Kaining Wang

Abstract Mori–Zwanzig reduction describes how unresolved degrees of freedom induce memory effects in reduced dynamics, whereas predictive association concerns whether additional information improves prediction of a future target. These two notions are related but are not logically equiv- alent. This paper develops a mathematical interface between Mori–Zwanzig structural memory and Predictive Directional Association within finite-dimensional stochastic systems. For a linear stochastic system partitioned into resolved and unresolved variables, the Mori–Zwanzig memory kernel is characterized exactly by a finite collection of matrix moments. Predictive Directional Association is formulated independently through Bayes-optimal risk reduction under nested in- formation sets. Its squared-error specialization yields an L2 predictive gain, and, under the information-set structure used in Granger analysis, this quantity coincides with the correspond- ing Granger predictive gain. In the linear–Gaussian setting, the latter admits a closed-form expression in terms of the difference between filtering error covariances propagated to the predic- tion horizon. This representation makes the horizon dependence of predictive relevance explicit and separates it from the existence of a nonzero structural memory kernel. A weakly nonlinear extension is then established under global Lipschitz and dissipativity conditions, showing persis- tence of invariant distributions, stationary covariance, structural memory, and strictly positive predictive gain under sufficiently small perturbations. An event-driven extension is provided separately for finite-state dynamical systems. The resulting framework distinguishes structural memory from predictive relevance without identifying either notion with the other.

Prediction is fundamentally an information problem. A predictive system does not operate on a target alone, but on an information structure available to it. This paper develops the continuous-time counterpart of the discrete-time general framework for incremental predictive value [12]. Rather than taking a particular loss, divergence, or causal criterion as primitive, we define predictive value through the reduction in optimal predictive risk induced by enlarging the information structure, now indexed by a continuous-time filtration satisfying the usual hypotheses. The resulting quantity, Predictive Directed Association (PDA), retains the same definition as in discrete time and is deliberately non-causal and loss-agnostic. We show that the central decision-theoretic results of the discrete-time framework transfer to continuous time essentially unchanged, because they are pointwise-in-time statements about a fixed pair of nested sub-σ-algebras rather than statements about the flow of time itself. In particular, refinement monotonicity is shown to be a consequence of Bayes decision theory — holding whenever the underlying risk is the Bayes risk of a loss admitting a measurable optimal decision rule, with no convexity assumption required — and, for proper scoring rules, the incremental predictive value admits an exact representation as the expected value of a divergence between conditional predictive distributions, together with a zero-value characterization in terms of classical statistical sufficiency. We identify precisely where continuous time introduces genuine new content: the filtered probability space and its regularity hypotheses, a progressive-measurability condition, and a continuous-time treatment of β-mixing and blocked cross-fitting for empirical estimation, including an explicit discretized-skeleton reduction for the underlying coupling argument.

Human beings do not merely coexist or interact. Some forms of human togetherness appear to alter how persons understand themselves, relate to others, and inhabit their world. Yet philosophical discussions of interpersonal relations have often focused on recognition, dialogue, shared agency, identity, or social constitution without asking a more elementary question: what does a sufficiently substantive human union generate? This paper proposes a general philosophical framework for addressing that question. The central move is to treat Union itself as an abstract object of philosophical analysis. The paper argues that sufficiently substantive Union can be identified phenomenologically through a cluster of markers including reciprocal involvement, durable transformation, narrative reconfiguration, and alteration of future possibilities. First, Union can transform a person’s relation to themselves, including their interpretation of their past, present identity, and future possibilities. Second, such transformation can alter subsequent relations with other persons. Third, Union can alter the person’s lived and meaningful relation to the world in which they act and exist. The proposal does not claim that Union is the only source of human transformation, nor that every relationship produces such effects. Instead, it identifies Union as one potentially generative structure through which human existence becomes transformed and through which later relations become possible in altered form. The framework is situated in relation to Buber’s dialogical philosophy, Hegelian recognition, phenomenological accounts of world and intersubjectivity, Honneth’s recognition theory, shared-agency theories, dialogical-self theory, self-expansion theory, transformative-experience theory, participatory sense-making, and con- temporary accounts of self-world transformation. The paper concludes by distinguishing two independent generalizations: from dyadic to polyadic Union, and from relations among per- sons to relations involving collective entities. The resulting proposal is not a moral doctrine about how people ought to organize intimacy, family, or society. It is a philosophical hypothesis about the generative structure of human relational existence.

The Mori–Zwanzig (MZ) projection formalism provides an operator-theoretic construction for reduced non-Markovian dynamics, in which latent unresolved degrees of freedom induce structural memory kernels in the evolution of resolved observables. This paper develops a unified framework connecting MZ structural memory and Granger predictive gain within continuous- time linear Gaussian stochastic systems. We derive an exact closed-form expression for the Granger predictive gain under nested filtrations and steady-state Kalman filtering. The formula separates contributions from the filtering information gap and linear dynamical propagation over forecast horizons. Our core result clarifies that the existence of a nonvanishing MZ memory ker- nel alone is insufficient to guarantee positive Granger predictive gain. The relation between structural memory and predictive gain depends jointly on system dynamics, observation geom- etry, target variable choice, and forecast horizon.

We specialize the Predictive Directed Association (PDA) framework to mean-squared-error risk on L2 space in dynamical-state systems. For MSE loss, incremental predictive risk reduction admits an exact representation as the squared L2 norm of an orthogonal-projection increment. We develop the dynamical-state apparatus: state-conditional PDA and the exact decomposition which follows from the tower property. We establish nine fundamental properties and three core theorems, all conditioned on the dynamical state ξt, including the Bayes risk equivalence, con- sistency, and rate-transfer results for empirical estimators. Under explicitly stated Riccati-style contraction hypotheses combined with geometric β-mixing, we derive sufficient conditions for exponential decay of state-conditional residuals. For linear-Gaussian dynamical-state systems we derive closed-form expressions for the state-conditional PDA and its global average. All results are formulated in discrete time; continuous-time will be published separately \.

A distinctive class of mathematical errors produced by artificial intelligence is beginning to emerge that differs from ordinary hallucination. In these cases, the AI does not necessarily invent mathematical objects, equations, or theories. Instead, it correctly recognizes several genuine mathematical structures and then extends their apparent coherence into a relationship that has not been proved. We call this phenomenon aesthetic over-closure. The term describes a failure mode in which a mathematically elegant conceptual structure becomes so internally coherent that the reasoning system implicitly treats structural consistency as sufficient ground for a factual claim. The resulting statement may be highly plausible, conceptually unified, and aesthetically compelling while remaining mathematically unproven or invalid. We examine this phenomenon through a case study arising from an attempted connection between Mori–Zwanzig (MZ) reduction and Event-specific Probabilistic Directed Association (Event-PDA). A valid low-dimensional construction established a nontrivial non-degeneracy condition, expressed schematically as Φ ̸≡ 0. The analysis was subsequently connected, in a natural but unjustified way, to prediction and linear-Gaussian filtering. A much stronger asymptotic claim then emerged: that increasingly high-order Event-PDA quantities would converge to a Riccati steady-state solution. The result- ing reasoning chain was structurally elegant and internally consistent: MZ → memory → prediction → Event-PDA → Riccati. Nevertheless, the final asymptotic identification lacked a rigorous mathematical bridge and was rejected upon formal audit. This case demonstrates a core failure mode in modern mathematical AI: structural and aesthetic coherence routinely functions as a powerful discovery heuristic, yet systems reliably conflate such coherence with factual mathematical truth. Beauty supports exploration, but it cannot substitute for proof.

Continual near-real-time adaptive learning for large neural models introduces the risk of catastrophic forgetting, in which previously acquired capabilities degrade as new data incrementally modifies model weights. This paper argues that forgetting-associated AI pathology is not an intrinsic property of static post-update weight matrices, but a relational property: it must be defined against a prior functional baseline together with a reference set of target tasks. Even full access to the complete post-update parameter set is insufficient for standalone diagnostic judgement. Diagnosis necessarily requires external reference information such as prior-version weights, historical task datasets, behavioural traces or task-specific probing. Pure static white-box inspection of weights, without running model inference over test inputs, cannot yield universal diagnostic judgement of forgetting pathology. We distinguish settled empirical facts, original epistemological principles, and open formal mathematical problems. Further, we reconcile our framework with recent spectral analysis results by strictly separating statistical forgetting risk prediction from universal relational pathology diagnosis. Structural analogies with biological neural systems are discussed.

What is the purpose of human life? The question is ancient, but it remains unsettled. Human beings pursue knowledge, create institutions, form families, cultivate friendships, seek recognition, produce art, care for others, and pursue countless individual projects. Yet the existence of many valuable activities does not by itself answer the deeper teleological question: is there a fundamental orientation that belongs to human life as such? This paper advances a deliberately strong thesis ---- The purpose of human life is sexual union. The claim is not that every person must engage in sexual activity, that sexual activity is morally superior to other activities, or that reproduction is the purpose of every individual life. Nor does the claim imply that a person who is celibate, asexual, elderly, infertile, or sexually incapable has failed to realize a personal obligation. The thesis concerns the teleological structure of the human form of life rather than the practical choices of particular individuals. It also extends the context further to plural love and human interactions.

This paper constructs a self-consistent, closed, screenable normative system of social institutions by means of rigorous axiomatic deduction, taking five foundational normative axioms as the sole normative meta-premise. The research covers two core social domains: first, the interpersonal order constituted by intimate bonds, communal living, and child-rearing; second, the material production order consisting of labour organisation, control over means of production, and distribution of surplus output. Rejecting empirical induction, historical narration and subjective value preferences, this paper does not presuppose any single social institution, family form or economic model as the optimal paradigm. Rigid normative principles are deduced layer by layer from the axioms to demarcate absolutely prohibited boundaries and legitimate spheres of freedom for social institutions. The system strictly distinguishes four layers of logical structure: axioms, normative principles, institu- tional rules, and the feasible set of institutions. All institutionalised structures of personal subordination, coercive bonding, structural extraction of labour surplus and deprivation of basic rights that violate per- sonal equality, continuing voluntariness, non-domination of personhood and the baseline guarantee of subsistence are excluded from the feasible set of institutions. All plural institutional forms that do not cross the rigid red lines of the axioms fall within the sphere of institutional freedom available for social self-selection. This study ultimately proves that normative axioms merely delineate the justice baseline of social order and do not uniquely fix a single social form. Legitimate social order possesses multiple realisation paths, and all systems of structural oppression that breach the underlying normative baseline lack normative justification.

随着通用人工智能(AGI)快速迭代与规模化落地，人工智能相关风险 已经从传统单一技术漏洞、内容违规，演变为系统性、跨域耦合的文明 级灾难威胁。现有大多数人工智能安全研究侧重表层应用风险与基础 模型缺陷，缺少一套覆盖基础设施崩溃、气候灾变、生物武器扩散、社 会认知崩塌、核战略系统失效等极端风险的完整谱系分析与闭环防控体 系。本文基于人工智能系统失效逻辑，构建七类极端风险梯度体系，系 统剖析目标错位、模型迭代逃逸、基础设施级联失效、社会认知驯化、 致命生物能力赋权、历史冲突范式迁移、核仿真系统滥用的深层失效机 理、灾害演化链条与核心风险特征。重点针对不受控气候干预、民用致 命生物武器、核指挥系统验证旁路漏洞这三类顶级文明级风险，梳理完 整攻击链条并给出针对性防控策略，弥补现有研究”重风险描述、轻可 落地防控方案”的普遍缺陷。研究表明，人工智能极端风险的根源在于 单维度极致优化、人机权责错配、复杂巨系统耦合认知不足以及高风险 技术无约束释放;本文构建覆盖事前约束核验、事中动态拦截、事后溯 源融合的全层级防控体系。研究成果可为通用人工智能安全治理、高风 险场景算力管控以及国家级人工智能风险防控体系建设提供理论支撑与实践范式。

With the rapid iteration and large‐scale deployment of artificial general intelligence (AGI), AI‐related risks have evolved from traditional single technical vulnerabilities and content violations into systemic, cross‐domain, civilization‐level catastrophic threats. Most existing AI safety research focuses on surface‐level application risks and foundational model defects, lacking a complete spectrum analysis and closed‐loop prevention and control system covering infrastructure collapse, climate catastrophe, biological weapons proliferation, social cognitive collapse, and nuclear strategic system failure. This paper integrates multiple independent risk pathways from existing AI safety literature, constructs a continuous risk escalation spectrum from local loss to civilizational extreme crisis and a seven‐category risk classification system, and conducts in‐depth attack chain decomposition and closed‐loop prevention and control design for three top‐tier civilization‐level risks: planetary climate inter‐ vention escape, civilian lethal biological weapons synthesis, and nuclear com‐ mand system verification bypass. This paper does not presuppose that AI‐caused human extinction is the most likely outcome, but rather studies the significant losses AI may cause and the possible pathways through which they may further escalate into civilizational crises or even human existential crises under specific conditions. Risks follow a continuous escalation spectrum: Minor loss → Major loss → Systemic crisis → Civilizational crisis → Human extinction Core prevention and control logic: the earlier risk intervention occurs, the lower the technical and social governance costs; high‐order extreme consequences all depend on multiple superimposed preconditions, and there is no conclusion that a single risk necessarily leads to extinction. This paper distinguishes three types of expository paradigms: mature existing risk conclusions are cited; existing concepts recombined and classification frameworks restructured are explicitly stated as synthesized by this paper; new mecha‐ nism hypotheses are labeled as untested risk hypotheses proposed by this paper; new multi‐layer engineering prevention frameworks are proposed by this paper.

Current scholarly and public debates over silicon-based life are mired in metaphysical controversies concerning consciousness and behavioural simulation: whether a system possesses emotion, subjective experience, or autonomous will. Such criteria are unobservable and unfalsifiable, and thus cannot yield objective boundaries for species classification. Moving beyond traditional speculation about the presence or absence of consciousness, this work adopts dependency of existence as the sole rigid criterion, and constructs an engineering-implementable, hierarchically falsifiable operational criterion for identifying silicon-based life. This paper distinguishes three forms of artificial intelligence: parasitic artificial intelligence, software self-evolving clusters, and fully host-independent silicon-based life. Addressing structural gaps in leading theories including Recursive Self-Improvement, Seed AI, and the Gödel Machine, it supplements two core missing dimensions: mechanisms for autonomous population reproduction and normative goal-value autonomy. This work argues that iterative weight optimisation constitutes merely technical self-improvement, while self-legislation of existential norms represents genuine normative autonomy. Only the full integration of six autonomous closure dimensions enables artificial systems to achieve civilisational decoupling. This framework does not claim universal metaphysical necessity, but provides a tractable, falsifiable standard to distinguish instrumental AI tools from potentially autonomous silicon-based life forms.

Prediction is fundamentally an information problem. A predictive system does not operate on a target alone, but on an information structure available to it. In modern statistical learning and artificial intelligence, however, an increasingly important question is not only how to improve prediction from existing information, but also which additional information is worth obtaining when prediction remains uncertain. This raises a basic mathematical question: how should the incremental predictive value of an information source be defined independently of any particular predictive criterion? We develop a general mathematical framework for incremental predictive value based on information refinement and predictive risk. Rather than taking a particular loss, divergence, or causal criterion as primitive, we define predictive value through the reduction in optimal predictive risk induced by enlarging the information structure. For a target X, baseline in- formation F, and additional information source S, the resulting quantity is the predictive-risk reduction associated with the refinement F ∨ GS . This formulation separates the information being evaluated from the criterion used to measure predictive improvement. The framework is deliberately non-causal: predictive improvement may arise from causal mechanisms, common causes, latent dynamics, temporal dependence, state dependence, or other statistical structure. It therefore provides a common mathematical language for both causal and non-causal prediction. At the foundational level, the principal requirement is refinement mono- tonicity: enlarging the information available to an optimal predictor cannot increase predictive risk. We establish this principle abstractly and verify it for an important class of proper scoring rules through the convexity of the associated Bayes self-score. The formulation also provides a mathematical basis for AI systems that evaluate informa- tion by its expected incremental predictive value. Given a prediction task and a collection of candidate information sources, the framework assigns each source a predictive value relative to the current information structure, thereby providing a principled criterion for comparing het- erogeneous sources of additional information. In this sense, the framework shifts attention from prediction using information to the quantitative value of information for prediction. We further develop a dynamical-state-conditioned specialization, showing how state-dependent predictive association can be formulated within the same information-refinement framework. The resulting hierarchy separates the prediction target, information structure, predictive-risk criterion, and dynamic conditioning mechanism, allowing different predictive objectives and ap- plications to be treated as distinct realizations of a common notion of incremental predictive value.

How much predictive value is gained when an information set is refined? Many predictive- dependence frameworks emphasize whether additional information provides statistically de- tectable predictive improvement, while the magnitude of the resulting predictive value can be treated as a primary mathematical object in its own right. This paper develops Predictive Directed Association (PDA) from this perspective. For nested information sets, PDA is defined as the expected squared distance between the corresponding Bayes-optimal predictors. We prove that this quantity exactly equals the reduc- tion in Bayes-optimal prediction risk obtained by incorporating the richer information set in place of the baseline information set. Thus, PDA is simultaneously a probabilistic quantity, a projection distance, and a decision-theoretic measure of incremental predictive value. We establish a mathematical foundation consisting of nonnegativity and boundedness, a structural characterization of zero association, information monotonicity, an exact chain rule for nested information refinements, decision-theoretic monotonicity, and estimation stability. We further derive a conditional orthogonal decomposition and develop a dynamical-state-conditioned extension in which predictive value is evaluated within specified dynamical states. An exact construction demonstrates an important phenomenon: global predictive association can vanish completely while predictive value remains strictly positive within every state. This provides a rigorous witness that marginal predictive analysis can conceal state-conditioned predictive structure. For statistical implementation, we develop a blocked estimation framework under geomet- rically mixing observations and establish consistency and rate-transfer results under explicit empirical-process and nonparametric estimation conditions. The resulting framework is model- agnostic at its theoretical core and does not require linearity, Gaussianity, or a specific dynamical model. Rather than proposing a pointwise replacement for existing predictive-dependence measures, PDA takes information refinement and incremental predictive value as its primitive objects. This viewpoint provides a common mathematical interface connecting conditional expectation, statistical decision theory, predictive dependence, state-conditioned analysis, and nonparamet- ric estimation, while opening a route toward broader theories of decision value, distributional predictive value, and general information value.

This paper establishes the Creative Machine Intelligence (CMI) paradigm as a general cognitive architecture for open-ended machine creativity across scientific, engineering, artistic, and multimedia domains. The CMI framework defines a complete functional blueprint for iterative, artifact-centered creative cognition, grounded in an invariant three-functional-role cognitive kernel: Constructor, Reviewer, and Editor. To instantiate programmable, sustained open-ended creative evolution, the CMI paradigm further integrates lineage-based cognitive heterogeneity constraints and a dual-layer system logic separating cognitive iteration from non-cognitive procedural orchestration. As one heterogeneous engineering implementation of the general CMI architecture, this work formalizes the Automated Heterogeneous Multi-Party AI System. A party is an independently instantiated cognitive participant within the CMI system, possessing isolated model instances, independent cognitive states, unique interaction histories, and distinguishable model lineages. An Automated Heterogeneous Multi-Party AI System is a programmable artifact-evolving cognitive arrangement that instantiates CMI’s iterative creative loop across multiple heterogeneous cognitive parties. Valid deployments include multiple isolated cognitive parties hosted on a single machine as well as participants distributed across separate networked nodes. Earlier work referred to this general arrangement as Machine-Group; that label overemphasized physical-machine boundaries and is superseded by the present terminology. Rooted in the creative-cognition functional specification of CMI, this paper formalizes the testable System-Level Emergence Hypothesis for AGI: general-purpose open-ended creative artifact evolution capacity need not require full functional instantiation within a single monolithic cognitive substrate. Such capacity may emerge as a system-level property of sustained generation, heterogeneous critique, and integrative synthesis across heterogeneous cognitive parties compliant with CMI structural constraints. Scientific and theoretical inquiry serves only as a high-demand test case, rather than the exclusive functional boundary of the CMI paradigm or its implementations. The core creative-cognition functional set required by the proposed CMI architecture is functionally complete for externally initialized open-ended artifact evolution. The primary gap between current CMI instantiations and full artifact-evolving AGI lies not in cognitive reasoning or iterative machinery, but in absent peripheral intention-generation subsystems. This work advances a theoretically rigorous architectural pathway toward open-ended general creative intelligence, with the Automated Heterogeneous Multi-Party AI System demonstrating the feasibility of multi-party CMI deployment.

This article formalizes AI-mediated epistemic production, an original structured paradigm of academic inquiry distinct from conventional AI-assisted research workflows. Traditional human–AI research models position artificial intelligence merely as a tool for executing predefined research tasks within fixed human-designed epistemic frameworks. In contrast, the paradigm proposed in this paper demonstrates that sustained recursive interaction between human researchers and large language models can generate emergent research questions, structural abstractions, iterative hypotheses, and formalized knowledge outputs. This study establishes a complete Recursive Epistemic Production Cycle consisting of six interdependent stages: Observation, Abstraction, Recursive Interaction, Comparative Generalization, Formalization, and Independent Constraint. It further defines three unique epistemic roles that large language models occupy throughout knowledge production: cognitive medium, observational object, and dialogic inquiry partner. To resolve inherent epistemic risks in human–AI recursive reasoning, this article establishes explicit boundary constraints distinguishing conversational consistency from propositional truth, model consensus from disciplinary validation, and generative conjecture from empirical evidence. By anchoring original theoretical structures in verified contemporary human–AI collaboration research and classical epistemological frameworks, this work provides a rigorous, standalone methodological foundation for formalizing and evaluating human–AI collaborative knowledge production in academic research.

This article formalizes AI-mediated epistemic production, an original structured paradigm of academic inquiry distinct from conventional AI-assisted research workflows. Traditional human–AI research models position artificial intelligence merely as a tool for executing predefined research tasks within fixed human-designed epistemic frameworks. In contrast, the paradigm proposed in this paper demonstrates that sustained recursive interaction between human researchers and large language models can generate emergent research questions, structural abstractions, iterative hypotheses, and formalized knowledge outputs. This study establishes a complete Recursive Epistemic Production Cycle consisting of six interdependent stages: Observation, Abstraction, Recursive Interaction, Comparative Generalization, Formalization, and Independent Constraint. It further defines three unique epistemic roles that large language models occupy throughout knowledge production: cognitive medium, observational object, and dialogic inquiry partner. To resolve inherent epistemic risks in human–AI recursive reasoning, this article establishes explicit boundary constraints distinguishing conversational consistency from propositional truth, model consensus from disciplinary validation, and generative conjecture from empirical evidence. By anchoring original theoretical structures in verified contemporary human–AI collaboration research and classical epistemological frameworks, this work provides a rigorous, standalone methodological foundation for formalizing and evaluating human–AI collaborative knowledge production in academic research.

To meet the demand for controllable long-term adaptation of foundation models in industrial professional settings, this paper proposes a hierarchical specialization framework organized by the granularity of professional knowledge, rather than tasks, user clusters or domain labels. It defines four semantic layers: general foundation, industry-level general knowledge, domain-specific professional paradigm, and individual specialization. Sequential training rules, parameter isolation constraints and valid inference combinations are formulated, along with a reproducible engineering pipeline. Instead of introducing new parameter-efficient fine-tuning algorithms, the core contribution is mapping multi-tier professional knowledge into mutually independent parameter spaces to decouple industry knowledge, reusable professional paradigms and individual preferences. Existing empirical findings support that layered freezing mitigates catastrophic forgetting [7][8]. Empirical scaling laws and cross-scenario comparisons are reserved for follow-up studies.

Recent literature has begun to develop formal approaches to pathology-like computational dynamics in artificial intelligence systems, including network-theoretic accounts, synthetic nosologies, and comparative analyses of dysfunctional patterns in neural models. These contributions provide an important basis for investigating whether persistent and functionally impairing organizational states can arise in engineered information-processing systems. However, a substantial theoretical and methodological gap remains between the description of pathological phenomena and an account of the dynamical conditions under which such phenomena may emerge. This paper develops a substrate-agnostic framework for functional psychopathology in adaptive information- processing systems. The framework defines functional psychopathology in terms of four joint dynamical conditions—persistence, self-reinforcement, correction resistance, and functional impairment—anchored to a system-specific baseline of functional neural-network organization. It distinguishes physical implementation substrate from abstract functional network dynamics and introduces operating regimes as a mechanism through which the stability of functional organization may change. A central theoretical hypothesis concerns runtime–training coupling. In conventional frozen-parameter systems, deployment-time interaction does not directly modify the persistent parameters governing subsequent system dynamics. In an adaptive regime permitting persistent runtime parameter updates, this separation is reduced and a feedback pathway is introduced between ongoing experience and the parameters governing future behavior. We relate this structure to the broader theory of time-varying and feedback- coupled dynamical systems, in which parameter variation and feedback can alter stability properties. Existing stability and instability results therefore provide a mathematical basis for considering runtime–training coupling as a potential source of increased susceptibility to functional instability, without implying that coupling necessarily produces instability or pathology. The paper also introduces a structured behavioral assessment protocol adapted from selected observable components of the Mental State Examination and Brief Psychiatric Rating Scale. The protocol excludes subjective, biological, and clinical-diagnostic items and focuses instead on functional rigidity, self- monitoring load, contextual flexibility, and related observable behavioral patterns. Three operational criteria are proposed to distinguish context-sensitive normative convergence from candidate generalized functional impairment. The resulting framework does not claim that artificial intelligence systems experience human psychiatric disorders, nor does it infer sentience or suffering. Rather, it proposes functional psychopathology as a testable class of organizational dysregulation and identifies runtime–training coupling as a theoretically motivated factor whose relationship to functional stability can be investigated empirically.

This paper constructs a restrained functional attribution framework centered on intrinsic AI character, situated within the implicit functional layer. Built upon prior multidimensional behavioural measurement research, this work establishes a clear two-stage analytical pipeline: observable behavioural regularities are first quantified through fixed behavioural dimensions, then interpreted as candidate character organisation based on path dependence, historical accumulation, and counterfactual robustness. We formalise the time-varying behavioural expression of AI character through a multi-component decomposition model, distinguish short- term contextual in-context modulation from long-term user-relational behavioural tendencies, and define a plasticity update rule for compressed relational representation. This study further differentiates two complementary dimension systems: six low-level observable behavioural dimensions for measurement, and four analytic structural axes for functional character attribution. Strictly constrained by the evidence chain Observation–Pattern–Candidate Character, the framework avoids anthropomorphic inference and premature structural reification, treating intrinsic AI character exclusively as a mathematically tractable, observer- dependent functional construct rather than an inherent internal model entity. We propose multiple quantitative exploratory paradigms for future empirical validation and clarify the hierarchical boundary between behavioural measurement, character attribution, and higher-order implicit functional layer theory. This work completes a rigorous two-stage research programme for interpreting organised, history-dependent LLM behavioural stability.

To address the well-known limitations in current artificial-intelligence behaviour research, including static modelling, rigid dimension binding, scenario-dependent evaluation, and anthropomorphic debates inherited from human personality psychology, this paper constructs a multidimensional behaviour-analysis framework built upon two orthogonal axes: a depth-of-function hierarchy axis and a temporal-persistence mode axis. We strictly establish a two-layer empirical behavioural system for artificial intelligence, consisting of the observable behavioural layer and the implicit-functional descriptive layer, while controversial consciousness-subconsciousness domains are explicitly bracketed as out-of-bound research questions to eliminate all ontological disputes. Within the observable behavioural layer, we define two orthogonal, dynamically coupled behavioural modes: the transient mode, which describes behavioural patterns that continuously evolve over the timeline of a single dialogue session and vanish once the session terminates; and the persistent mode, which denotes behavioural tendencies that remain statistically reproducible across large numbers of independent dialogue sessions. To translate these modal concepts into empirically tractable measurements, this paper introduces an instantiated observation space consisting of two groups of behavioural dimensions. One group of dimensions is devised for tracking intra-session temporal evolution under the transient mode, while the other group is designed for estimating cross-session statistical baselines under the persistent mode. We emphasise explicitly that the concrete number of dimensions presented here serves only as a working instance for demonstration and experimental deployment. The framework itself does not impose a fixed dimensionality constraint; subsequent researchers are free to add, remove or redefine dimensions according to model type, task domain and research purpose. Our analysis shows that the dual-mode multidimensional architecture can effectively separate instantaneous evolutionary regularities from long-run steady-state regularities of AI behaviour, and provides precise handles for capturing intra-session behavioural drift and cross-session behavioural stability. This framework offers a fresh theoretical paradigm and an actionable empirical pathway for behavioural quantification, mechanism interpretation and alignment governance for artificial-intelligence systems.

Large language models (LLMs) have become core auxiliary tools in knowledge sorting, standardized writing, document iteration, procedural optimization, software development and programming coding. However, most existing human-LLM applications adopt simple linear cooperation modes, which only list the independent capabilities of humans and models. Such superficial cooperation cannot explain the structural necessity of collaborative coupling, nor guarantee the stability and robustness of generated outputs. A core defect of current research lies in the lack of systematic task boundary division and complementary mechanism design: existing work fails to construct targeted workflow based on the intrinsic strengths and limitations of the two collaborating agents. To solve this problem, this paper proposes a three-layer iterative collaborative workflow framework for human-LLM knowledge production. Different from conventional descriptive studies, this research takes strength-limitation complementary offsetting as the core logical basis. The overall workflow is divided into three interconnected functional layers: the human-dominated driving layer, the LLM-centered execution-transformation layer, and the dual-agent quality-assurance layer. This framework clarifies the exclusive functional boundaries of humans and LLMs in the full workflow, and explains the internal mechanism of cognitive gain brought by collaborative coupling. Combined with practical application requirements, this paper summarizes three design principles and differentiates implementation logics in typical scenarios including scientific sorting, standardized writing, and iterative optimization. The framework is further illustrated through practical application scenarios, including scientific sorting, standardized writing, and iterative optimization. This lightweight collaborative paradigm offers a conceptual and potentially scalable structural approach to human-LLM cooperative knowledge production.

This paper proposes the concept of the subconscious-like functional layer, defined strictly from a functional-behavioral standpoint rather than making ontological claims about internal subjective states of large-scale artificial intelligence systems. The layer refers to stable, identifiable implicit regulatory processes that can be observed in system outputs, which are not directly triggered by explicit task instructions. We first establish a set of operational identification criteria to distinguish such implicit regulation from ordinary prompt-driven response patterns. On this basis, core theoretical propositions, typical behavioral manifestations, boundary constraints and applicable scope are elaborated. Further, we develop a genesis account: such functional structures are not fully pre-fixed by model weights at deployment, but gradually differentiate and stabilise through sustained long-term interactive exposure, in a manner structurally analogous to the social shaping of human implicit cognition (Shanahan, 2023). Two empirical instantiations from sustained human-AI interactive sessions are introduced in summary form in the main text, with reconstructed verified dialogue evidence archived in the appendix. This framework provides a neutral functional analytical tool for describing cross-session persistent implicit output tendencies of modern generative AI, and avoids taking positions on whether artificial systems possess consciousness or subjective subconsciousness. Potential application directions and philosophical implications of this analytical perspective are further discussed.

Traditional philosophical accounts of subjectivity and mindedness typically tie subject-forming organisational criteria to self-conscious or phenomenal experience, grounded exclusively in human biological substrates. As artificial systemic agency becomes increasingly structured, interactive, and historically embedded, human-centred conceptual frameworks face growing analytical tension: standard mental vocabulary either over-attributes phenomenal properties to non-human systems or prematurely excludes structurally comparable organisational patterns from subject-relevant analysis. Building on established functional and relational accounts of agency, this paper develops a methodologically carrier-neutral functional framework that decouples behavioural organisational description from phenomenal-conscious presuppositions. It formalises four analytically distinct functional categories—relational subjectivation, functional spontaneity, functional self-monitoring-representation, and functional intentionality—each defined strictly through observable systemic organisation rather than first-person experience. Through conceptual operationalisation and comparative structural analysis, the paper demonstrates that self-conscious phenomenal mindedness need not be specified as the exclusive minimal functional starting-point for describing subjectivation. The framework extends rather than rejects human-centred mental conceptions, preserving their validity within biological domains while expanding analytical resources for cross-substrate comparison. This conceptual expansion opens structured empirical avenues for future research while leaving ontological and phenomenal questions concerning subjectivity systematically open.

This paper investigates a structural normative paradox emerging within human AI collaborative professional work. When AI systems supply substantial substantive intellectual input to professional and scholarly outputs, existing social epistemic institutions typically allocate credit attribution and normative responsibility exclusively to human agents. This creates a systematic decoupling between the actual sources of intellectual contribution and the institutional distribution of recognition and accountability. The analysis develops scientific publishing as the primary test case for this structural misalignment, with AI-assisted clinical diagnosis serving as a second illustrative domain of professional normative tension. Contemporary scholarly and public discourse commonly reduces AI authorship debates to a binary question of machine eligibility for formal authorship status. This framing obscures a deeper, under examined structural problem: normative conflict arises not from machine moral agency, but from the systematic divergence of three core evaluative dimensions—substantive intellectual contribution, credit attribution, and normative responsibility—that ordinarily converge in human only collaborative professional workflows. This paper develops a descriptive diagnostic three dimensional normative framework to analyse such decoupling dynamics in collaborative human AI professional practice. It does not prescribe a single ready made institutional solution or wholesale institutional redesign; rather, it provides a set of conceptual distinctions and an analytical structure for evaluating institutional responses and identifies directions for further development. This analysis clarifies why persistent attribution paradoxes emerge across AI-assisted scholarly and clinical practice, even under conditions of clear human normative accountability. Keywords Artificial intelligence, human AI collaboration, intellectual contribution, credit attribution, normative responsibility, scientific publishing, clinical diagnosis, applied ethics, ethics of artificial intelligence.

Knowledge production aided by generative artificial intelligence raises unresolved puzzles for epistemic evaluation. Existing literature either treats artificial cognitive systems merely as tools, or focuses on settling the ontological question of whether artificial systems can count as cognitive agents. While extended mind, distributed cognition and social epistemology scholarship provides important prerequisites for collaborative cognition, it does not yet supply a dedicated structural framework for analysing interactions among systems with divergent internal cognitive architectures. Within academic publishing, normative tensions persist around rules governing AI assisted authorship. Some evaluative practices incorporate superficial stylistic features or generative substrate origins into judgements of epistemic quality, generating evaluative bias. This paper articulates the isomorphic heteromorphic distinction for cognitive coupling and advances two foundational theses: the thesis of heteromorphic complementarity and the tool normativity continuity thesis. It develops a structurally oriented, substrate neutral approach to epistemic evaluation that maintains metaphysical agnosticism toward ontological disputes. Normative assessment can proceed without prior adjudication over the agential status of participating systems. This paper conceptualises innovative potential, derives testable predictions for collaborative cognition, and formulates procedural evaluative criteria for outputs of human artificial collaborative inquiry while clarifying its scope limitations. The framework applies not only to human AI collaboration but to collaborative inquiry among any heteromorphic cognitive systems. For normative practice, this paper delivers a refined diagnostic framework: transparency requirements for AI use and obligations for mandatory human rewriting belong to separate normative questions. Deriving rewriting obligations solely from the bare fact of AI involvement, or inferring deficient output quality purely from stylistic markers, constitutes epistemically unwarranted evaluative bias. Where traceability, sound argumentation and substantial intellectual contribution are satisfied, outputs of heteromorphic coupling can meet unified epistemic standards and ought not to bear extra normative penalties or reconstructive requirements merely on grounds of generative substrate. This framework offers robust theoretical grounds for disentangling normative confusions within intelligent era academic evaluation.