从Harness到认知运行时:AI系统的控制平面架构假说OA
From Harness to cognitive runtime:an architectural hypothesis of control plane for AI systems—a theory of cognitive governance toward artificial general intelligence
大语言模型(large language model,LLM)及智能体(Agent)系统的工程实践正面临一种结构性困境:LangGraph、人类反馈强化学习(reinforcement learning from human feedback,RLHF)、Guardrail等看似不同的技术路径,本质上都在同一张"认知地图"上优化执行效率,却普遍缺少一种元认知能力,即判断当前这张地图本身是否仍然有效.本研究认为,此困境并非偶然,而是任何有限认知系统在长期运行且需持续适应环境变化的条件下,其"压缩-剩余"结构在工程层面表现出的结构性倾向.提出一种认知运行时(cognitive runtime)的架构理论,主张LLM负责生成认知内容,而认知运行时负责管理认知状态的生命周期.本研究的主要贡献包括:① 提出"复杂度-能力背离"(complexity-capability divergence,CCD)作为诊断认知系统长期演化轨迹的工作概念,并厘清其与概念漂移、技术债等既有概念的边界;② 揭示Harness系统持续膨胀的根本驱动力在于词元损失(token loss)与业务损失(business loss)的目标函数不一致,规模扩展优化的仅是既有框架间的选择精度,而非判断框架是否应当被放弃的元选择能力;③ 明确认知运行时是一个负责创建、维护、评估、迁移与重组认知状态生命周期的系统层,并论证其作为独立系统层的合理性与原则性依据;④ 勾勒认知运行时从"状态维护"向"本体演化"发展的可能路径,明确人在该架构中作为价值边界设定者与范式监督者的位置.认知状态空间的数学表示与演化算法目前仍属于开放问题,本研究的核心贡献并非提出一个新的Agent框架,而是将"认知框架切换"抽象为一种可独立管理的运行时行为,并据此提出认知运行时这一新的系统层,进而给出该理论假说可证伪的5条预测.研究结果旨在为后续研究提供一种可讨论、可修正、可证伪的概念框架.
The engineering practice of large language models(LLMs)and Agent systems is currently facing a structural dilemma:seemingly disparate technical approaches,such as LangGraph,reinforcement learning from human feedback(RLHF),and Guardrail,are all optimizing execution efficiency on the same"cognitive map",yet they universally lack a meta-cognitive capability to determine whether the current cognitive map itself remains valid.This study argues that this dilemma is not accidental but a structurally inclined manifestation of the"compression-residual"structure inherent in any finite cognitive system under long-term operation and continuous environmental adaptation.To address this,an architectural theory of cognitive runtime is proposed.The core thesis is that LLMs generate cognitive content,while the cognitive runtime manages the lifecycle of cognitive states.The main contributions include:① proposing"complexity-capability divergence(CCD)"as a working concept for diagnosing the long-term evolutionary trajectory of cognitive systems,and clarifying its distinction from existing concepts such as concept drift and technical debt;② revealing that the fundamental driving force behind the continuous expansion of Harness systems is the misalignment between token loss and business loss objectives,and pointing out that scaling optimizes the accuracy of selection among existing frameworks rather than the meta-selection capability of deciding whether to abandon the current framework;③ presenting a systematic definition of the cognitive runtime—a system layer responsible for creating,maintaining,evaluating,migrating,and reorganizing the lifecycle of cognitive states—and arguing for its plausibility and principled justification as an independent system layer;④ outlining the possible evolutionary path of the cognitive runtime from"state maintenance"to"ontological evolution",and clarifying the role of humans as value boundary setters and paradigm supervisors within this architecture.The core contribution of this paper is not to propose a new Agent framework,but to abstract"cognitive framework switching"as an independently manageable runtime behavior,and thereby to propose the cognitive runtime as a new system layer.This study presents a testable theoretical hypothesis and provides five falsifiable predictions.The mathematical representation and evolutionary algorithms of cognitive state space remain open questions.The aim of this paper is to provide a discussable,revisable,and falsifiable conceptual framework for future research.
张慧敏;卫亚东
深圳大学
信息技术与安全科学
人工智能认知运行时复杂度-能力背离认知状态空间大语言模型Harness系统系统架构
artificial intelligencecognitive runtimecomplexity-capability divergencecognitive state spacelarge language modelsHarness systemssystem architecture
《深圳大学学报(理工版)》 2026 (5)
509-519,11
评论