意象优先：大形象模型的理论架构与本体论白盒

中文摘要

当前以大语言模型（LLM）为核心的人工智能系统在创造性活动与物理世界理解方面存在结构性局限。本文诊断其根源为认知基元选择错误——以离散语言符号为基本单位，违背人类以多感官意象为思维载体的认知本质。提出"意象优先"假说：多感官意象而非语言符号是思维的底层运算载体，据此构建大形象模型（LIM）四层架构，以可运算矢量图（Vectok）为认知基元，给出意象物理引擎、体感锚点、拓扑嫁接创造的形式化定义，论证LIM的"本体论白盒"属性。在验证层面，实现首个可运行原型：推箱子任务上LIM体感引导成功率与A*最优解一致且无冗余步数，显著优于RL与LLM式规划；拓扑嫁接创造经多轮迭代收敛，生成新Vectok。结果表明，结构化矢量认知基元在物理推理与规划可靠性上具备可证伪的增量价值

Abstract

Current artificial intelligence systems centered on large language models (LLMs) exhibit structural limitations in creative activities and in understanding the physical world. This paper diagnoses the root cause as an error in the choice of cognitive primitives — taking discrete linguistic symbols as the basic unit violates the cognitive nature of human beings, who use multi-sensory imagery as the vehicle of thought. It proposes the "imagery-first" hypothesis: multi-sensory imagery, rather than linguistic symbols, is the underlying computational vehicle of thought. On this basis, it constructs the four-layer architecture of the Large Imagination Model (LIM), takes the computable vector graph (Vectok) as the cognitive primitive, gives formal definitions of an imagery physics engine, somatic anchors, and topological grafting creation, and argues for LIM's "ontological white-box" property. On the validation side, it implements the first runnable prototype: on the Sokoban task, LIM's somatic-guidance success rate is consistent with A* optimal solution with no redundant steps, significantly outperforming RL- and LLM-style planning; topological grafting creation converges through multiple rounds of iteration and generates new Vectoks. The results show that structured vector cognitive primitives offer falsifiable incremental value for physical reasoning and planning reliability.

关键词：意象优先；大形象模型；矢量图；全感官认知；体感锚点；拓扑嫁接；具

Keywords: Imagery first; Large Imagination Model; vector graph (Vectok); multi-sensory cognition; somatic anchor; topological grafting; embodied cognition