title 真正的辯論在找哪裡不同:共同域、分歧定位與因果區辨
english_title Real Debate Locates Where Models Diverge: Shared Domains, Divergence Localization, and Causal Discrimination
series AI 辯論學習方法論系列
series_english AI Debate Learning Methodology Series
series_id ADLM
paper_id ADLM-05
author Neo.K
organization EveMissLab
version 0.1.0
status Research Draft / Series Paper
date 2026-10-06
language zh-TW
canonical_source UTF-8 Markdown
math_source_rule inline math only with $...$ ; display math only with $$...$$
真正的辯論在找哪裡不同
共同域、分歧定位與因果區辨
Real Debate Locates Where Models Diverge: Shared Domains, Divergence Localization, and Causal Discrimination
系列: AI 辯論學習方法論系列(ADLM),Paper 05作者: Neo.K機構: EveMissLab版本: v0.1.0日期: 2026-10-06
公司: 一言諾科技有限公司
摘要
許多辯論在表面上呈現為:
M A v s M B . \boxed{
M_A
\quad vs\quad
M_B.
} M A v s M B .
參與者通常將工作重心放在:
找對方漏洞;
強化自己論點;
防守己方弱點;
爭取裁判或觀眾。
但如果辯論的目標是提高世界模型解析度,那真正重要的問題不是:
Who is wrong? \boxed{
\text{Who is wrong?}
} Who is wrong?
而是:
Where do the models first meaningfully diverge? \boxed{
\text{Where do the models first meaningfully diverge?}
} Where do the models first meaningfully diverge?
本文提出:
Locate the Divergence, not the Loser . \boxed{
\text{Locate the Divergence, not the Loser}.
} Locate the Divergence, not the Loser .
並建立一套從表面 disagreement 走向 causal discrimination 的流程:
Frame Alignment → Shared Domain Construction → Argument Decomposition → First Meaningful Divergence → Causal Crux → Discriminating Evidence . \boxed{
\text{Frame Alignment}
\rightarrow
\text{Shared Domain Construction}
\rightarrow
\text{Argument Decomposition}
\rightarrow
\text{First Meaningful Divergence}
\rightarrow
\text{Causal Crux}
\rightarrow
\text{Discriminating Evidence}.
} Frame Alignment → Shared Domain Construction → Argument Decomposition → First Meaningful Divergence → Causal Crux → Discriminating Evidence .
令兩個世界模型為:
M A , M B . M_A,
\quad
M_B. M A , M B .
它們通常不是從第一個節點就完全不同,而是共享大量結構:
s 0 → s 1 → ⋯ → s k . s_0
\rightarrow
s_1
\rightarrow
\cdots
\rightarrow
s_k. s 0 → s 1 → ⋯ → s k .
真正分叉可能只發生於:
s k → T A s A \boxed{
s_k
\xrightarrow{T_A}
s_{A}
} s k T A s A
與:
s k → T B s B . \boxed{
s_k
\xrightarrow{T_B}
s_{B}.
} s k T B s B .
本文將:
Δ ∗ = ( s k , T A , T B ) \boxed{
\Delta^\ast
=
(
s_k,
T_A,
T_B
)
} Δ ∗ = ( s k , T A , T B )
定義為 First Meaningful Divergence 。
辯論若能從「我們整個世界觀都不同」收斂成「我們其實只在這個 transition rule 不同」,即使雙方尚未達成共識,也已經產生實質 epistemic gain。
本文進一步提出 Causal Crux :若一個節點 c ∗ c^\ast c ∗ 的真假或 causal role 改變,足以使至少一方重新評估其終局模型,則:
c ∗ \boxed{
c^\ast
} c ∗
是高價值 crux。
真正高品質辯論接著應尋找:
E ∗ = arg max E D ( P ( E ∣ M A ) , P ( E ∣ M B ) ) \boxed{
E^\ast
=
\arg\max_E
D(
P(E\mid M_A),
P(E\mid M_B)
)
} E ∗ = arg E max D ( P ( E ∣ M A ) , P ( E ∣ M B ))
也就是最能區分兩模型的新證據、實驗、觀察或反事實。
本文因此主張:
Argument → Crux → Discriminating Evidence \boxed{
\text{Argument}
\rightarrow
\text{Crux}
\rightarrow
\text{Discriminating Evidence}
} Argument → Crux → Discriminating Evidence
應取代:
Argument → More Argument → Rhetorical Exhaustion . \boxed{
\text{Argument}
\rightarrow
\text{More Argument}
\rightarrow
\text{Rhetorical Exhaustion}.
} Argument → More Argument → Rhetorical Exhaustion .
本篇作為 ADLM 系列中段核心,將 ADLM-01 的世界模型比較、ADLM-02 的 claim typing、ADLM-03 的 predicate-domain discipline 與 ADLM-04 的 frame / invariant analysis 統一成一套可操作的辯論解析框架,並為 ADLM-06 的制度史與 ADLM-07 的 Green Model 建立直接前置。
關鍵詞: 共同域、分歧定位、Causal Crux、Discriminating Evidence、Shared Domain、World Model、Debate Resolution、Argument Mapping、因果區辨、AI Debate Learning
1. 問題起點:很多「世界觀衝突」其實沒有那麼大
兩個人可能說:
我們整個世界觀完全不同。
但真正展開後,常發現:
80 % shared + 20 % divergent . \boxed{
80\%\text{ shared}
+
20\%\text{ divergent}.
} 80% shared + 20% divergent .
甚至真正 decisive divergence 可能只剩:
1 – 3 causal transitions . \boxed{
1\text{–}3
\text{ causal transitions}.
} 1 – 3 causal transitions .
2. 表面分歧與結構分歧
本文區分:
Δ s u r f a c e \boxed{
\Delta_{\mathrm{surface}}
} Δ surface
與:
Δ s t r u c t . \boxed{
\Delta_{\mathrm{struct}}.
} Δ struct .
表面不同不一定是結構不同。
3. 表面分歧可能來自語言
例如:
A 使用:
Cause . \text{Cause}. Cause .
B 使用:
Ground . \text{Ground}. Ground .
但若兩者在實際模型中指向相同 relation,
則:
Δ s u r f a c e > 0 \boxed{
\Delta_{\mathrm{surface}}>0
} Δ surface > 0
但:
Δ s t r u c t ≈ 0. \boxed{
\Delta_{\mathrm{struct}}\approx0.
} Δ struct ≈ 0.
4. 也可能來自 frame
ADLM-04 已建立:
O A = Π A ( W ) , O_A=\Pi_A(W), O A = Π A ( W ) ,
O B = Π B ( W ) . O_B=\Pi_B(W). O B = Π B ( W ) .
若存在合法:
T A → B , T_{A\rightarrow B}, T A → B ,
那麼不同描述可能只是:
coordinate difference . \boxed{
\text{coordinate difference}.
} coordinate difference .
不是 substantive disagreement。
5. 真正的分歧
本文定義 substantive divergence:
Δ S ( M A , M B ) > 0 \boxed{
\Delta_S(M_A,M_B)>0
} Δ S ( M A , M B ) > 0
當且僅當經過:
semantic alignment;
frame alignment;
scope alignment;
後,模型仍在:
causal transition;
evidence interpretation;
modal structure;
value function;
上存在不可消去差異。
6. Shared Domain
本文定義:
D s h a r e d \boxed{
D_{\mathrm{shared}}
} D shared
為雙方可共同操作的最小世界模型底座。
至少包含:
D s h a r e d = ( V , R , O , I , S ) . D_{\mathrm{shared}}
=
(
V,
R,
O,
I,
S
). D shared = ( V , R , O , I , S ) .
其中:
V V V :shared variables;
R R R :shared relations;
O O O :shared observations;
I I I :shared inference rules;
S S S :shared scope / semantics。
7. Shared Domain 不是共識
重要:
D s h a r e d ≠ Consensus . \boxed{
D_{\mathrm{shared}}
\neq
\text{Consensus}.
} D shared = Consensus .
雙方可以強烈不同意結論,
但仍共享:
問題是什麼;
哪些觀察是真的;
某些符號代表什麼;
哪些推理規則可接受。
8. Shared Domain 也不是妥協
它不是:
大家各退一步。
而是:
construct a common coordinate system for disagreement . \boxed{
\text{construct a common coordinate system for disagreement}.
} construct a common coordinate system for disagreement .
9. 沒有共同域,辯論會退化
若:
D s h a r e d ≈ ∅ , D_{\mathrm{shared}}\approx\varnothing, D shared ≈ ∅ ,
雙方容易出現:
semantic cross-fire . \boxed{
\text{semantic cross-fire}.
} semantic cross-fire .
也就是:
每一句都像在反駁,但其實沒有作用在同一命題。
10. Shared Object First
辯論至少要先確認:
Are we talking about the same object? \boxed{
\text{Are we talking about the same object?}
} Are we talking about the same object?
否則:
P A ( x ) P_A(x) P A ( x )
與:
P B ( y ) P_B(y) P B ( y )
根本不能直接比較。
11. Shared Scope
同樣:
S c o p e A = S c o p e B \boxed{
Scope_A
=
Scope_B
} S co p e A = S co p e B
至少要能對齊。
ADLM-02 已指出:
local ≠ global . \text{local}
\neq
\text{global}. local = global .
如果一方談:
D l o c a l , D_{\mathrm{local}}, D local ,
另一方談:
Ω , \Omega, Ω ,
衝突可能是假的。
12. Shared Predicate Typing
ADLM-03 已指出 predicate 必須 well-typed。
所以:
T y p e ( P A ) ≈ T y p e ( P B ) \boxed{
Type(P_A)
\approx
Type(P_B)
} T y p e ( P A ) ≈ T y p e ( P B )
是進入 substantive debate 的前置。
13. Shared Frame
ADLM-04 已指出:
F A , F B F_A,F_B F A , F B
若不同,
先尋找:
T A → B . T_{A\rightarrow B}. T A → B .
因此真正辯論前:
semantic alignment + scope alignment + frame alignment \boxed{
\text{semantic alignment}
+
\text{scope alignment}
+
\text{frame alignment}
} semantic alignment + scope alignment + frame alignment
應先完成。
14. Debate Alignment Layer
本文將這一前處理稱:
A D = Debate Alignment Layer . \boxed{
\mathcal A_D
=
\text{Debate Alignment Layer}.
} A D = Debate Alignment Layer .
其輸出:
D s h a r e d . \boxed{
D_{\mathrm{shared}}.
} D shared .
15. 共同域越大越好嗎?
不一定。
若為了「和諧」把:
D s h a r e d D_{\mathrm{shared}} D shared
強行擴大,
可能產生:
false consensus . \boxed{
\text{false consensus}.
} false consensus .
所以共同域只收錄:
actually shared structures . \boxed{
\text{actually shared structures}.
} actually shared structures .
16. Honest Shared Domain
本文提出:
D s h a r e d h o n e s t \boxed{
D_{\mathrm{shared}}^{\mathrm{honest}}
} D shared honest
只包含:
明確接受;
可互相重建;
沒有 hidden disagreement;
的結構。
17. Agreement Verification
A 說:
我同意你的定義。
還不夠。
可以要求 A 重建 B:
D ^ B ( A ) . \hat D_B^{(A)}. D ^ B ( A ) .
若:
F ( D ^ B ( A ) , D B ) > τ , F(
\hat D_B^{(A)},
D_B
)
>\tau, F ( D ^ B ( A ) , D B ) > τ ,
才視為真正 semantic agreement。
18. Reciprocal Reconstruction
同樣:
D ^ A ( B ) \hat D_A^{(B)} D ^ A ( B )
也要通過。
這是:
reciprocal model reconstruction . \boxed{
\text{reciprocal model reconstruction}.
} reciprocal model reconstruction .
19. Shared Domain Fidelity
定義:
F D = F ( D ^ B ( A ) , D B ) + F ( D ^ A ( B ) , D A ) 2 . \boxed{
F_D
=
\frac{
F(
\hat D_B^{(A)},D_B
)
+
F(
\hat D_A^{(B)},D_A
)
}{2}.
} F D = 2 F ( D ^ B ( A ) , D B ) + F ( D ^ A ( B ) , D A ) .
若:
F D ≪ 1 , F_D\ll1, F D ≪ 1 ,
辯論應退回 alignment。
20. 不要急著 attack
傳統 debate:
Position → Attack . \text{Position}
\rightarrow
\text{Attack}. Position → Attack .
本文建議:
Position → Alignment → Decomposition → Attack . \boxed{
\text{Position}
\rightarrow
\text{Alignment}
\rightarrow
\text{Decomposition}
\rightarrow
\text{Attack}.
} Position → Alignment → Decomposition → Attack .
21. Argument Decomposition
將紅方模型:
M A M_A M A
拆成:
M A = ( P A , O A , I A , C A , V A ) . \boxed{
M_A=
(
P_A,
O_A,
I_A,
C_A,
V_A
).
} M A = ( P A , O A , I A , C A , V A ) .
其中:
P A P_A P A :premises;
O A O_A O A :observations;
I A I_A I A :inference;
C A C_A C A :causal relations;
V A V_A V A :value / objective。
22. 藍方也一樣
M B = ( P B , O B , I B , C B , V B ) . \boxed{
M_B=
(
P_B,
O_B,
I_B,
C_B,
V_B
).
} M B = ( P B , O B , I B , C B , V B ) .
然後做 component-wise comparison。
23. Debate Difference Vector
定義:
Δ A B = ( Δ P , Δ O , Δ I , Δ C , Δ V ) . \boxed{
\Delta_{AB}
=
(
\Delta_P,
\Delta_O,
\Delta_I,
\Delta_C,
\Delta_V
).
} Δ A B = ( Δ P , Δ O , Δ I , Δ C , Δ V ) .
24. Premise Divergence
Δ P \Delta_P Δ P
表示:
雙方前提不同。
這通常需要:
evidence;
definition;
prior;
檢查。
25. Observation Divergence
Δ O \Delta_O Δ O
表示:
雙方看到/接受的資料不同。
此時應優先:
data reconciliation . \boxed{
\text{data reconciliation}.
} data reconciliation .
不是繼續純推理。
26. Inference Divergence
若:
O A = O B , O_A=O_B, O A = O B ,
但:
I A ( O ) ≠ I B ( O ) , I_A(O)\neq I_B(O), I A ( O ) = I B ( O ) ,
則:
Δ I > 0. \Delta_I>0. Δ I > 0.
這是真正 reasoning disagreement。
27. Causal Divergence
若雙方接受同一 observation,
但:
C A : X → Y , C_A:
X\rightarrow Y, C A : X → Y ,
C B : Z → Y , C_B:
Z\rightarrow Y, C B : Z → Y ,
則:
Δ C > 0. \Delta_C>0. Δ C > 0.
需要 causal discrimination。
28. Value Divergence
若:
M A f a c t = M B f a c t M_A^{\mathrm{fact}}
=
M_B^{\mathrm{fact}} M A fact = M B fact
但:
V A ≠ V B , V_A\neq V_B, V A = V B ,
則:
debate has crossed into normative space . \boxed{
\text{debate has crossed into normative space}.
} debate has crossed into normative space .
不能繼續假裝只是 facts。
29. First Meaningful Divergence
本文提出核心結構:
Δ ∗ \boxed{
\Delta^\ast
} Δ ∗
即:
經過 semantic / scope / frame alignment 後,模型第一次產生足以影響後續結論的真正分叉。
30. Path Form
若 shared path:
s 0 → s 1 → ⋯ → s k , s_0
\rightarrow
s_1
\rightarrow
\cdots
\rightarrow
s_k, s 0 → s 1 → ⋯ → s k ,
之後:
s k → T A s A s_k
\xrightarrow{T_A}
s_A s k T A s A
與:
s k → T B s B , s_k
\xrightarrow{T_B}
s_B, s k T B s B ,
則:
Δ ∗ = ( s k , T A , T B ) . \boxed{
\Delta^\ast
=
(
s_k,
T_A,
T_B
).
} Δ ∗ = ( s k , T A , T B ) .
31. 為什麼 First Meaningful Divergence 很重要?
因為下游可能有:
100 100 100
個差異,
但都由:
Δ ∗ \Delta^\ast Δ ∗
產生。
若只攻擊 downstream symptoms,
辯論會非常冗長。
32. Divergence Compression
原本:
{ Δ 1 , … , Δ 100 } \{
\Delta_1,\ldots,\Delta_{100}
\} { Δ 1 , … , Δ 100 }
若可壓成:
Δ ∗ \boxed{
\Delta^\ast
} Δ ∗
就得到:
debate compression . \boxed{
\text{debate compression}.
} debate compression .
33. Root Divergence
本文也稱:
Δ r o o t \boxed{
\Delta_{\mathrm{root}}
} Δ root
為能生成大量 downstream disagreement 的根分歧。
34. Root ≠ Earliest Sentence
注意:
earliest verbal disagreement ≠ root divergence . \boxed{
\text{earliest verbal disagreement}
\neq
\text{root divergence}.
} earliest verbal disagreement = root divergence .
最早吵起來的句子可能只是表面症狀。
35. Causal Root Divergence
真正 root 應滿足:
Change ( Δ ∗ ) ⇒ many downstream differences collapse or change . \boxed{
\operatorname{Change}(\Delta^\ast)
\Rightarrow
\text{many downstream differences collapse or change}.
} Change ( Δ ∗ ) ⇒ many downstream differences collapse or change .
36. Sensitivity Test
對候選 divergence:
d i , d_i, d i ,
做:
do ( d i ′ ) . \operatorname{do}(d_i'). do ( d i ′ ) .
若大量 downstream conclusion:
C j C_j C j
隨之改變,
則:
d i d_i d i
可能是高價值 root。
37. Debate Sensitivity
定義:
S D ( d i ) = ∣ downstream conclusions changed ∣ ∣ total relevant conclusions ∣ . \boxed{
S_D(d_i)
=
\frac{
|\text{downstream conclusions changed}|
}{
|\text{total relevant conclusions}|
}.
} S D ( d i ) = ∣ total relevant conclusions ∣ ∣ downstream conclusions changed ∣ .
高:
S D S_D S D
代表高價值 crux candidate。
38. Crux
本文定義一般 crux:
c \boxed{
c
} c
為:
如果其 truth-value / causal role 改變,至少一方會顯著更新模型的命題或關係。
39. Causal Crux
若:
c ∗ c^\ast c ∗
直接位於 causal graph:
G C , G_C, G C ,
且改變它會影響 model prediction,
則:
c ∗ = Causal Crux . \boxed{
c^\ast
=
\text{Causal Crux}.
} c ∗ = Causal Crux .
40. Crux Is Not Merely a Weak Point
弱點:
w w w
可能只是 rhetoric vulnerability。
Crux:
c ∗ c^\ast c ∗
則是:
model-changing dependency . \boxed{
\text{model-changing dependency}.
} model-changing dependency .
41. Crux Value
定義:
V C ( c ) = P ( model update ∣ c resolved ) × impact . \boxed{
V_C(c)
=
P(
\text{model update}
\mid
c\text{ resolved}
)
\times
\text{impact}.
} V C ( c ) = P ( model update ∣ c resolved ) × impact .
42. High-Value Crux
高價值:
V C ( c ) ≫ 0. V_C(c)\gg0. V C ( c ) ≫ 0.
這意味:
解決這一個點,整場辯論就會大幅收斂。
43. Low-Value Argument
有些辯論花大量時間在:
wording;
side examples;
historical trivia;
即使解決也:
V C ≈ 0. V_C\approx0. V C ≈ 0.
這就是低收益 argument。
44. Debate Efficiency
定義:
η C = ∑ V C ( c i ) debate cost . \boxed{
\eta_C
=
\frac{
\sum V_C(c_i)
}{
\text{debate cost}
}.
} η C = debate cost ∑ V C ( c i ) .
高品質辯論應提高:
η C . \eta_C. η C .
45. Crux Search
本文提出:
S C \boxed{
\mathcal S_C
} S C
即:
在 argument graph 中主動搜尋最大 model-changing crux。
46. Crux Tree
可以建立:
T C \boxed{
T_C
} T C
每個節點是:
premise;
causal relation;
prediction;
value assumption。
從 outcome 向上追 dependency。
47. Backward Crux Tracing
從結論:
Y Y Y
反推:
Y ← C 1 ← C 2 ← P . Y
\leftarrow
C_1
\leftarrow
C_2
\leftarrow
P. Y ← C 1 ← C 2 ← P .
比較 A、B 的 path,
找 first divergence。
48. Forward Crux Testing
找到:
c ∗ c^\ast c ∗
後,
改變它:
c ∗ → c ′ c^\ast\rightarrow c' c ∗ → c ′
重新 propagate。
看:
M A , M B M_A,
M_B M A , M B
是否收斂。
49. Crux Agreement Test
問 A:
如果 c ∗ c^\ast c ∗ 是假的,你會改變結論嗎?
問 B:
如果 c ∗ c^\ast c ∗ 是真的,你會改變結論嗎?
若雙方都回答 yes,
則:
c ∗ \boxed{
c^\ast
} c ∗
非常適合作為 shared crux。
50. One-Sided Crux
有時只有:
A A A
會更新,
B 不會。
這仍有價值,
但不是:
double crux . \boxed{
\text{double crux}.
} double crux .
51. Double Crux
若:
c ∗ \boxed{
c^\ast
} c ∗
對 A、B 都是 decisive,
則:
c ∗ = Double Crux . \boxed{
c^\ast
=
\text{Double Crux}.
} c ∗ = Double Crux .
52. Crux ≠ Evidence
Crux 是:
which proposition matters . \boxed{
\text{which proposition matters}.
} which proposition matters .
Evidence 是:
what could resolve that proposition . \boxed{
\text{what could resolve that proposition}.
} what could resolve that proposition .
兩者必須分開。
53. Discriminating Evidence
對兩模型:
M A , M B , M_A,
M_B, M A , M B ,
理想 evidence:
E ∗ E^\ast E ∗
應使:
P ( E ∗ ∣ M A ) \boxed{
P(E^\ast\mid M_A)
} P ( E ∗ ∣ M A )
與:
P ( E ∗ ∣ M B ) \boxed{
P(E^\ast\mid M_B)
} P ( E ∗ ∣ M B )
差異最大。
54. Evidence Discrimination Score
定義:
D E ( E ) = d ( P ( E ∣ M A ) , P ( E ∣ M B ) ) . \boxed{
D_E(E)
=
d(
P(E\mid M_A),
P(E\mid M_B)
).
} D E ( E ) = d ( P ( E ∣ M A ) , P ( E ∣ M B )) .
55. Optimal Evidence
E ∗ = arg max E D E ( E ) . \boxed{
E^\ast
=
\arg\max_E
D_E(E).
} E ∗ = arg E max D E ( E ) .
這是辯論走向 experiment 的核心。
56. 不是所有 evidence 都有區辨力
如果:
P ( E ∣ M A ) ≈ P ( E ∣ M B ) , P(E\mid M_A)
\approx
P(E\mid M_B), P ( E ∣ M A ) ≈ P ( E ∣ M B ) ,
那看到:
E E E
也不會解決爭議。
57. Shared Evidence ≠ Discriminating Evidence
雙方都同意的 evidence:
E s E_s E s
很重要。
但如果:
E s E_s E s
被兩模型都預測,
它只能建立 shared domain。
不能 adjudicate。
58. Debate Should Search for Model-Separating Evidence
因此:
Evidence Volume ≠ Evidence Discrimination . \boxed{
\text{Evidence Volume}
\neq
\text{Evidence Discrimination}.
} Evidence Volume = Evidence Discrimination .
辯論不是比誰丟更多資料。
59. Evidence Quality
可寫:
Q E = F ( D E , R , C , V ) . \boxed{
Q_E
=
F(
D_E,
R,
C,
V
).
} Q E = F ( D E , R , C , V ) .
其中:
D E D_E D E :discrimination;
R R R :reliability;
C C C :cost;
V V V :verifiability。
60. Cheap but Weak Evidence
某 evidence 容易取得,
但:
D E ≈ 0. D_E\approx0. D E ≈ 0.
則對爭議價值很低。
61. Expensive but Decisive Evidence
某實驗成本高,
但:
D E ≫ 0. D_E\gg0. D E ≫ 0.
可能是最值得做的下一步。
62. Debate-to-Experiment Transition
本文提出:
Debate → Experiment \boxed{
\text{Debate}
\rightarrow
\text{Experiment}
} Debate → Experiment
的合法轉換點:
當:
Δ ∗ \Delta^\ast Δ ∗
與:
c ∗ c^\ast c ∗
已足夠清楚,
繼續語言辯論的 marginal gain 低於:
V ( E ∗ ) . V(E^\ast). V ( E ∗ ) .
63. Argument Saturation
定義:
G A ( n ) = new model-relevant gain at debate round n . \boxed{
G_A(n)
=
\text{new model-relevant gain at debate round }n.
} G A ( n ) = new model-relevant gain at debate round n .
若:
G A ( n ) → 0 , G_A(n)\rightarrow0, G A ( n ) → 0 ,
應停止口語攻防。
64. Argument Loop
若:
A 1 → B 1 → A 2 → B 2 A_1\rightarrow B_1\rightarrow A_2\rightarrow B_2 A 1 → B 1 → A 2 → B 2
只是改寫相同內容,
則:
Argument Loop . \boxed{
\text{Argument Loop}.
} Argument Loop .
65. Loop Detection
可用 semantic / structural similarity:
S i m ( A t , A t − k ) > τ . Sim(A_t,A_{t-k})>\tau. S im ( A t , A t − k ) > τ .
若長期成立,
表示 debate saturated。
66. Stop Rule
本文提出:
Stop arguing when no new crux, evidence, or model transition is being generated . \boxed{
\text{Stop arguing when no new crux, evidence, or model transition is being generated}.
} Stop arguing when no new crux, evidence, or model transition is being generated .
67. Debate Does Not Need Immediate Resolution
如果:
E ∗ E^\ast E ∗
尚未可得,
可以輸出:
Unresolved Pending Evidence . \boxed{
\text{Unresolved Pending Evidence}.
} Unresolved Pending Evidence .
68. Unresolved Is a High-Quality State
因為它保留:
shared domain;
crux;
evidence requirement。
比:
forced verdict \boxed{
\text{forced verdict}
} forced verdict
更有知識價值。
69. Evidence Ledger
本文建議辯論維護:
L E \boxed{
L_E
} L E
記錄:
evidence_id:
supports:
weakens:
shared_or_discriminating:
source:
reliability:
scope:
status:
70. Crux Ledger
同樣:
L C \boxed{
L_C
} L C
記錄:
crux_id:
statement:
type:
who_updates_if_true:
who_updates_if_false:
downstream_impact:
required_evidence:
status:
71. Shared Domain Ledger
L D \boxed{
L_D
} L D
記錄:
term:
definition:
scope:
frame:
accepted_by_A:
accepted_by_B:
confidence:
72. Debate State
完整外部化:
S D = ( M A , M B , D s h a r e d , Δ ∗ , C ∗ , E ∗ , H ) . \boxed{
S_D
=
(
M_A,
M_B,
D_{\mathrm{shared}},
\Delta^\ast,
C^\ast,
E^\ast,
H
).
} S D = ( M A , M B , D shared , Δ ∗ , C ∗ , E ∗ , H ) .
73. Revision History
H H H
保存:
M A ( 0 ) → M A ( 1 ) → ⋯ M_A^{(0)}
\rightarrow
M_A^{(1)}
\rightarrow
\cdots M A ( 0 ) → M A ( 1 ) → ⋯
與:
M B ( 0 ) → M B ( 1 ) → ⋯ M_B^{(0)}
\rightarrow
M_B^{(1)}
\rightarrow
\cdots M B ( 0 ) → M B ( 1 ) → ⋯
這可以直接成為 AI learning trace。
74. Revision Is Not Defeat
如果 A 因 evidence:
E E E
修改:
M A , M_A, M A ,
應記錄:
successful epistemic update . \boxed{
\text{successful epistemic update}.
} successful epistemic update .
不是:
A lost . \boxed{
\text{A lost}.
} A lost .
75. Shared Revision
最好情況可能:
M A → M A ′ , M_A\rightarrow M_A', M A → M A ′ ,
M B → M B ′ . M_B\rightarrow M_B'. M B → M B ′ .
兩者同時更新。
這為 ADLM-07 的 Green Model 鋪路。
76. Model Distance
定義:
d M ( M A , M B ) . \boxed{
d_M(
M_A,M_B
).
} d M ( M A , M B ) .
辯論後:
d M d_M d M
可能下降。
但這不是唯一成功指標。
77. Better Divergence Without Smaller Distance
有時:
d M d_M d M
不變,
但:
R D ↑ . R_D\uparrow. R D ↑ .
即雙方仍不同意,
但已清楚知道差異在哪。
這仍是成功。
78. Debate Resolution Gain
定義:
G R = R D a f t e r − R D b e f o r e . \boxed{
G_R
=
R_D^{\mathrm{after}}
-
R_D^{\mathrm{before}}.
} G R = R D after − R D before .
79. Causal Discrimination Gain
定義:
G C = D E ( E ∗ ) a f t e r − D E ( E ) b e f o r e . \boxed{
G_C
=
D_E(E^\ast)_{\mathrm{after}}
-
D_E(E)_{\mathrm{before}}.
} G C = D E ( E ∗ ) after − D E ( E ) before .
表示是否找到更好的區辨證據。
80. Shared-Domain Gain
定義:
G S = Q ( D s h a r e d a f t e r ) − Q ( D s h a r e d b e f o r e ) . \boxed{
G_S
=
Q(
D_{\mathrm{shared}}^{\mathrm{after}}
)
-
Q(
D_{\mathrm{shared}}^{\mathrm{before}}
).
} G S = Q ( D shared after ) − Q ( D shared before ) .
81. Debate Epistemic Gain
可綜合:
G d e b a t e = α G S + β G R + γ G C + δ G U , \boxed{
G_{\mathrm{debate}}
=
\alpha G_S
+
\beta G_R
+
\gamma G_C
+
\delta G_U,
} G debate = α G S + β G R + γ G C + δ G U ,
其中:
G U G_U G U
是 updateability gain。
82. 辯論可以零共識但高增益
可能:
Consensus = 0 \text{Consensus}=0 Consensus = 0
但:
G d e b a t e ≫ 0. G_{\mathrm{debate}}\gg0. G debate ≫ 0.
這是本文非常重要的結論。
83. 辯論也可以高共識但低增益
如果大家本來就同意,
最後只是互相重複,
則:
Consensus ≈ 1 \text{Consensus}\approx1 Consensus ≈ 1
但:
G d e b a t e ≈ 0. G_{\mathrm{debate}}\approx0. G debate ≈ 0.
84. Consensus Is Not the Objective Function
所以:
Consensus ≠ Epistemic Objective . \boxed{
\text{Consensus}
\neq
\text{Epistemic Objective}.
} Consensus = Epistemic Objective .
85. Shared Model Construction
真正求真型辯論應建立:
M s h a r e d \boxed{
M_{\mathrm{shared}}
} M shared
其中包含:
agreed structures;
disputed structures;
unknown structures。
86. Shared Model Can Contain Disagreement
M s h a r e d M_{\mathrm{shared}} M shared
不是:
one single answer . \boxed{
\text{one single answer}.
} one single answer .
它可以含:
{ H A , H B , ? } \boxed{
\{H_A,H_B,?\}
} { H A , H B , ?}
作為 open branches。
87. Branch-Preserving Model
本文提出:
M b r a n c h \boxed{
M_{\mathrm{branch}}
} M branch
允許:
s k → { T A , T B , T C } . s_k
\rightarrow
\{
T_A,
T_B,
T_C
\}. s k → { T A , T B , T C } .
直到 evidence 消除 branch。
88. 這比假共識更好
如果 evidence 不足,
保留:
multiple live branches \boxed{
\text{multiple live branches}
} multiple live branches
比硬選一邊更可靠。
89. Decision Under Branching
如果必須決策,
則:
A ∗ = arg max A ∑ i P ( M i ∣ E ) U ( A ∣ M i ) . A^\ast
=
\arg\max_A
\sum_i
P(M_i\mid E)
U(A\mid M_i). A ∗ = arg A max i ∑ P ( M i ∣ E ) U ( A ∣ M i ) .
所以 unresolved model 仍可支援行動。
90. Truth-Seeking Does Not Require Decision Paralysis
這再次說明:
uncertainty ≠ inaction . \boxed{
\text{uncertainty}
\neq
\text{inaction}.
} uncertainty = inaction .
91. Argument Graph
未來可將辯論表示成:
G A = ( V A , E A ) . \boxed{
G_A
=
(V_A,E_A).
} G A = ( V A , E A ) .
節點:
claims;
evidence;
assumptions;
causal relations。
邊:
supports;
contradicts;
depends-on;
refines。
92. Shared Argument Graph
A、B 圖合併:
G A B = G A ∪ G B . \boxed{
G_{AB}
=
G_A
\cup
G_B.
} G A B = G A ∪ G B .
然後標:
shared;
contested;
unique。
93. Divergence Cut
在:
G A B G_{AB} G A B
中尋找最小 cut:
C ∗ \boxed{
C^\ast
} C ∗
使去除/翻轉後大量 downstream paths 改變。
這是圖論版本的 crux search。
94. Minimal Disagreement Cut
本文提出候選概念:
Minimal Disagreement Cut . \boxed{
\text{Minimal Disagreement Cut}.
} Minimal Disagreement Cut .
即:
用最少的 contested nodes / edges 表示整場 substantive disagreement。
95. Debate Compression Ratio
定義:
C R D = ∣ raw disagreement statements ∣ ∣ minimal disagreement cut ∣ . \boxed{
CR_D
=
\frac{
|\text{raw disagreement statements}|
}{
|\text{minimal disagreement cut}|
}.
} C R D = ∣ minimal disagreement cut ∣ ∣ raw disagreement statements ∣ .
高:
C R D CR_D C R D
表示可以大幅壓縮。
96. AI 特別適合做這件事
人類工作記憶難以管理:
10 2 − 10 4 10^2-10^4 1 0 2 − 1 0 4
個 argument nodes。
AI 可以:
建 graph;
找 dependency;
聚類重複論點;
找 minimal cut。
這可能是 AI debate 真正比人類強的地方之一。
97. 但 AI 也可能 hallucinate shared domain
如果 AI 為了和諧,
虛構:
D s h a r e d , D_{\mathrm{shared}}, D shared ,
會產生:
Synthetic Consensus Error . \boxed{
\text{Synthetic Consensus Error}.
} Synthetic Consensus Error .
98. Synthetic Consensus Error
定義:
F S C \boxed{
F_{SC}
} F S C
為:
模型把實際仍有分歧的概念,錯誤標記成已共識。
99. Shared-Domain Verification
所以每個 shared item:
d i d_i d i
需要:
A confirms ∧ B confirms . A\text{ confirms}
\land
B\text{ confirms}. A confirms ∧ B confirms .
AI multi-agent 系統可做 reciprocal confirmation。
100. AI 也可能 hallucinate crux
某個看似核心的:
c c c
其實雙方都不會因它更新。
所以必須:
Crux Validation . \boxed{
\text{Crux Validation}.
} Crux Validation .
101. Crux Validation Protocol
對:
c , c, c ,
分別問:
Update A ( c = true/false ) ? \operatorname{Update}_A(c=\text{true/false})? Update A ( c = true/false )?
Update B ( c = true/false ) ? \operatorname{Update}_B(c=\text{true/false})? Update B ( c = true/false )?
若都不更新,
不是 crux。
102. AI Debate Role:Cartographer
本文首次正式引入:
C = Cartographer . \boxed{
C
=
\text{Cartographer}.
} C = Cartographer .
它不站紅藍。
它負責:
hypothesis space;
shared domain;
difference graph;
crux candidates。
103. Cartographer 不得求勝
其 reward 不包含:
P ( side wins ) . P(\text{side wins}). P ( side wins ) .
只包含:
mapping fidelity . \boxed{
\text{mapping fidelity}.
} mapping fidelity .
104. Red / Blue Role
紅藍仍很重要。
它們負責:
strongest adversarial model construction . \boxed{
\text{strongest adversarial model construction}.
} strongest adversarial model construction .
但必須接受:
C C C
的 shared state。
105. Evidence Agent
還可以有:
E = Evidence Agent . \boxed{
E
=
\text{Evidence Agent}.
} E = Evidence Agent .
專門搜尋:
E ∗ . E^\ast. E ∗ .
這讓 debate 不再只靠語言。
106. Verifier
最後:
V = Verifier . \boxed{
V
=
\text{Verifier}.
} V = Verifier .
判定:
source reliability;
experiment result;
formal validity。
107. ADLM-05 Runtime Topology
可寫:
C → R / B → C ∗ → E → V . \boxed{
C
\rightarrow
R/B
\rightarrow
C^\ast
\rightarrow
E
\rightarrow
V.
} C → R / B → C ∗ → E → V .
這還沒有 Green。
Green 會在 ADLM-07 正式加入。
108. Debate Learning Target
AI 不應只學:
produce counterargument . \boxed{
\text{produce counterargument}.
} produce counterargument .
還應學:
compress disagreement into causal cruxes . \boxed{
\text{compress disagreement into causal cruxes}.
} compress disagreement into causal cruxes .
109. Training Example Structure
一筆 ADLM training sample:
model_A:
model_B:
shared_domain:
surface_disagreements:
root_divergence:
validated_crux:
discriminating_evidence:
post_evidence_updates:
110. Better Training Signal
這比:
winner: A
資訊密度高得多。
111. ADLM-05 Reward
可暫定:
R 5 = α F D + β Q Δ ∗ + γ V C + δ D E + ϵ U R − η F S C . \boxed{
R_5
=
\alpha F_D
+
\beta Q_{\Delta^\ast}
+
\gamma V_C
+
\delta D_E
+
\epsilon U_R
-
\eta F_{SC}.
} R 5 = α F D + β Q Δ ∗ + γ V C + δ D E + ϵ U R − η F S C .
其中:
F D F_D F D :shared-domain fidelity;
Q Δ ∗ Q_{\Delta^\ast} Q Δ ∗ :divergence localization quality;
V C V_C V C :crux value;
D E D_E D E :evidence discrimination;
U R U_R U R :revision quality;
F S C F_{SC} F S C :false consensus penalty。
112. 第一原則:Shared-Domain Principle
Substantive debate should begin only after the object, scope, predicates, and frames are sufficiently aligned . \boxed{
\text{Substantive debate should begin only after the object, scope, predicates, and frames are sufficiently aligned}.
} Substantive debate should begin only after the object, scope, predicates, and frames are sufficiently aligned .
113. 第二原則:Divergence Localization Principle
The goal is to locate the smallest model-changing divergence, not maximize the number of objections . \boxed{
\text{The goal is to locate the smallest model-changing divergence, not maximize the number of objections}.
} The goal is to locate the smallest model-changing divergence, not maximize the number of objections .
114. 第三原則:Crux Principle
A high-value debate point is one whose resolution causes meaningful model revision . \boxed{
\text{A high-value debate point is one whose resolution causes meaningful model revision}.
} A high-value debate point is one whose resolution causes meaningful model revision .
115. 第四原則:Evidence Discrimination Principle
The best next evidence is not necessarily the largest body of evidence, but the evidence that most strongly separates competing models . \boxed{
\text{The best next evidence is not necessarily the largest body of evidence, but the evidence that most strongly separates competing models}.
} The best next evidence is not necessarily the largest body of evidence, but the evidence that most strongly separates competing models .
116. 第五原則:Argument Saturation Principle
When no new crux or discriminating evidence is generated, continued linguistic debate has diminishing epistemic value . \boxed{
\text{When no new crux or discriminating evidence is generated, continued linguistic debate has diminishing epistemic value}.
} When no new crux or discriminating evidence is generated, continued linguistic debate has diminishing epistemic value .
117. 第六原則:Unresolved-Legitimacy Principle
An explicitly unresolved branch is a legitimate and often superior output to forced consensus . \boxed{
\text{An explicitly unresolved branch is a legitimate and often superior output to forced consensus}.
} An explicitly unresolved branch is a legitimate and often superior output to forced consensus .
118. 第七原則:Revision Principle
Model revision in response to validated evidence is a success state . \boxed{
\text{Model revision in response to validated evidence is a success state}.
} Model revision in response to validated evidence is a success state .
119. 第八原則:Debate-as-Compression Principle
High-quality debate compresses many surface disagreements into a small set of causally decisive divergences . \boxed{
\text{High-quality debate compresses many surface disagreements into a small set of causally decisive divergences}.
} High-quality debate compresses many surface disagreements into a small set of causally decisive divergences .
120. 與 ADLM-01 的接口
ADLM-01:
Locate the Divergence, not the Loser . \boxed{
\text{Locate the Divergence, not the Loser}.
} Locate the Divergence, not the Loser .
本篇正式給出:
Δ ∗ \boxed{
\Delta^\ast
} Δ ∗
與:
c ∗ . \boxed{
c^\ast.
} c ∗ .
121. 與 ADLM-02 的接口
ADLM-02 處理:
claim scope , \text{claim scope}, claim scope ,
所以:
D s h a r e d D_{\mathrm{shared}} D shared
必須先做 quantifier / scope alignment。
122. 與 ADLM-03 的接口
ADLM-03 處理 predicate typing。
所以:
Δ ∗ \Delta^\ast Δ ∗
不能只是 ill-typed language conflict。
123. 與 ADLM-04 的接口
ADLM-04 處理 frame alignment。
所以:
Δ ∗ \Delta^\ast Δ ∗
必須在 coordinate differences 被消除後才合法。
124. 下一篇接口
ADLM-06 將研究:
為什麼人類明明知道求真與求勝不同,辯論制度卻常常仍獎勵求勝? \boxed{
\text{為什麼人類明明知道求真與求勝不同,辯論制度卻常常仍獎勵求勝?}
} 為什麼人類明明知道求真與求勝不同,辯論制度卻常常仍獎勵求勝?
並追溯:
Greek dialectic;
Nyāya vāda / jalpa / vitaṇḍā;
medieval disputatio;
competitive debate;
Structured Academic Controversy;
Double Crux;
adversarial collaboration。
下一篇:
《紅方、藍方與不存在的綠方:人類辯論制度為什麼獎勵求勝》
125. 本文不主張事項
本文不主張:
所有 disagreement 都能縮成一個 crux;
所有辯論都需要共同域很大;
所有 frame 都可轉換;
所有 causal disputes 都能實驗;
所有價值分歧都能用 evidence 解決;
consensus 是錯的;
unresolved 狀態永遠更好;
argument attack 沒有價值;
AI 一定能正確找 crux;
shared model 等於單一結論。
本文只主張:
truth-seeking debate should explicitly locate shared structure, decisive divergence, and discriminating evidence . \boxed{
\text{truth-seeking debate should explicitly locate shared structure, decisive divergence, and discriminating evidence}.
} truth-seeking debate should explicitly locate shared structure, decisive divergence, and discriminating evidence .
126. 結論
真正高品質的辯論,不應把時間主要花在:
How many weaknesses can I find in your position? \boxed{
\text{How many weaknesses can I find in your position?}
} How many weaknesses can I find in your position?
而應該逐步問:
What do we already share? \boxed{
\text{What do we already share?}
} What do we already share?
Where do our models first truly diverge? \boxed{
\text{Where do our models first truly diverge?}
} Where do our models first truly diverge?
Which divergence actually changes the downstream world model? \boxed{
\text{Which divergence actually changes the downstream world model?}
} Which divergence actually changes the downstream world model?
What evidence could resolve it? \boxed{
\text{What evidence could resolve it?}
} What evidence could resolve it?
兩個人可能一開始說:
我們整個世界觀都不一樣。
最後發現:
M A = S + T A , \boxed{
M_A
=
S
+
T_A,
} M A = S + T A ,
M B = S + T B , \boxed{
M_B
=
S
+
T_B,
} M B = S + T B ,
其中:
S S S
是龐大的 shared structure。
真正問題只有:
T A vs T B . \boxed{
T_A
\quad \text{vs}\quad
T_B.
} T A vs T B .
一旦找到這裡,辯論從:
position war \boxed{
\text{position war}
} position war
轉成:
causal discrimination problem . \boxed{
\text{causal discrimination problem}.
} causal discrimination problem .
這時真正有價值的下一句已經不是:
你錯了。
而是:
哪一個觀察、實驗、反事實或驗證,可以讓這兩個 transition rule 真正分開? \boxed{
\text{哪一個觀察、實驗、反事實或驗證,可以讓這兩個 transition rule 真正分開?}
} 哪一個觀察、實驗、反事實或驗證,可以讓這兩個 transition rule 真正分開?
如果找得到:
E ∗ , E^\ast, E ∗ ,
辯論就可以回到世界。
如果找不到,
也至少知道:
current evidence underdetermines the models . \boxed{
\text{current evidence underdetermines the models}.
} current evidence underdetermines the models .
這仍然是認識進步。
因此,求真型辯論最重要的產品並不是 winner。
而是:
D s h a r e d + Δ ∗ + c ∗ + E ∗ . \boxed{
D_{\mathrm{shared}}
+
\Delta^\ast
+
c^\ast
+
E^\ast.
} D shared + Δ ∗ + c ∗ + E ∗ .
也就是:
共同域、真正分歧、因果關鍵點、下一個能區分世界的證據。
到了這一步,辯論才不再只是論證的術。
它開始成為:
world-model debugging . \boxed{
\text{world-model debugging}.
} world-model debugging .
這就是 ADLM-05。
版本紀錄
v0.1.0 — 2026-10-06
建立 ADLM 系列 Paper 05;
定義 Shared Domain 與 Debate Alignment Layer;
提出 Reciprocal Model Reconstruction 與 Shared Domain Fidelity;
建立 Debate Difference Vector;
定義 First Meaningful Divergence;
提出 Divergence Compression 與 Root Divergence;
定義 Causal Crux、Crux Value 與 Crux Validation;
建立 Discriminating Evidence 與 Evidence Discrimination Score;
提出 Debate-to-Experiment Transition;
建立 Argument Saturation 與 Stop Rule;
提出 Shared / Crux / Evidence Ledgers;
定義 Debate Epistemic Gain;
提出 Branch-Preserving Model;
提出 Minimal Disagreement Cut 與 Debate Compression Ratio;
正式引入 Cartographer、Evidence Agent 與 Verifier;
建立八項核心原則;
為 ADLM-06 的辯論制度史與 ADLM-07 的 Green Model 建立直接前置。