TADC-06:關係優先認知與跨域連續性——從學科距離到關係距離的認知拓樸猜想
英文題名: Relation-First Cognition and Cross-Domain Continuity: From Disciplinary Distance to Relational Cognitive Distance系列: Topological Attention and Dynamic Cognitive Domains — Conjecture Series(TADC)中文系列名: 拓樸注意力與動態認知域命題系列編號: TADC-06版本: v0.1日期: 2026-08-17作者: Neo.K(許筌崴)協作: GPT-5.6 Sol文件性質: 理論命題/關係空間模型/可證偽研究綱領文獻檢索截點: 2026-08-17
摘要
跨領域思考通常以外部知識分類為基礎描述。例如,從數學轉向認知科學、從生物學轉向人工智慧,常被視為一次「跨域切換」。然而,這種描述預設了學科分類距離可以代表認知轉換距離。
本文提出相反的候選框架:
disciplinary distance ≠ cognitive relational distance . \boxed{
\text{disciplinary distance}
\neq
\text{cognitive relational distance}.
} disciplinary distance = cognitive relational distance .
一個表面上跨越多個學科的轉換,若兩個問題共享高度相似的因果結構、形式關係、程序結構、控制約束、類比映射或目標角色,則在認知系統的有效關係空間中可能仍是局部移動。反之,兩個被歸入同一學科的問題,若缺乏可重用的關係結構,對主體而言也可能需要高成本重建。
本文提出三個核心猜想:
Relation-First Cognition Conjecture(RFCC) :在部分高階推理中,關係結構比外部分類標籤更能預測有效認知鄰接與轉換成本;
Cross-Domain Continuity Conjecture(CDCC) :部分表面跨領域轉換,可在另一個 goal-conditioned relational topology 中形成連續或近連續的認知軌跡;
Relational Re-representation Conjecture(RRC) :跨域映射不只發現既有相似性,也可能透過 analogical alignment、abstraction 與 schema updating 改變後續有效關係表示。
本文定義外部分類距離:
d e x t ( x , y ) , d_{\mathrm{ext}}(x,y), d ext ( x , y ) ,
關係認知距離:
d r e l ( x , y ∣ G ) , d_{\mathrm{rel}}(x,y\mid G), d rel ( x , y ∣ G ) ,
以及跨域落差量:
Δ c r o s s = d e x t − d r e l . \boxed{
\Delta_{\mathrm{cross}}
=
d_{\mathrm{ext}}
-
d_{\mathrm{rel}}.
} Δ cross = d ext − d rel .
當:
Δ c r o s s ≫ 0 , \Delta_{\mathrm{cross}}\gg0, Δ cross ≫ 0 ,
稱為 Apparent Cross-Domain / Relationally Local Transition :外部分類看似遙遠,但內部關係上近鄰。
本文進一步使用 multiplex relational graph 表示語義、因果、類比、程序、時間、目標與控制關係,並提出 structural-alignment score、bridge density、translation cost、relational path cost 與 continuity ratio 等可測量量。
現有 analogical reasoning 研究已證明人類可跨不同 sensory modalities 進行 analogical mapping;實驗也觀察到 relational re-representation;計算模型則展示 relational representations 可支援跨 video games 與心理任務的 cross-domain generalization。2025/2026 年的 schema-drift 研究進一步顯示,抽象 relational schemas 在反覆類比使用後可以發生可測改變。同時,semantic-control 與 task-space cognitive-map 研究顯示,相同刺激可因目標與控制需求形成不同的大尺度認知/腦狀態,且 task-relevant abstract structure 可被建構為 cognitive maps。
這些結果支持「關係結構具有跨表面分類的認知作用」,但尚未證明 RFCC / CDCC。本文因此建立 taxonomy-only、surface-similarity、fixed-semantic-space、ordinary analogy、random-association 與 expertise-only 等競爭模型,並提出跨學科距離矩陣、relation-matched switching、bridge induction、re-representation、schema drift、path-return 與 multi-scale continuity 等實驗。
本文的核心主張不是「所有領域其實都一樣」,而是:
external categorical discontinuity ⇏ internal cognitive discontinuity . \boxed{
\text{external categorical discontinuity}
\not\Rightarrow
\text{internal cognitive discontinuity}.
} external categorical discontinuity ⇒ internal cognitive discontinuity .
如果關係距離不能比學科標籤、語義相似度與一般熟悉度提供額外預測,則本篇強版本應被拒絕。
關鍵詞: relational cognition;cross-domain reasoning;analogy;structural alignment;cognitive distance;cognitive topology;re-representation;cognitive maps;semantic control;TADC
0. 邊界聲明
本文不是要主張:
「所有知識本質上都是一樣的。」
也不是:
「學科分類沒有意義。」
學科分類對:
知識累積;
教育;
方法論;
社群;
審查;
工具;
專業規範;
都具有重要功能。
本文研究的是另一個問題:
Are disciplinary boundaries the same as cognitive transition boundaries? \boxed{
\text{Are disciplinary boundaries
the same as cognitive transition boundaries?}
} Are disciplinary boundaries the same as cognitive transition boundaries?
本文提出的答案只是:
not necessarily . \boxed{
\text{not necessarily}.
} not necessarily .
此外,「關係優先」不是指:
relations always dominate objects . \text{relations always dominate objects}. relations always dominate objects .
而是:
在部分需要 transfer、analogy、abstraction、cross-domain inference 與 problem reformulation 的高階 cognition 中,relational structure 可能比表面 category labels 更能預測認知鄰接。
1. 從 TADC-02 的問題重新開始
TADC-02 已提出:
d d i s c i p l i n e ( x , y ) ≠ d c o g n i t i v e ( x , y ) . d_{\mathrm{discipline}}(x,y)
\neq
d_{\mathrm{cognitive}}(x,y). d discipline ( x , y ) = d cognitive ( x , y ) .
但當時只是概念區分。
TADC-06 的目標是:
定義這兩種距離;
找到可觀察 proxy;
建立競爭模型;
決定是否真的需要「跨域連續性」這個構念。
2. 外部分類距離
令外部 taxonomy:
Π e x t = { D 1 , D 2 , … , D m } . \Pi^{\mathrm{ext}}
=
\{
D_1,D_2,\ldots,D_m
\}. Π ext = { D 1 , D 2 , … , D m } .
每個 cognitive object:
x x x
可具有一個或多個外部分類標籤:
L e x t ( x ) . L_{\mathrm{ext}}(x). L ext ( x ) .
最簡距離:
d e x t ( x , y ) = { 0 , L ( x ) = L ( y ) , 1 , L ( x ) ≠ L ( y ) . d_{\mathrm{ext}}(x,y)
=
\begin{cases}
0,&L(x)=L(y),\\
1,&L(x)\neq L(y).
\end{cases} d ext ( x , y ) = { 0 , 1 , L ( x ) = L ( y ) , L ( x ) = L ( y ) .
但真實 taxonomy 通常 hierarchical。
因此可用 tree distance:
d e x t t r e e ( x , y ) d_{\mathrm{ext}}^{\mathrm{tree}}
(
x,y
) d ext tree ( x , y )
或 ontology graph distance。
3. 外部分類距離的限制
兩個 concepts:
x , y x,y x , y
可能分別屬於:
D A , D B . D_A,
D_B. D A , D B .
即使:
d e x t ( x , y ) ≫ 0 , d_{\mathrm{ext}}(x,y)
\gg0, d ext ( x , y ) ≫ 0 ,
它們可能共享:
feedback;
invariance;
conservation;
recursion;
optimization;
hierarchy;
phase transition;
information bottleneck;
control loop。
因此:
external label distance \boxed{
\text{external label distance}
} external label distance
不必等於:
structural reasoning distance . \boxed{
\text{structural reasoning distance}.
} structural reasoning distance .
4. Multiplex Relational Graph
定義 cognitive relational graph:
G t = ( V t , E t ( S ) , E t ( C ) , E t ( A ) , E t ( P ) , E t ( T ) , E t ( G ) , E t ( K ) ) . \mathcal G_t
=
(
V_t,
E_t^{(S)},
E_t^{(C)},
E_t^{(A)},
E_t^{(P)},
E_t^{(T)},
E_t^{(G)},
E_t^{(K)}
). G t = ( V t , E t ( S ) , E t ( C ) , E t ( A ) , E t ( P ) , E t ( T ) , E t ( G ) , E t ( K ) ) .
其中:
E ( S ) E^{(S)} E ( S ) :semantic relations;
E ( C ) E^{(C)} E ( C ) :causal relations;
E ( A ) E^{(A)} E ( A ) :analogical relations;
E ( P ) E^{(P)} E ( P ) :procedural relations;
E ( T ) E^{(T)} E ( T ) :temporal relations;
E ( G ) E^{(G)} E ( G ) :goal relations;
E ( K ) E^{(K)} E ( K ) :control / constraint relations。
因此 cognition 不在單一 semantic embedding 中發生,
而可能在:
multiplex relational space \boxed{
\text{multiplex relational space}
} multiplex relational space
中發生。
5. 關係認知距離
每種 relation:
r r r
有權重:
w r ( G t ) . w_r(G_t). w r ( G t ) .
定義 aggregate accessibility:
κ t ( x , y ∣ G t ) = F ( κ S , κ C , κ A , κ P , κ T , κ G , κ K ) . \kappa_t(x,y\mid G_t)
=
F
\left(
\kappa_S,
\kappa_C,
\kappa_A,
\kappa_P,
\kappa_T,
\kappa_G,
\kappa_K
\right). κ t ( x , y ∣ G t ) = F ( κ S , κ C , κ A , κ P , κ T , κ G , κ K ) .
若 path:
γ = x 0 → x 1 → ⋯ → x n , \gamma
=
x_0\rightarrow x_1\rightarrow\cdots\rightarrow x_n, γ = x 0 → x 1 → ⋯ → x n ,
可定義:
K r e l ( γ ∣ G ) = − ∑ i = 0 n − 1 log κ ( x i , x i + 1 ∣ G ) . K_{\mathrm{rel}}(\gamma\mid G)
=
-\sum_{i=0}^{n-1}
\log
\kappa(
x_i,x_{i+1}\mid G
). K rel ( γ ∣ G ) = − i = 0 ∑ n − 1 log κ ( x i , x i + 1 ∣ G ) .
因此:
d r e l ( x , y ∣ G ) = min γ : x ⇝ y K r e l ( γ ∣ G ) . \boxed{
d_{\mathrm{rel}}(x,y\mid G)
=
\min_{\gamma:x\leadsto y}
K_{\mathrm{rel}}(\gamma\mid G).
} d rel ( x , y ∣ G ) = γ : x ⇝ y min K rel ( γ ∣ G ) .
6. 為什麼使用 goal-conditioned distance?
同兩個 concepts:
x , y x,y x , y
在不同問題下可能:
d r e l ( x , y ∣ G 1 ) ≪ d r e l ( x , y ∣ G 2 ) . d_{\mathrm{rel}}(x,y\mid G_1)
\ll
d_{\mathrm{rel}}(x,y\mid G_2). d rel ( x , y ∣ G 1 ) ≪ d rel ( x , y ∣ G 2 ) .
例如:
在:
G 1 = compare feedback structures G_1
=
\text{compare feedback structures} G 1 = compare feedback structures
時,
兩個不同學科系統可能非常近。
但在:
G 2 = compare empirical measurement protocols G_2
=
\text{compare empirical measurement protocols} G 2 = compare empirical measurement protocols
時,
它們可能非常遠。
因此:
d r e l = d r e l ( x , y ∣ G ) . \boxed{
d_{\mathrm{rel}}
=
d_{\mathrm{rel}}(x,y\mid G).
} d rel = d rel ( x , y ∣ G ) .
7. Relation-First Cognition Conjecture(RFCC)
RFCC 宣稱:
在至少部分高階推理與知識轉移任務中,goal-conditioned relational distance 比 external disciplinary distance 更能預測 transition cost、inference success 與 transfer。
形式:
Pred ( d r e l ) > Pred ( d e x t ) \boxed{
\operatorname{Pred}
(
d_{\mathrm{rel}}
)
>
\operatorname{Pred}
(
d_{\mathrm{ext}}
)
} Pred ( d rel ) > Pred ( d ext )
在適當控制:
familiarity;
word similarity;
education;
exposure;
motor / perceptual demands;
後成立。
8. 「關係優先」不是「語義優先」
semantic similarity:
s s e m ( x , y ) s_{\mathrm{sem}}(x,y) s sem ( x , y )
只是:
G \mathcal G G
的一層。
兩個表面語義很遠的情境:
s s e m ↓ s_{\mathrm{sem}}\downarrow s sem ↓
可能仍有:
s s t r u c t ↑ . s_{\mathrm{struct}}\uparrow. s struct ↑ .
這正是 analogy literature 長期研究的核心現象之一。
所以:
relation-first ≠ semantic-similarity-first . \boxed{
\text{relation-first}
\neq
\text{semantic-similarity-first}.
} relation-first = semantic-similarity-first .
9. Analogical Mapping 的既有基礎
analogical reasoning 經常要求:
source ↔ target \text{source}
\leftrightarrow
\text{target} source ↔ target
表面 items 不同,
但:
relational structure \text{relational structure} relational structure
相似。
因此:
surface dissimilarity + structural similarity \boxed{
\text{surface dissimilarity}
+
\text{structural similarity}
} surface dissimilarity + structural similarity
不是新發現。
TADC-06 的新問題是:
這種 structural similarity 是否能被提升為對 cognitive transition distance 的一般描述?
10. 跨 modality analogy
Weinberger 等人(2022)研究 analogical mapping across sensory modalities,顯示 analogy ability 並不只局限於單一資訊 modality。
這支持:
relational mapping can cross surface representational format . \boxed{
\text{relational mapping can cross
surface representational format}.
} relational mapping can cross surface representational format .
但跨 modality 不等於跨 disciplinary domain。
因此只是鄰接證據。
11. Cross-Domain Generalization 的計算模型
Doumas、Puebla、Martin 與 Hummel(2022)提出 relation-learning / cross-domain generalization theory。
該模型能:
從簡單 visual stimuli 學 relational representations;
在 Breakout 與 Pong 等不同 domains 間 generalize;
在不同心理任務之間做 relational transfer。
這提供一個重要 possibility proof:
cross-domain transfer can be modeled through structured relational representations . \boxed{
\text{cross-domain transfer can be modeled
through structured relational representations}.
} cross-domain transfer can be modeled through structured relational representations .
但 computational sufficiency 不等於 human cognitive necessity。
12. Relational Re-representation
Lu、Wu 與 Holyoak 等人的 analogical re-representation work 顯示:
analogical mapping 過程中,
relation representation 本身可能發生改變,
以允許原本不完全相同的 relations 被重新對齊。
因此 analogy 不一定只是:
find existing isomorphism . \text{find existing isomorphism}. find existing isomorphism .
也可能:
transform representations so alignment becomes possible . \boxed{
\text{transform representations so alignment becomes possible}.
} transform representations so alignment becomes possible .
這與 TADC-01 的 ASTC 非常接近。
13. Relational Re-representation Conjecture(RRC)
本文提出更一般版本:
若:
x , y x,y x , y
原本:
d r e l ( t ) ( x , y ) ≫ 0 , d_{\mathrm{rel}}^{(t)}(x,y)
\gg0, d rel ( t ) ( x , y ) ≫ 0 ,
經:
alignment / abstraction / schema induction \text{alignment / abstraction / schema induction} alignment / abstraction / schema induction
後:
d r e l ( t + 1 ) ( x , y ) < d r e l ( t ) ( x , y ) , d_{\mathrm{rel}}^{(t+1)}(x,y)
<
d_{\mathrm{rel}}^{(t)}(x,y), d rel ( t + 1 ) ( x , y ) < d rel ( t ) ( x , y ) ,
則:
cognitive distance itself has been transformed . \boxed{
\text{cognitive distance itself has been transformed}.
} cognitive distance itself has been transformed .
RRC 宣稱:
cross-domain reasoning can sometimes change the relational metric it uses . \boxed{
\text{cross-domain reasoning can sometimes
change the relational metric it uses}.
} cross-domain reasoning can sometimes change the relational metric it uses .
14. Schema Drift
Vagnino 與 Walker 的 2025/2026 Cognition work 研究:
abstract relational schemas \text{abstract relational schemas} abstract relational schemas
是否會在類比使用中改變。
結果指出 abstract schemas 在特定條件下確實會 drift。
這是一個很重要的鄰接結果:
relational abstraction itself can be plastic . \boxed{
\text{relational abstraction itself can be plastic}.
} relational abstraction itself can be plastic .
但 schema drift 不自動證明:
attention topology . \text{attention topology}. attention topology .
它只使:
R t ≠ R t + 1 \mathcal R_t
\neq
\mathcal R_{t+1} R t = R t + 1
變得更加實驗上可信。
15. Cross-Domain Continuity Conjecture(CDCC)
令一條 trajectory:
γ = ( x 0 , x 1 , … , x n ) . \gamma
=
(x_0,x_1,\ldots,x_n). γ = ( x 0 , x 1 , … , x n ) .
外部 taxonomy:
L e x t ( x i ) L_{\mathrm{ext}}(x_i) L ext ( x i )
可能頻繁改變。
如果:
d r e l ( x i , x i + 1 ∣ G ) ≤ ϵ d_{\mathrm{rel}}
(
x_i,x_{i+1}
\mid G
)
\leq
\epsilon d rel ( x i , x i + 1 ∣ G ) ≤ ϵ
對大部分 (i) 成立,
則在 relational topology 中:
γ \gamma γ
仍是一條 low-cost path。
因此:
external domain discontinuity ⇏ relational cognitive discontinuity . \boxed{
\text{external domain discontinuity}
\not\Rightarrow
\text{relational cognitive discontinuity}.
} external domain discontinuity ⇒ relational cognitive discontinuity .
16. 跨域落差量
定義:
Δ c r o s s ( x , y ∣ G ) = d ^ e x t ( x , y ) − d ^ r e l ( x , y ∣ G ) \boxed{
\Delta_{\mathrm{cross}}
(
x,y\mid G
)
=
\widehat d_{\mathrm{ext}}(x,y)
-
\widehat d_{\mathrm{rel}}(x,y\mid G)
} Δ cross ( x , y ∣ G ) = d ext ( x , y ) − d rel ( x , y ∣ G )
其中兩距離先標準化。
17. 四種情形
Type I — Same-domain / Relationally Local
d e x t ↓ , d r e l ↓ . d_{\mathrm{ext}}\downarrow,
\qquad
d_{\mathrm{rel}}\downarrow. d ext ↓ , d rel ↓ .
最普通。
Type II — Cross-domain / Relationally Local
d e x t ↑ , d r e l ↓ . d_{\mathrm{ext}}\uparrow,
\qquad
d_{\mathrm{rel}}\downarrow. d ext ↑ , d rel ↓ .
即:
Δ c r o s s ≫ 0. \Delta_{\mathrm{cross}}\gg0. Δ cross ≫ 0.
這是 TADC-06 最關注情形。
Type III — Same-domain / Relationally Distant
d e x t ↓ , d r e l ↑ . d_{\mathrm{ext}}\downarrow,
\qquad
d_{\mathrm{rel}}\uparrow. d ext ↓ , d rel ↑ .
外部分類相同,
但 cognition 可能需要大幅重建。
Type IV — Cross-domain / Relationally Distant
d e x t ↑ , d r e l ↑ . d_{\mathrm{ext}}\uparrow,
\qquad
d_{\mathrm{rel}}\uparrow. d ext ↑ , d rel ↑ .
真正的高成本跨域。
18. 所以「跨領域能力」本身也要重新拆
一般說某人:
cross-domain ability ↑ . \text{cross-domain ability}\uparrow. cross-domain ability ↑ .
可能至少表示:
A. Low Relational Distance
他已經建立很多跨 domain bridge。
B. High Translation Ability
即使 distance 大,
也能做 chart translation。
C. High Search Ability
能找出 hidden alignment。
D. High Re-representation Ability
能改造表示以產生 alignment。
E. High Re-indexing Ability
能換尺度重新描述問題。
因此:
cross-domain cognition ≠ one scalar talent . \boxed{
\text{cross-domain cognition}
\neq
\text{one scalar talent}.
} cross-domain cognition = one scalar talent .
19. Structural Alignment Score
對兩個 local structures:
U A , U B , U_A,
U_B, U A , U B ,
定義候選:
S a l i g n ( U A , U B ) S_{\mathrm{align}}
(
U_A,U_B
) S align ( U A , U B )
考慮:
role correspondence;
relation type correspondence;
causal ordering;
graph motif;
constraint structure;
goal role。
例如:
S a l i g n = α S R + β S C + γ S G + δ S M . S_{\mathrm{align}}
=
\alpha S_R
+
\beta S_C
+
\gamma S_G
+
\delta S_M. S align = α S R + β S C + γ S G + δ S M .
20. 高 alignment 不要求 item similarity
可能:
S i t e m ≈ 0 S_{\mathrm{item}}\approx0 S item ≈ 0
但:
S a l i g n ≈ 1. S_{\mathrm{align}}\approx1. S align ≈ 1.
這正是:
deep analogy . \boxed{
\text{deep analogy}.
} deep analogy .
所以 TADC 的 distance 應讓:
S a l i g n S_{\mathrm{align}} S align
降低:
d r e l . d_{\mathrm{rel}}. d rel .
21. Bridge Object / Bridge Relation
TADC-02 已定義 bridge object。
本文加入:
B α β = { b : b supports low-cost relational translation } . B_{\alpha\beta}
=
\{
b:
b
\text{ supports low-cost relational translation}
\}. B α β = { b : b supports low-cost relational translation } .
bridge 不一定是一個 object。
它可能是一個 relation pattern:
r ∗ . r^*. r ∗ .
例如:
r ∗ = feedback loop . r^*
=
\text{feedback loop}. r ∗ = feedback loop .
則:
r ∗ r^* r ∗
可以同時連接:
biology;
control engineering;
economics;
cognition。
22. Bridge Density
定義:
ρ B ( U A , U B ) = ∣ B A B ∣ ∣ U A ∣ + ∣ U B ∣ \rho_B
(
U_A,U_B
)
=
\frac{
|B_{AB}|
}{
|U_A|+|U_B|
} ρ B ( U A , U B ) = ∣ U A ∣ + ∣ U B ∣ ∣ B A B ∣
或 weighted version:
ρ B w = ∑ b ∈ B w b . \rho_B^w
=
\sum_{b\in B}
w_b. ρ B w = b ∈ B ∑ w b .
候選預測:
ρ B ↑ ⇒ K s w i t c h ↓ . \rho_B\uparrow
\Rightarrow
K_{\mathrm{switch}}\downarrow. ρ B ↑⇒ K switch ↓ .
23. Translation Cost
沿用 TADC-02:
K α β = K r e t r i e v a l + K r e c o d e + K c o n t e x t + K l o s s . K_{\alpha\beta}
=
K_{\mathrm{retrieval}}
+
K_{\mathrm{recode}}
+
K_{\mathrm{context}}
+
K_{\mathrm{loss}}. K α β = K retrieval + K recode + K context + K loss .
加入:
K a l i g n . K_{\mathrm{align}}. K align .
因此:
K α β = K R + K C + K L + K A . \boxed{
K_{\alpha\beta}
=
K_R+K_C+K_L+K_A.
} K α β = K R + K C + K L + K A .
若:
S a l i g n ↑ , S_{\mathrm{align}}\uparrow, S align ↑ ,
通常預測:
K A ↓ . K_A\downarrow. K A ↓ .
24. Cognitive Discontinuity Index
定義:
CDI ( x , y ) = α K c o n t e x t + β K a l i g n + γ K r e t r i e v a l + δ d r e l . \operatorname{CDI}(x,y)
=
\alpha K_{\mathrm{context}}
+
\beta K_{\mathrm{align}}
+
\gamma K_{\mathrm{retrieval}}
+
\delta d_{\mathrm{rel}}. CDI ( x , y ) = α K context + β K align + γ K retrieval + δ d rel .
如果:
C D I ↓ CDI\downarrow C D I ↓
即使:
d e x t ↑ , d_{\mathrm{ext}}\uparrow, d ext ↑ ,
實際 transition 可能仍然流暢。
25. 跨域連續不是「沒有切換」
重要限制:
如果:
L e x t ( x ) ≠ L e x t ( y ) , L_{\mathrm{ext}}(x)\neq L_{\mathrm{ext}}(y), L ext ( x ) = L ext ( y ) ,
外部分類確實改變。
CDCC 不是否認這件事。
而是說:
categorical switch ≠ high cognitive discontinuity . \boxed{
\text{categorical switch}
\neq
\text{high cognitive discontinuity}.
} categorical switch = high cognitive discontinuity .
26. Relation-First 與 Category-First
定義兩種理想化 policy。
Category-First
先:
L ( x ) L(x) L ( x )
再找 domain-specific method:
M L ( x ) . M_{L(x)}. M L ( x ) .
流程:
x → D i → M i . x
\rightarrow
D_i
\rightarrow
M_i. x → D i → M i .
Relation-First
先抽:
R ( x ) \mathcal R(x) R ( x )
再找:
R ( y ) ≈ R ( x ) . \mathcal R(y)
\approx
\mathcal R(x). R ( y ) ≈ R ( x ) .
流程:
x → r ∗ → { y 1 , y 2 , … } . x
\rightarrow
r^*
\rightarrow
\{y_1,y_2,\ldots\}. x → r ∗ → { y 1 , y 2 , … } .
27. 兩者不應被道德化
Category-first 在很多領域很有效:
法律;
medicine;
engineering standards;
taxonomy;
regulated procedure。
Relation-first 對:
analogy;
theory transfer;
creativity;
model reuse;
abstraction;
可能更有效。
所以:
relation-first ≠ universally superior . \boxed{
\text{relation-first}
\neq
\text{universally superior}.
} relation-first = universally superior .
它是一種 task-dependent policy。
28. Relation-First Policy Selector
令:
P R F = P ( relation-first ∣ G , X , R , C ) . P_{\mathrm{RF}}
=
P(
\text{relation-first}
\mid
G,
X,
R,
C
). P RF = P ( relation-first ∣ G , X , R , C ) .
如果 goal:
G G G
要求:
novel transfer;
cross-domain inference;
structural discovery;
則:
P R F ↑ P_{\mathrm{RF}}\uparrow P RF ↑
可能更有效。
若要求:
strict compliance;
exact domain convention;
fixed procedure;
則:
P R F ↓ P_{\mathrm{RF}}\downarrow P RF ↓
可能更有效。
29. Semantic Control 的鄰接證據
Wang 等人(2024)使用相同或高度匹配 stimuli,
比較:
global semantic association;
semantic feature matching;
non-semantic control tasks。
其結果顯示:
不同 retrieval demands 形成不同 macroscale brain-state configurations。
這支持:
same inputs can induce different effective relational retrieval states depending on goal . \boxed{
\text{same inputs can induce different effective
relational retrieval states depending on goal}.
} same inputs can induce different effective relational retrieval states depending on goal .
但這不是 RFCC 的直接證明。
30. Task-Space Cognitive Maps
Tan 等人(2025)顯示 medial / lateral OFC 等區域對 task-space cognitive map 有互補表徵,
支持:
task-relevant abstract state structure can be explicitly represented . \boxed{
\text{task-relevant abstract state structure
can be explicitly represented}.
} task-relevant abstract state structure can be explicitly represented .
這使:
d r e l ( x , y ∣ G ) d_{\mathrm{rel}}(x,y\mid G) d rel ( x , y ∣ G )
不再只是語言遊戲。
但仍需行為與 neural data 共同驗證。
31. 關係空間可以不是唯一的
同一:
X X X
可有:
G ( c a u s a l ) , \mathcal G^{(causal)}, G ( c a u s a l ) ,
G ( s e m a n t i c ) , \mathcal G^{(semantic)}, G ( se man t i c ) ,
G ( p r o c e d u r a l ) . \mathcal G^{(procedural)}. G ( p r oce d u r a l ) .
因此:
one object set ≠ one cognitive topology . \boxed{
\text{one object set}
\neq
\text{one cognitive topology}.
} one object set = one cognitive topology .
goal:
G G G
決定哪一層 weighting 上升。
32. Relation Layer Switching
如果從:
E ( S ) E^{(S)} E ( S )
切到:
E ( C ) , E^{(C)}, E ( C ) ,
即由 semantic similarity 看問題,
改成 causal structure 看問題,
這是一種:
relation-layer switch . \boxed{
\text{relation-layer switch}.
} relation-layer switch .
它可能:
X t = X t + 1 X_t=X_{t+1} X t = X t + 1
但:
d r e l ( t ) ≠ d r e l ( t + 1 ) . d_{\mathrm{rel}}^{(t)}
\neq
d_{\mathrm{rel}}^{(t+1)}. d rel ( t ) = d rel ( t + 1 ) .
這又是一種 TADC-01 式 space transformation。
33. 跨域推理的三階段
候選模型:
Stage 1 — Retrieval
找:
U B U_B U B
作為 candidate source。
Stage 2 — Alignment
計算:
S a l i g n ( U A , U B ) . S_{\mathrm{align}}(U_A,U_B). S align ( U A , U B ) .
Stage 3 — Re-representation
必要時:
R A , R B → R ~ A , R ~ B \mathcal R_A,\mathcal R_B
\rightarrow
\widetilde{\mathcal R}_A,
\widetilde{\mathcal R}_B R A , R B → R A , R B
使共通結構可被抽取。
34. 第四階段:Transfer
建立:
M : U A ⇝ U B . M:
U_A
\rightsquigarrow
U_B. M : U A ⇝ U B .
如果:
M M M
支援 novel inference,
才算真正 cross-domain transfer。
單純說:
「這兩個很像。」
不夠。
35. 第五階段:Schema Update
從 source / target 對齊後形成:
S ∗ . S^*. S ∗ .
而:
S ∗ S^* S ∗
會影響未來:
d r e l . d_{\mathrm{rel}}. d rel .
因此:
cross-domain transfer can alter future cross-domain distance . \boxed{
\text{cross-domain transfer
can alter future cross-domain distance}.
} cross-domain transfer can alter future cross-domain distance .
這是 RRC / schema-drift 方向最重要的推論。
36. Distance Plasticity
定義:
Δ d r e l = d r e l p o s t − d r e l p r e . \Delta d_{\mathrm{rel}}
=
d_{\mathrm{rel}}^{\mathrm{post}}
-
d_{\mathrm{rel}}^{\mathrm{pre}}. Δ d rel = d rel post − d rel pre .
若 alignment learning 後:
Δ d r e l < 0 , \Delta d_{\mathrm{rel}}<0, Δ d rel < 0 ,
則 cognitive distance 被壓縮。
若錯誤 analogy 被修正:
Δ d r e l > 0 , \Delta d_{\mathrm{rel}}>0, Δ d rel > 0 ,
則 cognitive distance 被拉開。
所以:
d r e l itself may be plastic . \boxed{
d_{\mathrm{rel}}
\text{ itself may be plastic}.
} d rel itself may be plastic .
37. Relation Compression
多個 domains:
D 1 , … , D n D_1,\ldots,D_n D 1 , … , D n
若共享:
r ∗ , r^*, r ∗ ,
則可以 coarse-grain 成:
[ r ∗ ] \boxed{
[r^*]
} [ r ∗ ]
這是一個 relation-centered super-domain。
例如:
{ D 1 , D 2 , D 3 } → U r ∗ . \{D_1,D_2,D_3\}
\rightarrow
U_{r^*}. { D 1 , D 2 , D 3 } → U r ∗ .
這是一種 TADC-04 Re-indexing。
38. 所謂「沒有領域」的嚴格版本
本文不使用:
沒有任何領域。
而使用:
No fixed single partition is assumed to be cognitively privileged across all goals . \boxed{
\text{No fixed single partition is assumed
to be cognitively privileged across all goals}.
} No fixed single partition is assumed to be cognitively privileged across all goals .
中文:
不假定存在一套對所有目標都具有認知優先性的固定領域分割。
這比「領域不存在」精確得多。
39. 多個合法 atlas
同一:
X X X
可能有:
A ( d i s c i p l i n a r y ) , \mathfrak A^{(disciplinary)}, A ( d i sc i pl ina r y ) ,
A ( c a u s a l ) , \mathfrak A^{(causal)}, A ( c a u s a l ) ,
A ( f o r m a l ) , \mathfrak A^{(formal)}, A ( f or ma l ) ,
A ( p r o c e d u r a l ) . \mathfrak A^{(procedural)}. A ( p r oce d u r a l ) .
不同 atlas 可能都有效,
但適用 goal 不同。
因此:
multiple valid atlases ≠ no structure . \boxed{
\text{multiple valid atlases}
\neq
\text{no structure}.
} multiple valid atlases = no structure .
40. Relation-First Cognition 的風險:錯誤 analogy
若看到:
r A ≈ r B r_A\approx r_B r A ≈ r B
就過度 transfer,
可能造成:
false structural equivalence . \boxed{
\text{false structural equivalence}.
} false structural equivalence .
例如:
surface relation 類似但 causal mechanism 不同;
mathematical form 相同但 domain assumptions 不同;
control architecture 類似但 scale / noise regime 不同。
所以 RFCC 需要:
alignment + constraint checking . \boxed{
\text{alignment}
+
\text{constraint checking}.
} alignment + constraint checking .
41. Constraint-Preserving Mapping
一個合法 mapping:
M : U A → U B M:U_A\rightarrow U_B M : U A → U B
不能只保 relation names。
還要測:
C A → C B . C_A
\rightarrow
C_B. C A → C B .
定義:
Q M = S a l i g n − λ L c o n s t r a i n t . Q_M
=
S_{\mathrm{align}}
-
\lambda
L_{\mathrm{constraint}}. Q M = S align − λ L constraint .
其中:
L c o n s t r a i n t L_{\mathrm{constraint}} L constraint
是 constraint violation。
只有:
Q M > θ Q_M>\theta Q M > θ
才接受 transfer。
42. Cross-Domain Hallucination
在 human / AI reasoning 中,
若:
S s u r f a c e ↑ S_{\mathrm{surface}}\uparrow S surface ↑
但:
S c o n s t r a i n t ↓ , S_{\mathrm{constraint}}\downarrow, S constraint ↓ ,
仍強行 Gluing,
會產生:
cross-domain hallucination . \boxed{
\text{cross-domain hallucination}.
} cross-domain hallucination .
這裡 hallucination 只是廣義錯誤結構對齊,
不是精神醫學術語。
43. Detachment 在跨域研究中的必要性
TADC-03 的:
D = Detachment D
=
\text{Detachment} D = Detachment
在這裡尤其重要。
一個 analogy:
A ↔ B A\leftrightarrow B A ↔ B
部分成立,
不能因此所有 relations 都被 Gluing。
需要:
D [ R i n v a l i d ] . D[
R_{\mathrm{invalid}}
]. D [ R invalid ] .
因此成熟 relation-first cognition 不是:
G → G → G G\rightarrow G\rightarrow G G → G → G
而是:
G → test → D → G v a l i d . \boxed{
G
\rightarrow
\text{test}
\rightarrow
D
\rightarrow
G_{\mathrm{valid}}.
} G → test → D → G valid .
44. Cross-Domain Continuity Ratio
對 trajectory:
γ = ( x 1 , … , x n ) , \gamma=(x_1,\ldots,x_n), γ = ( x 1 , … , x n ) ,
定義:
C C R ( γ ) = ∑ i 1 [ d r e l ( x i , x i + 1 ∣ G ) < θ ] n − 1 . CCR(\gamma)
=
\frac{
\sum_i
\mathbf 1[
d_{\mathrm{rel}}(x_i,x_{i+1}\mid G)<\theta
]
}{
n-1
}. C C R ( γ ) = n − 1 ∑ i 1 [ d rel ( x i , x i + 1 ∣ G ) < θ ] .
如果:
C C R → 1 , CCR\rightarrow1, C C R → 1 ,
即使 external labels 頻繁改變,
trajectory 仍具有高 relational continuity。
45. External Switching Ratio
定義:
E S R ( γ ) = ∑ i 1 [ L e x t ( x i ) ≠ L e x t ( x i + 1 ) ] n − 1 . ESR(\gamma)
=
\frac{
\sum_i
\mathbf 1[
L_{\mathrm{ext}}(x_i)
\neq
L_{\mathrm{ext}}(x_{i+1})
]
}{
n-1
}. E S R ( γ ) = n − 1 ∑ i 1 [ L ext ( x i ) = L ext ( x i + 1 )] .
關鍵 signature:
E S R ↑ ∧ C C R ↑ . \boxed{
ESR\uparrow
\quad\land\quad
CCR\uparrow.
} E S R ↑ ∧ C C R ↑ .
也就是:
表面高跨域,內部高連續。
46. Relation-First Signature
因此候選 relation-first trajectory:
Θ R F = ( E S R , C C R , S a l i g n , K t r a n s l a t i o n , P r e t u r n , Q M ) . \Theta_{RF}
=
(
ESR,
CCR,
S_{\mathrm{align}},
K_{\mathrm{translation}},
P_{\mathrm{return}},
Q_M
). Θ R F = ( E S R , C C R , S align , K translation , P return , Q M ) .
如果:
E S R ↑ , C C R ↑ , K t r a n s l a t i o n ↓ , Q M ↑ , ESR\uparrow,
CCR\uparrow,
K_{\mathrm{translation}}\downarrow,
Q_M\uparrow, E S R ↑ , C C R ↑ , K translation ↓ , Q M ↑ ,
比「random switching」更符合 RFCC。
47. 與 TADC-05 的連接
TADC-05:
H m a c r o ↓ ∧ H m i c r o ↑ . H_{\mathrm{macro}}\downarrow
\land
H_{\mathrm{micro}}\uparrow. H macro ↓ ∧ H micro ↑ .
TADC-06 現在補:
即使 micro topics 跨 external disciplines:
H e x t − d o m a i n ↑ , H_{\mathrm{ext-domain}}\uparrow, H ext − domain ↑ ,
只要:
C C R ↑ CCR\uparrow C C R ↑
仍可屬於同一 higher-order cognitive complex。
因此:
domain hyperfocus \boxed{
\text{domain hyperfocus}
} domain hyperfocus
不必等於:
discipline hyperfocus . \boxed{
\text{discipline hyperfocus}.
} discipline hyperfocus .
48. Null Model 1:Taxonomy-Only
假設:
K s w i t c h = f ( d e x t ) . K_{\mathrm{switch}}
=
f(
d_{\mathrm{ext}}
). K switch = f ( d ext ) .
如果外部學科距離已穩定預測:
RT;
error;
transfer;
re-entry;
則 RFCC 沒必要。
49. Null Model 2:Surface Similarity
假設:
K s w i t c h = f ( 1 − s s u r f a c e ) . K_{\mathrm{switch}}
=
f(
1-s_{\mathrm{surface}}
). K switch = f ( 1 − s surface ) .
如果普通 lexical / perceptual / semantic similarity 已完整解釋,
不用 structural relation。
50. Null Model 3:Fixed Semantic Embedding
假設存在:
Z s e m Z_{\mathrm{sem}} Z sem
固定。
只需 embedding cosine distance:
d Z ( x , y ) d_Z(x,y) d Z ( x , y )
即可。
如果:
d r e l d_{\mathrm{rel}} d rel
無法提供增量,
multiplex relational topology 沒必要。
51. Null Model 4:Ordinary Analogy
也可能 TADC-06 只是:
analogy theory renamed . \boxed{
\text{analogy theory renamed}.
} analogy theory renamed .
這是一個真正風險。
若 RFCC 只在 explicit analogy tasks 成立,
而無法預測:
spontaneous topic transition;
switching cost;
domain return;
research problem navigation;
那它應被限制為 analogy subtheory。
52. Null Model 5:Expertise Only
專家跨域流暢,
可能只是:
more knowledge . \text{more knowledge}. more knowledge .
如果控制:
expertise , familiarity , retrieval fluency \text{expertise},
\text{familiarity},
\text{retrieval fluency} expertise , familiarity , retrieval fluency
後,
relation distance 不再有預測力,
RFCC 被削弱。
53. Null Model 6:Random Association
高 branching:
ν s w i t c h ↑ \nu_{\mathrm{switch}}\uparrow ν switch ↑
可能只是 diffuse association。
如果 transitions:
d r e l d_{\mathrm{rel}} d rel
並不比隨機 baseline 短,
則:
C C R CCR C C R
不成立。
54. 實驗一:Crossed Distance Matrix
建立四類 pairs:
same-domain / relationally near;
same-domain / relationally far;
cross-domain / relationally near;
cross-domain / relationally far。
控制:
familiarity;
word frequency;
surface similarity;
task difficulty。
測:
R T , A c c u r a c y , T r a n s f e r , M e m o r y . RT,
Accuracy,
Transfer,
Memory. R T , A cc u r a cy , T r an s f er , M e m or y .
55. 關鍵比較
RFCC 預測:
K ( cross-domain, relationally near ) < K ( same-domain, relationally far ) . K(
\text{cross-domain, relationally near}
)
<
K(
\text{same-domain, relationally far}
). K ( cross-domain, relationally near ) < K ( same-domain, relationally far ) .
如果反覆成立,
就是非常強的證據:
d r e l > d e x t in predictive importance . \boxed{
d_{\mathrm{rel}}
>
d_{\mathrm{ext}}
\text{ in predictive importance}.
} d rel > d ext in predictive importance .
56. 實驗二:Bridge Induction
先測:
d r e l p r e ( U A , U B ) . d_{\mathrm{rel}}^{pre}
(
U_A,U_B
). d rel p r e ( U A , U B ) .
再教一個 shared relational schema:
r ∗ . r^*. r ∗ .
之後測:
d r e l p o s t . d_{\mathrm{rel}}^{post}. d rel p os t .
RRC 預測:
d r e l p o s t < d r e l p r e . d_{\mathrm{rel}}^{post}
<
d_{\mathrm{rel}}^{pre}. d rel p os t < d rel p r e .
57. 實驗三:False Bridge
建立 surface-similar 但 structurally invalid pair。
若 relation-first cognition 成熟,
經 constraint feedback 後應:
d r e l ↑ d_{\mathrm{rel}}\uparrow d rel ↑
或:
Q M ↓ . Q_M\downarrow. Q M ↓ .
這直接測:
Detachment of false analogy . \boxed{
\text{Detachment of false analogy}.
} Detachment of false analogy .
58. 實驗四:Re-representation
提供兩個 relations:
r A , r B r_A,r_B r A , r B
初始不容易對齊。
經 analogy task 後測:
predicate interpretation;
relation categorization;
transfer;
change detection。
若 representations 本身改變:
R t → R t + 1 , \mathcal R_t
\rightarrow
\mathcal R_{t+1}, R t → R t + 1 ,
支持 RRC。
59. 實驗五:Spontaneous Cross-Domain Navigation
不提示 analogy。
給 open-ended problem solving task。
記錄:
x 1 → x 2 → ⋯ x_1\rightarrow x_2\rightarrow\cdots x 1 → x 2 → ⋯
並事後建立:
external labels;
semantic distance;
structural distance;
causal distance;
goal relevance。
測:
K s w i t c h K_{\mathrm{switch}} K switch
到底最受哪個 distance 預測。
60. 實驗六:Return Path
如果:
x i x_i x i
跨到:
y j y_j y j
再返回:
x i + k , x_{i+k}, x i + k ,
測:
P r e t u r n P_{\mathrm{return}} P return
是否被 structural bridge 預測。
RFCC 預測:
S a l i g n ↑ ⇒ P r e t u r n ↑ . S_{\mathrm{align}}\uparrow
\Rightarrow
P_{\mathrm{return}}\uparrow. S align ↑⇒ P return ↑ .
61. 實驗七:Multi-Atlas Task
同一 object set:
X X X
先按:
A ( d i s c i p l i n a r y ) \mathfrak A^{(disciplinary)} A ( d i sc i pl ina r y )
操作,
再按:
A ( c a u s a l ) \mathfrak A^{(causal)} A ( c a u s a l )
或:
A ( f o r m a l ) \mathfrak A^{(formal)} A ( f or ma l )
操作。
測:
performance;
switch cost;
novel inference;
neural representational geometry。
如果某 goal 下:
A ( c a u s a l ) \mathfrak A^{(causal)} A ( c a u s a l )
明顯勝過:
A ( d i s c i p l i n a r y ) , \mathfrak A^{(disciplinary)}, A ( d i sc i pl ina r y ) ,
支持 multiple-valid-atlas model。
62. 實驗八:Schema Drift
對同一 relational schema:
S S S
反覆應用到不同 domains:
D 1 , D 2 , … , D n . D_1,D_2,\ldots,D_n. D 1 , D 2 , … , D n .
追蹤:
S 0 → S 1 → ⋯ . S_0\rightarrow S_1\rightarrow\cdots. S 0 → S 1 → ⋯ .
測其:
boundary;
prototype;
inference pattern;
transfer errors。
如果 schema 會因應用 history 改變,
支持:
relation-space plasticity . \boxed{
\text{relation-space plasticity}.
} relation-space plasticity .
63. 八個核心可證偽命題
TADC6-H1 — Relational Distance Predicts Switching
控制 external taxonomy 後:
d r e l d_{\mathrm{rel}} d rel
仍預測:
K s w i t c h . K_{\mathrm{switch}}. K switch .
TADC6-H2 — Cross-Domain Near Can Beat Same-Domain Far
存在:
K C N < K S F , K_{CN}
<
K_{SF}, K C N < K S F ,
其中:
(CN):cross-domain relationally near;
(SF):same-domain relationally far。
TADC6-H3 — Bridge Learning Compresses Distance
d r e l p o s t < d r e l p r e . d_{\mathrm{rel}}^{post}
<
d_{\mathrm{rel}}^{pre}. d rel p os t < d rel p r e .
TADC6-H4 — False Analogy Can Be Detached
constraint feedback 後:
Q M ↓ Q_M\downarrow Q M ↓
且 erroneous transfer 減少。
TADC6-H5 — Relational Re-representation
analogical alignment 後:
R t ≠ R t + 1 \mathcal R_t
\neq
\mathcal R_{t+1} R t = R t + 1
具有獨立 behavioral signature。
TADC6-H6 — High ESR Can Coexist with High CCR
存在 trajectories:
E S R ↑ ∧ C C R ↑ . ESR\uparrow
\land
CCR\uparrow. E S R ↑ ∧ C C R ↑ .
TADC6-H7 — Multiple Atlases Have Goal-Specific Value
不同:
A i \mathfrak A_i A i
對不同:
G j G_j G j
具有不同 predictive utility。
TADC6-H8 — Relation Model Adds Prediction
multiplex relational model:
M R M_R M R
必須在 out-of-sample prediction 上優於:
taxonomy;
surface similarity;
semantic embedding;
familiarity;
模型。
64. 什麼會殺掉 RFCC / CDCC?
F1 — Taxonomy Wins
若:
d e x t d_{\mathrm{ext}} d ext
穩定優於:
d r e l , d_{\mathrm{rel}}, d rel ,
RFCC 強版失敗。
F2 — Semantic Embedding Is Enough
若:
d Z d_Z d Z
完全吸收 structural relation effect,
multiplex model 過度複雜。
F3 — Relation Effect Disappears after Familiarity Control
若:
expertise / familiarity \text{expertise / familiarity} expertise / familiarity
解釋全部效果,
relation-first 不需要。
F4 — No High-ESR / High-CCR Trajectories
若 external domain switching 高時,
relational continuity 必然低,
CDCC 被否定。
F5 — Re-representation Is Not Needed
若 alignment 只是在固定 representation 中匹配,
RRC 強版失敗。
F6 — Schema Drift Has No Functional Consequence
若 schema representation 改變,
但不影響 future inference / transfer,
distance-plasticity 說法被削弱。
F7 — Relation-First Only Works in Explicit Analogy Tasks
則本篇應縮減為 analogy theory extension,
不能宣稱一般 cognitive-domain continuity。
65. Relation-First 的最小模型
令:
G t = ( V , E ( 1 ) , … , E ( m ) ) . \mathcal G_t
=
(V,E^{(1)},\ldots,E^{(m)}). G t = ( V , E ( 1 ) , … , E ( m ) ) .
goal-conditioned weighting:
w t ( G ) = ( w 1 , … , w m ) . \mathbf w_t(G)
=
(
w_1,\ldots,w_m
). w t ( G ) = ( w 1 , … , w m ) .
有效 graph:
G t G = ∑ r w r ( G ) E ( r ) . \mathcal G_t^G
=
\sum_r
w_r(G)
E^{(r)}. G t G = r ∑ w r ( G ) E ( r ) .
認知距離:
d r e l G = d ( G t G ) . d_{\mathrm{rel}}^G
=
d(
\mathcal G_t^G
). d rel G = d ( G t G ) .
因此:
goal changes ⇒ effective topology changes . \boxed{
\text{goal changes}
\Rightarrow
\text{effective topology changes}.
} goal changes ⇒ effective topology changes .
66. Cross-Domain Continuity 的最小判準
對 trajectory:
γ . \gamma. γ .
若:
E S R ( γ ) ≥ θ E ESR(\gamma)\geq\theta_E E S R ( γ ) ≥ θ E
且:
C C R ( γ ) ≥ θ C , CCR(\gamma)\geq\theta_C, C C R ( γ ) ≥ θ C ,
則稱:
surface-cross-domain but relationally continuous trajectory . \boxed{
\text{surface-cross-domain but relationally continuous trajectory}.
} surface-cross-domain but relationally continuous trajectory .
這只是描述性分類。
要升級成認知機制,
仍需:
C C R CCR C C R
預測 behavior / neural dynamics。
67. 與 TADC-03 六算子的統一
Relation-first cognition 可以直接映射到:
Expansion
發現新 source domain:
E . E. E .
Traversal
沿 relational path:
T . T. T .
Gluing
建立 analogy bridge:
G . G. G .
Detachment
移除 invalid transfer:
D . D. D .
Re-indexing
把兩個具體 domains 壓成:
r ∗ r^* r ∗
的 abstract schema:
R − . R^-. R − .
再把 schema 套回新 domain:
R + . R^+. R + .
68. 一條典型跨域推理路徑
x A ⟶ E U B ⟶ G B A B ⟶ R − S ∗ ⟶ R + U C ⟶ D S v a l i d . x_A
\overset{E}{\longrightarrow}
U_B
\overset{G}{\longrightarrow}
B_{AB}
\overset{R^-}{\longrightarrow}
S^*
\overset{R^+}{\longrightarrow}
U_C
\overset{D}{\longrightarrow}
S_{\mathrm{valid}}. x A ⟶ E U B ⟶ G B A B ⟶ R − S ∗ ⟶ R + U C ⟶ D S valid .
這比:
A → B → C A\rightarrow B\rightarrow C A → B → C
的「換領域」描述多出了真正結構內容。
69. 與 TADC-04 的統一
若:
U A , U B , U C U_A,U_B,U_C U A , U B , U C
在 fine scale 是不同 domains,
但:
R − R^- R −
後都變成:
z r ∗ , z_{r^*}, z r ∗ ,
則 coarse scale:
d r e l ( U A , U B ) → 0. d_{\mathrm{rel}}
(
U_A,U_B
)
\rightarrow0. d rel ( U A , U B ) → 0.
因此:
cross-domain distance can collapse under relational coarse-graining . \boxed{
\text{cross-domain distance can collapse
under relational coarse-graining}.
} cross-domain distance can collapse under relational coarse-graining .
這是 TADC-04 scale relativity 在跨域認知上的直接結果。
70. 與 TADC-05 的統一
若 long-focus trajectory:
γ \gamma γ
跨:
D 1 , D 2 , … , D n D_1,D_2,\ldots,D_n D 1 , D 2 , … , D n
但:
C C R ( γ ) ↑ CCR(\gamma)\uparrow C C R ( γ ) ↑
且:
G C ↑ , G_C\uparrow, G C ↑ ,
則這些 external switches 不一定破壞:
Topological Hyperfocus . \boxed{
\text{Topological Hyperfocus}.
} Topological Hyperfocus .
也就是:
disciplinary diversity ≠ attentional fragmentation . \boxed{
\text{disciplinary diversity}
\neq
\text{attentional fragmentation}.
} disciplinary diversity = attentional fragmentation .
71. Cross-Domain Entropy
外部分類 entropy:
H e x t = − ∑ D p ( D ) log p ( D ) . H_{\mathrm{ext}}
=
-\sum_Dp(D)\log p(D). H ext = − D ∑ p ( D ) log p ( D ) .
relation-family entropy:
H R = − ∑ r p ( r ) log p ( r ) . H_R
=
-\sum_rp(r)\log p(r). H R = − r ∑ p ( r ) log p ( r ) .
一條 trajectory 可以:
H e x t ↑ H_{\mathrm{ext}}\uparrow H ext ↑
但:
H R ↓ , H_R\downarrow, H R ↓ ,
如果它一直在追同一種 structure。
例如跨很多 domains,
但都在追:
r ∗ = feedback . r^*
=
\text{feedback}. r ∗ = feedback .
72. 反過來也可能
外部 domain:
D D D
固定,
但 relation families:
r 1 , r 2 , … r_1,r_2,\ldots r 1 , r 2 , …
一直亂跳。
所以:
H e x t ↓ H_{\mathrm{ext}}\downarrow H ext ↓
但:
H R ↑ . H_R\uparrow. H R ↑ .
這再次說明:
discipline entropy ≠ cognitive relation entropy . \boxed{
\text{discipline entropy}
\neq
\text{cognitive relation entropy}.
} discipline entropy = cognitive relation entropy .
73. Relation-Locked Attention
本文提出一個不等於 hyperfocus 的中性描述:
Relation-Locked Attention(RLA) . \boxed{
\text{Relation-Locked Attention(RLA)}.
} Relation-Locked Attention ( RLA ) .
若:
P ( r ∗ ∣ γ ) → 1 , P(r^*\mid\gamma)\rightarrow1, P ( r ∗ ∣ γ ) → 1 ,
即使:
P ( D i ) P(D_i) P ( D i )
分散,
可說 attention 長期鎖定某個 relation family。
這個構念比:
domain lock \text{domain lock} domain lock
更抽象。
74. RLA 與 THF 的差別
THF:
persistent evolving cognitive complex . \boxed{
\text{persistent evolving cognitive complex}.
} persistent evolving cognitive complex .
RLA:
persistent relation family . \boxed{
\text{persistent relation family}.
} persistent relation family .
RLA 可以是 THF 的一個 invariant:
I R ( K t ) = r ∗ . I_R(\mathcal K_t)=r^*. I R ( K t ) = r ∗ .
如果:
r ∗ r^* r ∗
在多次 domain changes 中保持,
它可能就是 TADC-04 所求的 cross-scale invariant 候選之一。
75. 關係不變量
假設:
M i : U i → U i + 1 . M_i:
U_i\rightarrow U_{i+1}. M i : U i → U i + 1 .
如果:
I R ( U i ) = I R ( U i + 1 ) , I_R(U_i)
=
I_R(U_{i+1}), I R ( U i ) = I R ( U i + 1 ) ,
則:
I R I_R I R
是 relational invariant。
例如:
feedback structure \boxed{
\text{feedback structure}
} feedback structure
可能在 biological / computational / social systems 中保持。
但真正合法 transfer 還要保:
constraints . \text{constraints}. constraints .
76. Relation Invariant ≠ Mechanism Identity
兩個 systems 共享:
I R I_R I R
不表示:
mechanism A = mechanism B . \text{mechanism}_A
=
\text{mechanism}_B. mechanism A = mechanism B .
因此:
structural invariance ≠ ontological identity . \boxed{
\text{structural invariance}
\neq
\text{ontological identity}.
} structural invariance = ontological identity .
這是避免跨域理論過度膨脹的重要限制。
77. 可驗證的「領域沒有領域」版本
最強但仍合法的命題不是:
there are no domains . \text{there are no domains}. there are no domains .
而是:
No single external domain partition is assumed to dominate cognitive adjacency for every high-level goal. \boxed{
\text{No single external domain partition
is assumed to dominate cognitive adjacency
for every high-level goal.}
} No single external domain partition is assumed to dominate cognitive adjacency for every high-level goal.
若實驗最後顯示:
Π e x t \Pi^{\mathrm{ext}} Π ext
在所有 task 上都最有預測力,
這句就被否定。
78. 研究方法上的重要後果
若 RFCC 可能成立,
研究 cross-domain cognition 時不能只記:
how many labels changed . \text{how many labels changed}. how many labels changed .
還要記:
relational distance;
mapping quality;
bridge structure;
constraint preservation;
re-representation;
return path;
goal continuity。
否則會把:
relational traversal \boxed{
\text{relational traversal}
} relational traversal
誤判成:
cognitive fragmentation . \boxed{
\text{cognitive fragmentation}.
} cognitive fragmentation .
79. TADC-06 最小預測模型
定義 switch cost:
K i → j = β 0 + β 1 d e x t + β 2 d s e m + β 3 d r e l + β 4 F + β 5 E + ϵ . K_{i\rightarrow j}
=
\beta_0
+
\beta_1d_{\mathrm{ext}}
+
\beta_2d_{\mathrm{sem}}
+
\beta_3d_{\mathrm{rel}}
+
\beta_4F
+
\beta_5E
+
\epsilon. K i → j = β 0 + β 1 d ext + β 2 d sem + β 3 d rel + β 4 F + β 5 E + ϵ .
其中:
(F):familiarity;
(E):expertise。
RFCC 最小要求:
β 3 > 0 \boxed{
\beta_3>0
} β 3 > 0
且加入:
d r e l d_{\mathrm{rel}} d rel
後,
out-of-sample prediction 提升。
80. 強版本預測
更強:
∣ β 3 ∣ > ∣ β 1 ∣ |\beta_3|
>
|\beta_1| ∣ β 3 ∣ > ∣ β 1 ∣
在 cross-domain transfer task 中成立。
但本文不把這個不等式當必要條件。
只要:
Δ Prediction d r e l > 0 \Delta\operatorname{Prediction}_{d_{\mathrm{rel}}}>0 Δ Prediction d rel > 0
就有增量價值。
81. 最小資料集
一個未來 dataset 至少應有:
{ x i , L i , G i , R i j , d s e m , d r e l , t i , K i , Y i } . \{
x_i,
L_i,
G_i,
R_{ij},
d_{\mathrm{sem}},
d_{\mathrm{rel}},
t_i,
K_i,
Y_i
\}. { x i , L i , G i , R ij , d sem , d rel , t i , K i , Y i } .
其中:
cognitive state;
external label;
goal;
relation type;
distances;
timestamp;
switch cost;
outcome。
這樣才能真正驗證:
d e x t d_{\mathrm{ext}} d ext
與:
d r e l d_{\mathrm{rel}} d rel
誰更重要。
82. 不能只靠自我報告
一個 participant 說:
我覺得這兩個領域其實一樣。
不是證據。
必須有:
RT;
error;
transfer;
novel inference;
eye tracking;
transition sequence;
neural representation;
predictive model;
等 independent observables。
因此:
subjective continuity ≠ measured continuity . \boxed{
\text{subjective continuity}
\neq
\text{measured continuity}.
} subjective continuity = measured continuity .
83. 也不能只靠 embedding
LLM embedding:
z ( x ) z(x) z ( x )
可以估 semantic distance,
但:
embedding similarity ≠ human relational distance . \boxed{
\text{embedding similarity}
\neq
\text{human relational distance}.
} embedding similarity = human relational distance .
尤其:
causal;
procedural;
analogical;
goal;
關係可能和一般 semantic embedding 不一致。
embedding 可以是 proxy,
不能直接當 ground truth。
84. 人–AI 情境暫時不在本篇證明
AI 可以:
找 analogy;
補 bridge;
保存 context;
降低 retrieval cost;
建 relational graphs。
這可能大幅改變:
d r e l e f f e c t i v e . d_{\mathrm{rel}}^{\mathrm{effective}}. d rel effective .
但那是 TADC-07 的主題。
TADC-06 先建立:
relation-first cognition \boxed{
\text{relation-first cognition}
} relation-first cognition
本身。
85. 核心理論總結
本文提出:
G t G = ∑ r w r ( G ) E ( r ) \boxed{
\mathcal G_t^G
=
\sum_r
w_r(G)E^{(r)}
} G t G = r ∑ w r ( G ) E ( r )
以及:
d r e l ( x , y ∣ G ) = min γ : x ⇝ y K r e l ( γ ∣ G ) . \boxed{
d_{\mathrm{rel}}(x,y\mid G)
=
\min_{\gamma:x\leadsto y}
K_{\mathrm{rel}}(\gamma\mid G).
} d rel ( x , y ∣ G ) = γ : x ⇝ y min K rel ( γ ∣ G ) .
與外部分類:
d e x t ( x , y ) d_{\mathrm{ext}}(x,y) d ext ( x , y )
比較。
86. 三個根猜想
RFCC
d r e l adds predictive value beyond d e x t . \boxed{
d_{\mathrm{rel}}
\text{ adds predictive value beyond }
d_{\mathrm{ext}}.
} d rel adds predictive value beyond d ext .
CDCC
E S R ↑ ∧ C C R ↑ \boxed{
ESR\uparrow
\land
CCR\uparrow
} E S R ↑ ∧ C C R ↑
的 trajectory 可以存在。
RRC
d r e l p o s t ≠ d r e l p r e \boxed{
d_{\mathrm{rel}}^{post}
\neq
d_{\mathrm{rel}}^{pre}
} d rel p os t = d rel p r e
因 alignment / re-representation / schema change 而發生。
87. 結論
本文將「跨領域」從一個單一類別事件:
D i → D j D_i\rightarrow D_j D i → D j
重寫為兩個彼此可分離的問題:
第一:
external taxonomy changed? \boxed{
\text{external taxonomy changed?}
} external taxonomy changed?
第二:
effective relational state changed discontinuously? \boxed{
\text{effective relational state changed discontinuously?}
} effective relational state changed discontinuously?
兩者不必相同。
因此:
external categorical discontinuity ⇏ internal cognitive discontinuity . \boxed{
\text{external categorical discontinuity}
\not\Rightarrow
\text{internal cognitive discontinuity}.
} external categorical discontinuity ⇒ internal cognitive discontinuity .
本文提出:
Δ c r o s s = d e x t − d r e l \boxed{
\Delta_{\mathrm{cross}}
=
d_{\mathrm{ext}}
-
d_{\mathrm{rel}}
} Δ cross = d ext − d rel
來描述外部分類與內部關係距離的落差。
當:
Δ c r o s s ≫ 0 , \Delta_{\mathrm{cross}}\gg0, Δ cross ≫ 0 ,
外界可能看到:
大幅跨領域。
但 cognition 可能只是:
traversing a short relational path . \boxed{
\text{traversing a short relational path}.
} traversing a short relational path .
現有 analogical reasoning、cross-domain generalization、re-representation、schema drift、semantic control 與 task-space cognitive-map literature 都說明:
surface category 並非唯一有效結構;
relational mapping 可以跨 representation format;
relational representations 可以支援 transfer;
relation representations 本身可能改變;
task goal 可以重配置有效認知/腦狀態。
但這些仍不足以證明:
relation-first topology is the general architecture of human cognition . \boxed{
\text{relation-first topology
is the general architecture of human cognition}.
} relation-first topology is the general architecture of human cognition .
所以本篇接受明確反證:
如果:
external taxonomy;
semantic embedding;
familiarity;
ordinary analogy;
已能完整預測 switching、transfer 與 return,
則 RFCC 不需要。
如果:
E S R ↑ ESR\uparrow E S R ↑
必然伴隨:
C C R ↓ , CCR\downarrow, C C R ↓ ,
則 CDCC 失敗。
如果 relational representations 在 cross-domain use 後沒有功能性塑形,
RRC 強版本失敗。
因此 TADC-06 的最終命題不是:
「領域不存在。」
而是:
No single fixed domain partition should be assumed to define cognitive adjacency before relational structure is measured. \boxed{
\text{No single fixed domain partition
should be assumed to define cognitive adjacency
before relational structure is measured.}
} No single fixed domain partition should be assumed to define cognitive adjacency before relational structure is measured.
中文:
在測量關係結構之前,不應先假定任何單一固定領域分割必然等於認知鄰接結構。
若這個命題成立,
「跨領域能力」就不再只是:
ability to jump far . \boxed{
\text{ability to jump far}.
} ability to jump far .
它可能同時包括:
ability to discover that two apparently distant regions were structurally adjacent all along . \boxed{
\text{ability to discover that two apparently distant regions
were structurally adjacent all along}.
} ability to discover that two apparently distant regions were structurally adjacent all along .
而這正是 TADC-07 的入口:
如果 AI、外部記憶與多智能體系統可以主動保存 context、發現 bridge、建立 relation graph,那它們是否不只是「幫人算得更快」,而是在直接改變有效認知距離與可達拓樸?
參考文獻
Wang X, Krieger-Redwood K, Cui Y, Smallwood J, Du Y, Jefferies E, et al. Macroscale brain states support the control of semantic cognition. Communications Biology . 2024;7:926. doi:10.1038/s42003-024-06630-7.
Tan L, Qiu Y, Qiu L, et al. The medial and lateral orbitofrontal cortex jointly represent the cognitive map of task space. Communications Biology . 2025;8:163. doi:10.1038/s42003-025-07588-w. PMID: 39900714.
Qiu Y, Li H, Liao J, et al. Forming cognitive maps for abstract spaces: the roles of the human hippocampus and orbitofrontal cortex. Communications Biology . 2024;7:517. doi:10.1038/s42003-024-06214-5. PMID: 38693344.
Weinberger AB, Gallagher NM, Warren T, Green AE. Analogical mapping across sensory modalities and evidence for a general analogy factor. Cognition . 2022;223:105029. doi:10.1016/j.cognition.2022.105029. PMID: 35091260.
Kroczek B, Ciechanowska I, Chuderski A. Uncovering the course of analogical mapping using eye tracking. Cognition . 2022;225:105140. doi:10.1016/j.cognition.2022.105140. PMID: 35483161.
Doumas LAA, Puebla G, Martin AE, Hummel JE. A theory of relation learning and cross-domain generalization. Psychological Review . 2022;129(5):999–1041. doi:10.1037/rev0000346. PMID: 35113620.
Evidence of analogical re-representation from a change detection task. Cognition . 2019. doi:10.1016/j.cognition.2019.04.031. PMID: 31075695.
Vagnino R, Walker C. Schema drift: Relational concepts and conceptual change. Cognition . 2026;271:106418. doi:10.1016/j.cognition.2025.106418. PMID: 41494364.
Garvert MM, Dolan RJ, Behrens TEJ. A map of abstract relational knowledge in the human hippocampal-entorhinal cortex. eLife . 2017;6:e17086. doi:10.7554/eLife.17086.
Park SA, Miller DS, Nili H, Ranganath C, Boorman ED. Map Making: Constructing, Combining, and Inferring on Abstract Cognitive Maps. Neuron . 2020;107(6):1226–1238.e8. doi:10.1016/j.neuron.2020.06.030.
Behrens TEJ, Muller TH, Whittington JCR, et al. What Is a Cognitive Map? Organizing Knowledge for Flexible Behavior. Neuron . 2018;100(2):490–509. doi:10.1016/j.neuron.2018.10.002.
Holyoak KJ, Thagard P. The analogical mind. American Psychologist . 1997;52(1):35–44. PMID: 9017931.
Semantic Structural Alignment of Neural Representational Spaces Enables Translation between English and Chinese Words. Journal of Cognitive Neuroscience . 2016. doi:10.1162/jocn_a_01000. PMID: 27315264.
Qiu Y, et al. Dynamic changes in orbitofrontal-hippocampal connectivity linked to cognitive map formation in humans. NeuroImage . 2025;318:121415. PMID: 40780573.
與系列的關係
已完成:
TADC-01:《注意力不是單點選擇——可變認知空間與注意—空間轉換猜想》
TADC-02:《動態認知域——領域作為局部座標圖》
TADC-03:《拓樸注意力六算子——展開、收斂、遍歷、黏合、切離與重索引》
TADC-04:《嵌套注意域與觀察尺度——宏觀/微觀的相對性與多尺度重索引》
TADC-05:《從單點超專注到拓樸超專注——域級持續性、內部高熵遍歷與可控退出》
TADC-06:《關係優先認知與跨域連續性——從學科距離到關係距離的認知拓樸猜想》
下一篇:
TADC-07:《外部認知支架與人—AI 認知拓樸》
後續:
TADC-08:《拓樸注意力的測量、反證與工程化》
狀態: TADC-06 v0.1原始人體/臨床數據: 無理論狀態: 猜想/研究綱領;未經一般性實驗驗證跨域狀態: RFCC / CDCC / RRC 均為待驗證命題拓樸狀態: relational topology 為 goal-conditioned multiplex accessibility 的候選形式,不宣稱已建立嚴格數學拓樸