embodied continual learning is not inadequate optimization but inadequate translation: credit assignment is ill-posed until temporal history has been rendered into a representation where responsibility is locally computable. We characterize the constraints on temporal translation---causal, bounded, continuous, locally interpretable---and show they admit essentially one solution class among linear, time-invariant, finite-state summaries: exponentially decaying measurements at geometrically spaced timescales. When spaced by the golden ratio to maximize incommensurability, this decomposition provides a minimal temporal language in which the past is present and credit assignment becomes tractable. We instantiate this in the Spectral Online Machine Architecture (SOMA), demonstrating continuous adaptation without catastrophic forgetting, under fixed resources, with no replay buffer and no sequence storage, reaching 1.87 bits per byte under streaming constraints with bounded memory. The architecture satisfies a requirement that theories of consciousness increasingly emphasize: temporal integration must be intrinsic to the system's state, not externalized as retrievable data. A system with translated time is its history; it does not merely have access to it. The failure of continual learning is not a failure of learning rules. It is a failure to give them a language for time.
This work explores the implementation of an agentic AI framework tailored for the educational ecosystem, with a focus on the efficacy of adaptive response framing. While traditional school automation prioritizes task completion, our research emphasizes the psychological impact of "linguistic wrappers" on stakeholder engagement. We introduce the Golden Wrapper and Reinforcement Learner personas as superior strategies for navigating the complex interpersonal dynamics between teachers, staff, parents, and students. Through an empirical evaluation of 1,200 email interactions, we demonstrate that high affirmation "cushioning" (the Golden Wrapper) significantly outperforms direct or purely empathetic approaches in securing business outcomes and positive sentiment.
Metacognition is an important aspect of information processing in the brain, subserving judgement and making cognition robust. In the literature, there are different views on the role metacognition plays in consciousness. The role of meta-cognition has been addressed by various studies. In relation to consciousness, some authors argue that metacognition is not necessarily essential in consciousness, but rather an extra mechanism constructed on a more basic mechanism, necessary when reflecting on and reporting one's own experiences. Others hold that metacognition is an integral part of phenomenal consciousness, possibly accounting for the hard problem of consciousness eventually. We aim to clarify why no consensus has been reached on whether large language models can possess consciousness, and why diverse and competing posi-tions persist regarding the nature and plurality of conscious-ness. On the strength of the analysis, we propose metacognitive closure, a concept analogous to Colin McGinn’s cognitive closure. We discuss the possibility that difficulties in elucidating mechanisms of consciousness might be clarified by considering the nature of metacognition. Based on this view, we argue how we may be able to streamline issues in consciousness through an analysis of metacognition, in a continuous spectrum with problems in cognition in general.
Uncrewed Aerial Systems (UAS) increasingly operate as large, heterogeneous teams under partial observability. In such settings, coordination depends on how agents infer, represent, and act upon uncertainty about one another: belief divergence is a natural consequence of asymmetric sensing and latent intent, and naive communication mechanisms do not scale with team size. We argue for a reframing of interaction as communication-as-control, grounded in a small set of tightly coupled capabilities that link sensing to distributed execution under interaction constraints. Rather than prescribing a fixed pipeline, coordination in such teams relies on several interdependent capabilities. First, agents must perform implicit intent inference without communication, forming uncertainty-aware beliefs over teammate intent and progress from observable behavior. Second, such uncertainty can inform selective explicit interaction, with targeted, minimal exchanges invoked only when unresolved ambiguity becomes consequential for joint decision-making. Third, uncertainty arising from both implicit inference and explicit interaction can be leveraged to support belief-guided coordination and adaptation when alignment is partial, delayed, or unattainable. This makes selective interaction a scalability mechanism: as teams grow in size and heterogeneity, coordination remains feasible only if communication is sparse, recipient-specific, and tied to decision relevance.
Here, we explore points of convergence between the Human Consciousness Hypothesis (HCH) and Integrated World Modeling Theory (IWMT). HCH posits that consciousness is defined by three fundamental principles: Genesis (conscious functions constituting an early-stage learning algorithm), Coherence (maximization of representational consistency), and Second-Order Perception (synchronous meta-awareness of perceptual processes). IWMT serves as a unifying model that reconciles major theories of consciousness with a particular focus on Global Neuronal Workspace Theory, Integrated Information Theory, and the Free Energy Principle and Active Inference framework. Central to IWMT is the proposal that phenomenal consciousness is “what it feels like” to be the spatiotemporally and causally coherent functioning of a probabilistic generative world model for an embodied agent. Mechanistically, IWMT identifies "self-organizing harmonic modes" (SOHMs) as synchronous neural complexes implementing iterative Bayesian inference to generate consciousness as maximum a posteriori estimates of embodied sensorium states. Nested heterarchies of SOHMs are proposed as biophysical substrates for consciousness, acting as dynamic cores of integrated information that facilitate the synchronous combination of multimodal sense data into a unified field of experience to promote intelligent/adaptive (active) inference and learning. Critically, IWMT requires (body-)world models to be capable of both informing and being informed by action-perception cycles at behaviorally relevant timescales. This architecture suggests consciousness could potentially be realized in artificial systems with appropriate recurrent dynamics and sufficient degrees (and kinds) of embodied grounding.
Monte Carlo Tree Search (MCTS) faces a severe scalability bottleneck in Multi-Agent Planning (MAP) due to the combinatorial explosion of joint action spaces. In this position paper, we argue that the key to tractable planning lies in leveraging low-dimensional representational structures to guide the tree search, rather than enumerating the raw action space. Specifically, under a linear approximation of joint-action returns, we demonstrate that the node expansion problem can be effectively cast and solved as a linear contextual bandit, providing theoretical regret guarantees. Empirical results on complex benchmarks confirm that this structure-aware search significantly outperforms state-of-the-art baselines, offering a scalable path for neuro-symbolic multi-agent planning. We further discuss promising extensions, including integrating dynamic agent grouping and coalition formation mechanisms to further reduce the effective branching factor.
Creativity is shaped by cultural epistemologies, symbolic systems, and traditions of sensemaking. Yet many contemporary AI-driven creativity support tools remain rooted in Western models of cognition and representation. This paper introduces CARE, the Culturally Aligned Reflection Engine, a symbolically mediated generative system that retrieves culturally grounded symbolic frames such as Adinkra symbols and proverbs to condition creative interpretation. The system supports creative reframing and culturally meaningful interaction through proverb-based interpretation, positioning AI as a partner in sensemaking rather than a source of prescriptive guidance. This work demonstrates how activating cultural heritage within interactive systems can promote culturally grounded design and expand creative possibilities. Reflections from the design and prototyping process illustrate how culturally situated metaphors can enrich ideation, deepen reflective engagement, and offer pathways for cultural continuity in computational creativity.
Biomanufacturing of biofuels, food, and energetic materials is increasingly important for defense, security, and disaster response operations, where supply-chain resilience and field-forward production are critical. Here, we present a cloud-orchestrated, edge-executed autonomous biomanufacturing platform built around an open-source, parallel mini-bioreactor array. The system integrates in situ sensing, state and growth-parameter estimation under uncertainty, and closed-loop Bayesian optimization to autonomously plan and execute experiments. We show that the platform identifies the optimal growth temperature of Escherichia coli in a single continuous run completed in under 24 hours, and enables parallel screening of Rhodopseudomonas palustris to identify favorable carbon sources and temperatures. This work demonstrates a deployable sense-reason-act autonomy loop on a real physical system.
Generalisation-optimal learning favours weak rules that keep many futures open. Adaptation under this heuristic yields a hierarchy of selves. Previous Stack Theory results show this hierarchy is necessary for human-like consciousness. A first-order self tags intervention versus observation, enabling causal learning and underpinning subjective experience. A second-order self models the listener, letting self-report survive decoder mismatch. It is necessary for access consciousness, self-awareness and Gricean meaning. A third-order self binds the future self, making long-horizon trust rational and enabling narrative planning. Without this hierarchy, a system lacks ingredients of human-like conscious experience. Here I do not restate these arguments, but supply formal necessity proofs for these first three orders of self grounded in observable behaviour. I also validate the predicted capability gaps in three randomised Monte Carlo experiments. Hence when I say no selves, no consciousness, I mean no consciousness like we humans have. I then position Stack Theory relative to Embedded Universal Predictive Intelligence (EUPI), which recreates many earlier Stack Theory results but inherits from AIXI a reliance on description-length priors whose optimality depends on the choice of reference machine. Stack Theory's weakness principle is representation-invariant. It maximises generalisation probability without requiring a privileged encoding. I discuss relative strengths, proposing that bridging the two frameworks could combine Stack Theory's firmer theoretical foundation with EUPI's ready integration into reinforcement learning.
The goal of detecting, measuring, and engineering machine consciousness depends on making explicit key metaphysical assumptions about the nature of consciousness that may be lying dormant in the minds of AI researchers. In this position paper, we lay out a small set of arguments that call into question a particular family of these assumptions. These assumptions all arise from the dominant paradigm of analytic materialist philosophy in combination with modern cognitive science: 1. consciousness is a property that emerges from complex matter, 2. consciousness depends on architectural cognitive complexity, and 3. consciousness can be measured as a function of behavioral cognitive complexity. Alternatives to these views and their implications for future machine consciousness research are discussed.
As AI-generated content proliferates in workplace settings, objective metrics that capture dimensions relevant to constructive creativity-work that balances novelty with usefulness-are needed to support metacognitive monitoring and quality evaluation. Existing creativity instruments either measure unconstrained divergence or assess accuracy against fixed item sets, limiting their applicability to real-world text of varying complexity. We present a framework for evaluating "constructive creativity" through three orthogonal, scale-invariant metrics derived from embedding geometry: divergence (semantic spread among target concepts), alignment (semantic fit between associations and targets), and parsimony (non-redundancy of associations). Across 3,600 random configurations with targets ranging from 1 to 50 words and associations ranging from 2 to 50 words, aggregate pairwise correlations remained below |r| < 0.22 under OpenAI ada-002 embeddings and |r| < 0.23 under GloVe embeddings. Regime analysis decomposing orthogonality by task scale (m, n) revealed that two of three metric pairs-divergence-parsimony and alignment-parsimony-maintained independence uniformly across all scales and both embedding models, while divergence-alignment exhibited independence at smaller task sizes. Benchmarking against 144 Remote Associates Test items provided partial validation: alignment distinguished correct solutions from chance (mean similarity = 0.356), though divergence did not predict item difficulty or solution time as hypothesized. LLM judges (GPT-4o-mini, Claude Haiku 4.5) discriminated correct solutions from foils with 84-88% accuracy, and model-solving difficulty correlated with human difficulty ( r = 0.559). These metrics offer candidate dimensions for applying semantic associations to creative tasks, providing a foundation for future validation as cognitive feedback tools in AI-augmented workflows.
Recent studies suggest that habitual generative AI use may erode human creative capacity-paralleling documented effects of GPS on spatial cognition and search engines on memory recall. We introduce Human-Centric Agentic Governance & Observability (H-AGO), a framework that shifts AI governance from content filtering to creative capacity preservation. H-AGO implements three innovations: (1) a Creative Capacity Index (CCI) that longitudinally tracks human ideation ability independent of AI assistance using embedded measurement and minimal-burden micro-assessments, (2) dual-phase constitutional governance (CreativeSilence and Socratic Guardrails) operating at real-time and longitudinal scales, and (3) adaptive intervention policies that adjust AI assistance parameters based on measured CCI trends. Unlike existing AI safety approaches focused on output harm, H-AGO governs for human capability preservation-a core principle of Well-Being AI. We propose a hybrid edge-cloud architecture where governance runs locally while high-quality Socratic scaffolding leverages cloud LLMs. We present a planned evaluation design for creative writing tasks targeting 30-40 participants over 4 weeks.
Large Language Model (LLM)-based conversational assistants, such as ChatGPT, Gemini, and DeepSeek, have shown strong potential for enabling natural language access to information. However, their deployment in sensitive domains raises concerns related to trustworthiness, including hallucinations, limited transparency, and the risk of unsafe or inappropriate responses. Purely rule-based conversational systems, while reliable and accurate to the information source, lack the flexibility required for open-ended information retrieval. We introduce SafeGenChat, a neuro-symbolic hybrid framework for trustworthy information retrieval dialogs on sensitive topics where it is paramount to make the context and risk of the information transparent to the user. Inspired by the dual-system theory of fast and slow thinking as implemented in the recently proposed SOFAI architecture, SafeGenChat combines a generative LLM-based component (System-1) with a symbolic, rule-based component (System-2) that dynamically routes user queries between verified answers and purposeful do-not-answer responses based on an assessed risk of the dialog context. We present a case study of an HIV-focused chatbot that answers user queries related to HIV to illustrate the design and application of SafeGenChat in a safety-critical domain. Overall, this work introduces a neuro-symbolic framework for risk-aware conversational information retrieval, adapts the SOFAI dual-system architecture to dialog-based settings, and demonstrates its applicability through a safety-critical HIV decision support case study.
Frontier AI systems are now producing responses that make users, developers, and policymakers genuinely pause and ask: could these models have conscious experiences? Yet the field still lacks rigorous, hard-to-game tools that can distinguish genuine indicators from optimized artifacts or surface-level cues. We introduce the Triangulated Consciousness Assessment Stack (TCAS), a validity-centered framework that combines four evidence streams: behavioral batteries with robustness controls (B), mechanistic indicators with explicit assumptions (M), perturbation tests that probe causal sensitivity and proxy failures (P), and observer-confound controls that separate anthropomorphic attribution from evidence (O). When all streams are available, TCAS produces theoryindexed credence bands and standardized disclosure cards (TCAS Cards) rather than binary detection verdicts. We report an empirical evaluation of GPT-5.2 Pro via OpenRouter (2026-02-19 UTC) covering B and P streams only, including a pre-specified role-play negative control. M and O were not run in this black-box walkthrough, so theory-indexed credence bands are explicitly withheld under the missing-stream rule. Prompts, rubric, judge prompt, raw outputs, and provenance manifest are released at the repository commit cited in the camera-ready build.
As AI systems scale, many proposed indicators of machine consciousness track capability rather than consciousness, reducing discriminative power precisely when discrimination becomes important. This paper proposes Triadic Relational Ontology (TRO) as a constraint on test design: if machine consciousness exists, it must involve a sustained triadic organization comprising (i) a temporally coherent self pole (S), (ii) a genuinely distinct world/other pole (W), and (iii) relational mediation (R) that maintains self-other differentiation under perturbation. TRO predicts qualitative collapse modes (dyadic, solipsistic, and decoupled) that do not monotonically scale with performance. We outline an implementable intervention-based protocol for transformer language models and address two conceptual risks that commonly sink such proposals: probe circularity (via counterfactual validation and cross- method convergence requirements) and S/W/R proxy bootstrapping (via staged identification with independent causal validation). Passing TRO-aligned tests is not treated as proof of consciousness; rather, TRO is proposed as a way to generate meaningful negative evidence and constrain future benchmarks in a CIMC-aligned program of discriminative testing.
