D
Domain Algebra Official Site
Author Essay作者文章

Where Does Reasoning Happen?推理在哪里发生

This essay is both a brief history of ideas and a statement of position. I want to clarify the historical line in which my work belongs, and why it is not something invented out of nowhere.这篇文章是一个思想史梳理,也是一个立场声明。我试图说清楚,我的工作在一条什么样的历史线上,以及为什么它不是凭空捏造的。

For two and a half millennia, people have discussed the form, limits, semantics, and world-structure of reasoning. Yet the question of the semantic space in which a reasoning step is valid has remained outside the system. Here I retrace that historical line, then explain what my work tries to place inside the reasoning unit itself.两千五百年来,人类一直在讨论推理的形式、边界、语义和世界结构,但“推理在哪个语义空间里发生”这个问题始终被放在系统外部。这里我沿着那条历史线回看它,然后说明我的工作试图把什么真正放进推理单元内部。

Domain AlgebraDomain Algebra Logic History逻辑史 Category Theory范畴论 Neural-Symbolic AI神经符号 AI
@D

Chao Li
Principal Investigator
AI EDtech Governance Trust
Deepleap.ai
lichao@deepleap.ai
Chao Li
首席研究员(PI)
AI EDtech Governance Trust(人工智能教育科技治理信托)
Deepleap.ai
lichao@deepleap.ai

Manifesto宣言

The world is not divided into mathematical problems and human problems. The real divide is between relations that have entered structure, and meanings that are still left outside reasoning.这个世界并不分成数学问题与人文问题。真正的分界在于:哪些关系已经进入结构,哪些意义仍被留在推理之外。

Essay文章

Can a label on the door open the door?贴在门上的标签,能打开门吗?

This is the question the essay tries to answer. For two and a half millennia, human beings have kept attaching labels to reasoning. This fact belongs to physics; that conclusion belongs to medicine. But labels are not structure, and domains have never been part of reasoning itself.这是这篇文章要回答的问题。两千五百年来,人类一直在给推理贴标签。这个事实属于物理学,那个结论属于医学。但标签从来不是结构,域从来不是推理的一部分。

Until now.直到现在。

A label says: “this fact belongs to a certain domain.” It is an annotation about the fact, and the reasoning engine may choose to ignore it.标签说的是:“这个事实属于某个域。”这是关于事实的注释,推理引擎可以选择忽略它。

A structure says: “this relation has meaning only inside this domain.” That is part of the relation's own definition, and the reasoning engine cannot route around it.结构说的是:“这个关系在这个域里才有意义。”这是关系本身的定义,推理引擎无法绕过它。

A domain-labeled triple (apple, is_a, fruit)[Biology] and a four-place relation in which the domain is built into the relation itself, is_a(apple, fruit, @Biology), are computationally very different things. In the former, is_a is global and the domain is an annotation. In the latter, is_a@Biology is a relation different from is_a@Business, and the domain is part of the definition.一个带域标签的三元组 (苹果, is_a, 水果)[Biology] 和一个把域编进关系的四元组 is_a(苹果, 水果, @Biology),在计算上是完全不同的东西。前者的 is_a 是全局的,域是注释。后者的 is_a@Biology 是一个不同于 is_a@Business 的关系,域是定义的一部分。

@D is the key that opens the door of reasoning, not the label pasted on the door.而 @D 是打开推理之门的钥匙,而不是贴在门上的标签。

A question never answered head-on一个从未被正面回答的问题

Human beings have been thinking about reasoning for two and a half thousand years.人类思考推理已经两千五百年了。

Over that time, we have accumulated extraordinarily rich results. We know the formal structure of reasoning, the limits of its computational power, the foundations of its semantics, and how it varies across possible worlds.在这两千五百年里,我们积累了极其丰富的成果。我们知道推理的形式结构是什么,知道推理的计算能力边界在哪里,知道推理的语义基础怎么建立,知道推理在不同可能世界里如何变化。

Where does reasoning happen?推理在哪里发生?

Not “what is reasoning?” Aristotle already answered that at the level of form. Not “what can reasoning compute?” Turing already answered that at the level of limits. The question is this: when a reasoning step occurs, in what semantic space is it valid, and who determines that?不是“推理是什么” 这件事,亚里士多德已经回答了形式。不是“推理能算什么”这件事,图灵已经回答了边界。而是:当一个推理步骤发生的时候,它在哪个语义空间里有效?谁决定了这件事?

The reason this question was never addressed directly is not that it was unimportant, but that every tradition hid its answer outside the system: in human assumptions, in query conditions, in application code, in the namespace of the knowledge base. It never entered the reasoning unit itself.这个问题之所以没有被正面回答,不是因为它不重要,而是因为所有的传统都把它的答案藏在了系统的外部。在人的假设里,在查询条件里,在应用代码里,在知识库的命名空间里。它从未进入推理单元本身。

I want to walk through that historical line first, and then say where I stand, what I found, and what I built.我想先把这条历史线走一遍,然后说我在哪里、发现了什么、做了什么。

First stop: Aristotle and the form of reasoning第一站:亚里士多德,推理的形式

Aristotle's syllogistic was the first systematic theory of reasoning in human history. Its central insight was that reasoning has form, and that form is independent of content.亚里士多德的三段论是人类历史上第一个系统化的推理理论。它的核心洞察是:推理有形式,这个形式独立于内容。

All humans are mortal. Socrates is human. Therefore Socrates is mortal.所有人都会死。苏格拉底是人。所以苏格拉底会死。

This reasoning is valid not because of Socrates as a person, but because of its form. No matter what A, B, and C are, if the premises are true, the conclusion follows.这个推理之所以有效,不是因为苏格拉底这个人,而是因为它的形式。无论 A、B、C 是什么,只要前提为真,结论就为真。

But Aristotle made a silent assumption here: all premises hold in the same world. Who supplied that semantic space, and how was it fixed? He did not ask. Two and a half thousand years later, that unasked question becomes decisive.但亚里士多德在这里做了一个无声的假设:所有前提在同一个世界里成立。这个语义空间是谁给的,怎么确定的,他没有问。这个没被问的问题,两千五百年后成了关键。

Second stop: Frege and Russell, the formalization of reasoning第二站:弗雷格与罗素,推理的形式化

Frege's Begriffsschrift in 1879 constructed the first complete formal logical system. Reasoning was no longer argument in natural language, but transformation inside a symbolic system. Russell and Whitehead pushed this program to its limit in Principia Mathematica.弗雷格 1879 年的《概念文字》构造了第一个完整的形式逻辑系统。推理不再是自然语言里的论证,而是符号系统中的变换。罗素和怀特海在《数学原理》里把这个计划推向极致。

The achievement of this movement is incomparable. It gave reasoning precision, verifiability, and the possibility of mechanization.这个运动的成就是无与伦比的。它给推理带来了精确性、可验证性、可机械化的可能。

But the cost was that semantics was suspended. The symbols inside the formal system had no meaning by themselves; meaning was supplied from outside, by human interpretation and by application context. The reasoning engine did not know what it was talking about. It knew only rules.但它的代价是:语义被悬置了。形式系统里的符号本身没有意义,意义由外部的人来赋予,由应用上下文来决定。推理引擎本身不知道自己在说什么,它只知道规则。

Here a seed was planted: the separation between form and semantics.这里埋下了一颗种子:形式和语义的分离。

Third stop: Turing and the computational boundary of reasoning第三站:图灵,推理的计算边界

Turing drew the boundary of the computational power of reasoning: what is computable is what a Turing machine can halt on and answer. This is one of the central intellectual achievements of the twentieth century.图灵划定了推理的计算能力边界:可计算的东西,就是图灵机能停机并给出答案的东西。这是二十世纪最重要的智识成就之一。

But the Turing machine has a structural property: its rules are global. There is no distinction between one set of rules for medical contexts and another for physical contexts.但图灵机有一个结构性的特点:它的规则是全局的。没有“在医学语境里用这些规则,在物理学语境里用那些规则”的区分。

The Turing machine is a domain-free reasoning machine. This is not a flaw; it was its design goal. But once we tried to use this framework to understand language and cognition, the absence of domains began to create friction.图灵机是一个无域的推理机器。这不是缺陷,是它的设计目标。但当我们后来试图用这个框架来理解人类语言和认知的时候,无域这件事就开始产生摩擦了。

Fourth stop: Tarski and the semantic foundation of reasoning第四站:塔斯基,推理的语义基础

Tarski brought semantics back in. His central idea was that truth is satisfaction in a model. This built a bridge between formal systems and the world, and became the foundation of modern logical semantics.塔斯基把语义找回来了。他的核心思想是真值是模型中的满足关系。这给形式系统和世界之间建立了桥梁,成为现代逻辑语义学的基础。

But Tarski's model is fixed and global. Within a single Tarskian model, an “atom” cannot mean the excitation of a quantum field in physics and an indivisible operation in computer science. To handle such polysemy, you need multiple different models, but Tarski's framework does not provide tools for the relations among them.但塔斯基的模型是固定的、全局的。你不能在同一个塔斯基模型里让“原子”在物理学语境里指量子场的激发,在计算机科学语境里指不可中断的操作。如果你想处理这种多义性,你需要多个不同的模型,但这些模型之间的关系,塔斯基的框架没有给出工具。

Fifth stop: Kripke and the semantic worlds of reasoning第五站:克里普克,推理的语义世界

Kripke's possible-world semantics is the step closest to my own work.克里普克的可能世界语义学是最接近我的工作的一步。

Its central idea is that truth is relative to possible worlds. A frame is a pair $(W, R)$, where $W$ is the set of worlds and $R$ is the accessibility relation between them.他的核心思想是真值是相对于可能世界的。一个框架是一对 $(W, R)$,$W$ 是世界的集合,$R$ 是世界之间的可达关系。

Kripke allowed us to say that the same concept may have different properties in different worlds. That is a major advance.克里普克让我们能够说:同一个概念在不同世界里可以有不同的性质。这是一个巨大的进步。

But Kripke's framework still has a limit: the accessibility relation $R$ is global. All worlds share the same relation. There is no distinction of the form “these two worlds are accessible in a medical context, but not in a physical context.”但克里普克的框架仍然有一个局限:可达关系 $R$ 是全局的。所有世界共享同一套可达关系。没有“在医学语境里这两个世界可达,在物理学语境里它们不可达”的区分。

Sixth stop: Gabbay and the relativization of reasoning第六站:加登福斯,推理的相对化

In 1975, Dov Gabbay did something crucial: he made the rules of reasoning themselves relative.Dov Gabbay 在 1975 年的工作中做了一件关键的事情:他让推理规则本身变成了相对的。

Reasoning was no longer one fixed formal system, but a parameterized family of systems. The validity of a reasoning step changes with context, and the valid set of rules changes with context as well.推理不是一个固定的形式系统,而是一个参数化的系统族。同一个推理步骤,在不同的上下文里有效性不同;有效的推理规则集,随着上下文变化。

This was the first time someone challenged, at the level of formal systems, the assumption that reasoning rules are global.这是第一次,有人在形式系统的层面上质疑“推理规则是全局的”这个假设。

But even Gabbay did not make context a first-class citizen of the reasoning unit. Context was still external, still meta-level. The reasoning engine did not know in which context it was operating.但即使是 Gabbay,也没有把上下文本身作为推理的一级公民。上下文仍然是外部的、元层的。推理引擎自己不知道自己在哪个上下文里。

Seventh stop: Grothendieck and Lawvere, the fibring of reasoning第七站:Grothendieck、Lawvere,推理的纤维化

In category theory, Grothendieck and Lawvere did something that looks abstract but is extraordinarily powerful: they made structure itself relative, internal, and computable.Grothendieck 和 Lawvere 在范畴论中做了一个看起来很抽象但极其强大的事情:他们让结构本身变成了相对的、内化的、可计算的。

Grothendieck's work on fibrations in the 1960s tells us that the existence and properties of objects are relative to a base space. The same concept may have different fibers over different base points.Grothendieck 在 1960 年代发展的纤维化思想告诉我们:对象的存在和性质,相对于一个基空间。同一个概念在不同的基点上有不同的纤维。

Lawvere brought this insight into the foundations of logic. Semantics is not absolute, but relative: relative to a domain, a context, a set of possible worlds. Rules of inference, truth, and even predicates themselves may all be defined relatively on fibers.Lawvere 把这个思想融入了逻辑的基础。他说,语义不是绝对的,而是相对的,相对于一个域、一个上下文、一个可能世界的集合。推理规则、真值、甚至谓词本身,都可以在纤维上相对化地定义。

This means that the domain can move from the outside of the system to the inside. From a label to a structure.这意味着:域可以从系统的外部,变成系统的内部。从标签变成结构。

Now: why the line stopped there现在:为什么这条线在这里停住了

Why did this line not continue after Gabbay and category theory? One practical reason is that, before large-scale knowledge systems existed, the question “in which domain does reasoning happen?” was a philosophical problem rather than an engineering problem.这条线走到 Gabbay 和范畴论之后,为什么没有继续走下去?一个很实际的原因是:在没有大规模知识系统的时代,“推理在哪个域里发生”是一个哲学问题,不是一个工程问题。

That is no longer the case. Large language models, knowledge graphs, and neuro-symbolic systems now process cross-domain reasoning every day at unprecedented scale. Yet their underlying architectures still inherit the unchallenged assumption of the past two and a half millennia: reasoning relations are global, and domains remain outside the system.现在不一样了。大型语言模型、知识图谱、神经符号系统每天都在处理跨域推理,规模是前所未有的。但它们的底层架构仍然继承了那个两千五百年来从未被质疑的假设:推理关系是全局的,域在系统外部。

LLM hallucination has many causes. One of its deepest causes is architectural: causal relations that are valid in medicine are not valid in meteorology; principles that hold in quantum mechanics do not hold for everyday objects. If the domain is not inside the reasoning unit, the system has no structural way to block this cross-domain contamination.LLM 的“幻觉”有很多来源。其中一个根本来源就是这个架构性问题:在医学域里成立的因果关系,在气象学域里不成立;在量子力学里成立的叠加原理,在日常物体上不成立。但因为域不在推理单元里,系统没有能力用推理结构本身来阻止这种跨域污染。

This is not a training-data problem, nor a model-size problem. It is an architectural problem.这不是训练数据的问题,也不是模型大小的问题,而是架构的问题。

What my work adds to this line我的工作在这条线上做了什么

I propose a framework: Domain-Contextualized Computation, or the relativized reasoning unit.我提出了一个框架:相对化推理单元,也就是 Domain-Contextualized Computation。

Its central claim is simple: reasoning relations are not global. The validity of reasoning is relative to a domain. That domain is not outside the system, but inside the reasoning unit itself, as a parameter of the relation.它的核心主张很简单:推理关系不是全局的。推理的有效性,是相对于一个域的。这个域,不在系统外部,而是在推理单元本身里,作为关系的一个参数。

$$c_1 \xrightarrow{r} c_2$$
$$c_1 \xrightarrow{r@d} c_2$$

This is not annotation. It is structure. @d is not a label; it changes the meaning, validity, and transmissibility of the relation.这不是注释。这是结构。@d 不是标签,它改变了关系的含义、有效性和可传播性。

At the level of knowledge representation: it becomes possible to express, inside one knowledge base, that an apple is a fruit in the biological domain but a company in the commercial domain, without contradiction.在知识表示上:能够在同一个知识库里表达“在生物学域内苹果是水果,但在商业域内苹果是一个公司”,而不产生逻辑矛盾。

At the level of reasoning: it becomes possible to block cross-domain contamination automatically, inside the reasoning process itself. Medical inference rules are not mistakenly applied to physical problems, because the fibers of the two domains are separated.在推理上:能够自动地、在推理过程中本身阻止跨域污染。医学推理规则不会被错误地应用到物理问题上,因为这两个域的纤维是分离的。

At the level of computation: the relativized semantics can be compiled into executable code using $\tau$-typed Galois connections and well-founded recursion. It has already been realized in memristive systems, affective computation, and language generation.在计算上:能够用 $\tau$-typed Galois 连接和 well-founded 递归,把这个相对化的语义编译成可执行的代码。在忆阻器、在情感计算、在语言生成上,都已经实现。

At the physical level: it can be implemented directly in silicon chips or memristor arrays through symmetry breaking and state encoding. Reasoning no longer remains a software simulation; it becomes a physical phenomenon.在物理上:能够用对称性破缺和状态编码,在硅芯片或忆阻器阵列上直接实现。推理不再是软件仿真,而是物理现象。

The hidden common root and the answer共同根源与答案

There is one hidden common root in the history of reasoning over the past two and a half millennia: the globality of reasoning relations was never itself questioned.两千五百年的推理史,有一个隐藏的共同根源:推理关系是全局的这件事从来没有被质疑。

The assumption was so natural that it did not even appear as an assumption. It was background, air, part of reasoning itself.这个假设太自然了,以至于它从来不被看作一个假设。它就是背景,是空气,是推理本身的一部分。

But like Aristotle's assumption that all premises hold in one world, once it is questioned, the whole framework must be reconsidered.但它就像亚里士多德关于“所有前提在同一世界”的假设一样,一旦被质疑,整个框架就必须被重新思考。

The answer is not complicated. The answer is simply this: let the domain enter. Not as annotation, but as structure. Not as metadata, but as part of arity. Reasoning relations are not merely binary; they are defined relative to domains.答案不复杂。答案就是:让域进来。不是作为注释,而是作为结构。不是作为元信息,而是作为 arity 的一部分。推理关系不是二元的,而是相对于域被定义的。

Once you make that change, everything else follows naturally. Grothendieck already showed how to formalize this structure, Lawvere already showed what properties it has, and Turing already showed how to compute it. My work is to assemble these fragments into a concrete system that is computable and physically realizable.一旦你做了这个改变,所有的后续都自然而然地跟着来了。Grothendieck 已经告诉我们如何形式化这个结构,Lawvere 已经告诉我们这个结构有什么性质,图灵已经告诉我们如何计算。我的工作,是把这些碎片在一个具体的、可计算的、可物理实现的系统里真正组装起来。

In one sentence一句话总结

Grothendieck built the engine, Lawvere proved that it can run, and this line of work is the first time it has been mounted into a vehicle and driven on the road, across cognitive, linguistic, affective, and physical domains at once.Grothendieck 造了发动机,Lawvere 证明了它能跑,这套工作是第一次把它装进车里开上路,在认知域、语言域、情感域、物理域里同时开。