学术论文
从硬件认知架构到 AGI 身份理论，再到实证验证——三篇论文构成一条完整的研究线索。
Rethinking AI Hardware: A Three-Layer Cognitive Architecture for Autonomous Agents
Abstract
The next generation of autonomous AI systems will not be constrained by model capability alone, but by how intelligence is structured and distributed across heterogeneous hardware. Existing paradigms—cloud-centric AI, on-device inference, and edge-cloud pipelines—treat planning, reasoning, and execution as a monolithic computational concern, yielding avoidable inefficiencies in latency, energy, and behavioral continuity.
We propose the Tri-Spirit Architecture, a three-layer cognitive framework that explicitly separates planning (Super Layer), reasoning (Agent Layer), and execution (Reflex Layer) across heterogeneous compute substrates, coordinated via an asynchronous message bus. We introduce a formal system model, a parameterised routing policy, a habit-compilation mechanism that promotes repeated reasoning paths into zero-inference execution policies, a memory model with convergence semantics, and safety constraints.
We evaluate the architecture through a reproducible simulation study (N = 2,000 synthetic tasks) comparing Tri-Spirit against cloud-centric and edge-only baselines. Against a cloud-centric baseline, Tri-Spirit reduces mean task latency by 75.6% and energy consumption by 71.1%, with 30% fewer LLM invocations and 77.6% offline task completability.
中文摘要
下一代自主 AI 系统的瓶颈不在模型能力，而在智能如何被结构化地分配到异构硬件上。本文提出三灵架构（Tri-Spirit Architecture），将认知功能显式分离为规划层（Super Layer）、推理层（Agent Layer）和执行层（Reflex Layer），通过异步消息总线协调。该架构包含形式化系统模型、参数化路由策略、习惯编译机制和带收敛语义的记忆模型。
模拟实验（N=2000）表明，相比云端基线，三灵架构平均任务延迟降低 75.6%，能耗降低 71.1%，LLM 调用次数减少 30%，离线任务完成率达 77.6%。
- 首次将认知功能（规划/推理/执行）与硬件层显式解耦
- 习惯编译机制：重复推理路径 → 零推理执行策略
- 2000 任务仿真验证跨延迟、能耗、离线连续性四项指标
From Task-Solving Agents to Identity-Forming Intelligence: A Unified Framework for Experience-Driven AGI
Abstract
Recent advances in artificial intelligence have produced systems of remarkable breadth in task performance. Yet these systems remain fundamentally organized around externally specified objectives. They optimize tasks, but they do not possess a persistent internal structure that unifies behaviour across time. This paper argues that such a structure is necessary for Artificial General Intelligence (AGI).
We propose a theory-first framework in which AGI is defined by the capacity to form and evolve identity through experience. Five interrelated constructs organize the framework: experience as identity-relevant difference; identity as a dynamic constraint field over action distributions; WILL as developmental pressure induced by distance to the identity attractor; identity attractors as emergent stable configurations; and self-generated non-self as the capacity for open-ended development.
The paper makes three claims: the central divide between contemporary AI and AGI is the emergence of an identity structure; systems organized purely by external prompts can display competence without self-constraint; and identity formation supplies a unifying developmental principle. The framework yields falsifiable predictions and a viable research agenda, recasting AGI as a developmental problem.
中文摘要
当代 AI 系统在任务表现上已极为强大，但仍围绕外部目标组织，缺乏跨时间统一行为的内在结构。本文提出一个理论优先的框架，将 AGI 定义为通过经验形成和演化身份的能力。框架由五个核心概念组织：经验=身份相关的差异；身份=行动分布上的动态约束场；WILL=身份状态到最近吸引子的距离产生的演化压力；身份吸引子=身份动力学的涌现稳定构型；自我生成的非自我=开放发展的必要条件。
本文核心主张：当代 AI 与 AGI 的分水岭不是能力扩展，而是身份结构的涌现；单纯由外部 prompt/reward 组织的系统可以有能力但缺乏强意义上的自我约束；身份形成为行动、学习、连续性和开放成长提供了统一的发展原则。
- 将 AGI 重新定义为"身份形成智能"而非"能力扩展"
- 形式化五元框架：经验 · 身份 · WILL · 吸引子 · 自生非我
- 界定身份缺失/僵化/不稳定三种失败机制，给出可证伪预测
Longitudinal Identity Dynamics in Tri-Spirit LLM Agents: A Cross-Model Ablation Study
Abstract
Large language model (LLM) agents are increasingly evaluated as interactive systems, yet most evaluations remain short-horizon and task-centered. This paper studies a different question: whether an agent architecture can support measurable longitudinal identity dynamics across repeated interaction.
Building on the Tri-Spirit Architecture and the Identity-Driven AGI Framework, we report Study3 v4, a 100-round controlled ablation study replicated across ChatGPT-5.5 and DeepSeek V4 Pro under four conditions: full Tri-Spirit, identity-only, static persona, and Tri-Spirit without check-loop feedback. Across 800 canonical events, the full Tri-Spirit condition produced the strongest final trajectory distance and development score in both models. Strict breakthrough candidates appeared only in the full Tri-Spirit condition: 84 rounds in ChatGPT-5.5 and 63 rounds in DeepSeek V4 Pro.
These findings provide evidence that architectural decomposition and feedback structure can shape protocol-defined longitudinal self-representation trajectories of LLM agents.
中文摘要
LLM Agent 的评估多停留在短期任务层面。本文回答一个不同的问题：Agent 架构是否能支撑可测量的纵向身份动力学？基于三灵架构和身份驱动 AGI 框架，我们完成了 Study3 v4——一项 100 轮对照消融实验，在 ChatGPT-5.5 和 DeepSeek V4 Pro 上跨模型复现，设置四种条件：完整三灵、仅身份、静态人格、无检查回路三灵。
共 800 个标准事件的结果显示：完整三灵条件在两个模型上均产生最强的最终轨迹距离和发展得分；严格突破候选仅出现在完整三灵条件（ChatGPT-5.5: 84轮, DeepSeek V4 Pro: 63轮）。结果表明，架构分解和反馈结构可塑造 LLM Agent 的纵向自我表征轨迹。
- 首次用 100 轮对照消融实验测量 Agent 身份动力学
- 跨 ChatGPT-5.5 / DeepSeek V4 Pro 双模型复现
- 六维身份向量 + 轨迹距离 + 突破检测器，全面量化