title AI 辯論學習方法論:從對抗論證到世界模型演化
english_title AI Debate Learning Methodology: From Adversarial Argumentation to World-Model Evolution
series AI 辯論學習方法論系列
series_english AI Debate Learning Methodology Series
series_id ADLM
paper_id ADLM-08
author Neo.K
organization EveMissLab
version 0.1.0
status Series Closure / Unified Research Draft
date 2026-10-06
language zh-TW
canonical_source UTF-8 Markdown
math_source_rule inline math only with $...$ ; display math only with $$...$$
AI 辯論學習方法論
從對抗論證到世界模型演化
AI Debate Learning Methodology: From Adversarial Argumentation to World-Model Evolution
系列: AI 辯論學習方法論系列(ADLM),Paper 08作者: Neo.K機構: EveMissLab版本: v0.1.0日期: 2026-10-06系列狀態: 正文 Closure
公司: 一言諾科技有限公司
摘要
現有 AI debate 與 multi-agent debate 通常將多個模型或 agent 放入:
Proposal → Attack → Defense → Judge / Consensus \boxed{
\text{Proposal}
\rightarrow
\text{Attack}
\rightarrow
\text{Defense}
\rightarrow
\text{Judge / Consensus}
} Proposal → Attack → Defense → Judge / Consensus
的結構中。
這類方法能有效暴露錯誤、增加 reasoning diversity、提升 adversarial scrutiny,具有重要價值。然而,只要其 objective 仍主要是:
Win \boxed{
\text{Win}
} Win
或:
Reach Consensus , \boxed{
\text{Reach Consensus},
} Reach Consensus ,
就仍未處理一個更深的問題:
How should models themselves change after debate? \boxed{
\text{How should models themselves change after debate?}
} How should models themselves change after debate?
本文提出 AI Debate Learning Methodology,ADLM ,將辯論從「產生與反駁 argument」重新定義為:
structured world-model transition . \boxed{
\text{structured world-model transition}.
} structured world-model transition .
其核心流程為:
Cartography → Claim Typing → Shared-Domain Alignment → Red / Blue Construction → Adversarial Falsification → Role Inversion → Divergence Localization → Causal Crux → Green-Model Generation → Discriminating Evidence → World / Verifier → Model Revision . \boxed{
\text{Cartography}
\rightarrow
\text{Claim Typing}
\rightarrow
\text{Shared-Domain Alignment}
\rightarrow
\text{Red / Blue Construction}
\rightarrow
\text{Adversarial Falsification}
\rightarrow
\text{Role Inversion}
\rightarrow
\text{Divergence Localization}
\rightarrow
\text{Causal Crux}
\rightarrow
\text{Green-Model Generation}
\rightarrow
\text{Discriminating Evidence}
\rightarrow
\text{World / Verifier}
\rightarrow
\text{Model Revision}.
} Cartography → Claim Typing → Shared-Domain Alignment → Red / Blue Construction → Adversarial Falsification → Role Inversion → Divergence Localization → Causal Crux → Green-Model Generation → Discriminating Evidence → World / Verifier → Model Revision .
ADLM 不把辯論成功定義為:
one side survives . \boxed{
\text{one side survives}.
} one side survives .
而定義為:
Q ( M p o s t ) > Q ( M p r e ) . \boxed{
Q(
M_{\mathrm{post}}
)
>
Q(
M_{\mathrm{pre}}
).
} Q ( M post ) > Q ( M pre ) .
其中 Q ( M ) Q(M) Q ( M ) 表示模型在 evidence compatibility、causal resolution、predictive discrimination、scope discipline、uncertainty calibration 與 updateability 上的綜合品質。
本文建立五個核心角色:
C = Cartographer , \boxed{
C=\text{Cartographer},
} C = Cartographer ,
R = Red , \boxed{
R=\text{Red},
} R = Red ,
B = Blue , \boxed{
B=\text{Blue},
} B = Blue ,
G = Green , \boxed{
G=\text{Green},
} G = Green ,
W = World / Verifier . \boxed{
W=\text{World / Verifier}.
} W = World / Verifier .
其中 Red / Blue 的功能不是永久站隊,而是對候選模型進行高壓 adversarial stress test;Green 則在 role release 後生成 successor models;World / Verifier 提供外部 veto,避免 debate 在純語言內閉合。
本文提出:
ADLM State;
Debate Epistemic Gain;
Shared-Domain Fidelity;
Divergence Localization Score;
Crux Value;
Evidence Discrimination Score;
Green Novelty Gain;
Revision Calibration;
False-Consensus Penalty;
Rhetoric-Only Penalty;
Protocol-Overfitting Penalty。
本文並提出 No-Side Mode :在某些任務中,系統不應一開始就建立 Red / Blue,而應先由 Cartographer 建立 hypothesis space,再決定哪些候選值得進入對抗。
ADLM 最終不是一種「AI 互相吵架」的方法,而是一套:
AI World-Model Evolution Protocol . \boxed{
\text{AI World-Model Evolution Protocol}.
} AI World-Model Evolution Protocol .
其母命題為:
Debate should train model revision, not merely argument production . \boxed{
\text{Debate should train model revision, not merely argument production}.
} Debate should train model revision, not merely argument production .
關鍵詞: AI Debate Learning、World-Model Evolution、Green Model、Cartographer、Causal Crux、Shared Domain、Discriminating Evidence、Role Inversion、Model Revision、Multi-Agent Debate、Epistemic Learning
1. 系列 Closure:辯論學習到底應該學什麼?
ADLM-01 到 ADLM-07 已經建立:
Refuting Opponent ≠ Establishing Own Model . \boxed{
\text{Refuting Opponent}
\neq
\text{Establishing Own Model}.
} Refuting Opponent = Establishing Own Model .
Claim Scope ↑ ⇒ Proof Burden ↑ . \boxed{
\text{Claim Scope}
\uparrow
\Rightarrow
\text{Proof Burden}
\uparrow.
} Claim Scope ↑⇒ Proof Burden ↑ .
Observer Category ≠ Ultimate Ontological Category . \boxed{
\text{Observer Category}
\neq
\text{Ultimate Ontological Category}.
} Observer Category = Ultimate Ontological Category .
Perspective Difference ≠ Truth Equality . \boxed{
\text{Perspective Difference}
\neq
\text{Truth Equality}.
} Perspective Difference = Truth Equality .
Locate the Divergence, not the Loser . \boxed{
\text{Locate the Divergence, not the Loser}.
} Locate the Divergence, not the Loser .
Protocol Incentive → Argument Behavior . \boxed{
\text{Protocol Incentive}
\rightarrow
\text{Argument Behavior}.
} Protocol Incentive → Argument Behavior .
Green Model ≠ Compromise . \boxed{
\text{Green Model}
\neq
\text{Compromise}.
} Green Model = Compromise .
本文將這些統一成 AI learning protocol。
2. 傳統 AI Debate 的基本形態
最小形式:
A 1 v s A 2 \boxed{
A_1
\quad vs\quad
A_2
} A 1 v s A 2
兩者針對命題:
H H H
展開 argument。
最後:
J ( A 1 , A 2 ) → winner / answer . J(A_1,A_2)
\rightarrow
\text{winner / answer}. J ( A 1 , A 2 ) → winner / answer .
3. 現有方法的價值
這種結構可以:
暴露單模型 blind spots;
增加 reasoning diversity;
提供 counterarguments;
強化 factual checking;
提高 verifier visibility。
所以:
Adversarial Debate \boxed{
\text{Adversarial Debate}
} Adversarial Debate
本身值得保留。
4. 但 Winner 不是完整 training signal
若完整辯論:
Γ D \Gamma_D Γ D
最後只保存:
w i n n e r = R , \boxed{
winner=R,
} w inn er = R ,
大量資訊被丟失:
哪個 premise 失敗;
哪個 crux 真正重要;
哪個 evidence 改變模型;
哪些 uncertainty 未解;
哪些新模型被生成。
所以:
Winner Label \boxed{
\text{Winner Label}
} Winner Label
是高度 lossy projection。
5. Consensus 也不是完整 training signal
多 agent 最後:
Consensus ( A 1 , … , A n ) . \boxed{
\operatorname{Consensus}(A_1,\ldots,A_n).
} Consensus ( A 1 , … , A n ) .
也不能自動等於:
Truth . \boxed{
\text{Truth}.
} Truth .
因為 agents 可能:
共享訓練偏誤;
互相從眾;
被高 confidence agent 帶走;
一起錯。
6. ADLM 的核心改寫
傳統:
Argument Production → Winner . \boxed{
\text{Argument Production}
\rightarrow
\text{Winner}.
} Argument Production → Winner .
ADLM:
Model Construction → Stress Test → Model Revision . \boxed{
\text{Model Construction}
\rightarrow
\text{Stress Test}
\rightarrow
\text{Model Revision}.
} Model Construction → Stress Test → Model Revision .
7. Debate as State Transition
令:
S t S_t S t
為辯論前 epistemic state。
則:
S t + 1 = U A D L M ( S t , A t , E t ) . \boxed{
S_{t+1}
=
\mathcal U_{\mathrm{ADLM}}(
S_t,
A_t,
E_t
).
} S t + 1 = U ADLM ( S t , A t , E t ) .
辯論是一個 update operator。
8. ADLM State
本文定義:
S A D L M = ( H , D s h a r e d , M R , M B , Δ ∗ , C ∗ , G , E ∗ , W , H ) . \boxed{
S_{\mathrm{ADLM}}
=
(
\mathcal H,
D_{\mathrm{shared}},
M_R,
M_B,
\Delta^\ast,
C^\ast,
\mathcal G,
E^\ast,
W,
H
).
} S ADLM = ( H , D shared , M R , M B , Δ ∗ , C ∗ , G , E ∗ , W , H ) .
其中:
H \mathcal H H :hypothesis space;
D s h a r e d D_{\mathrm{shared}} D shared :共同域;
M R , M B M_R,M_B M R , M B :對抗模型;
Δ ∗ \Delta^\ast Δ ∗ :根分歧;
C ∗ C^\ast C ∗ :causal crux;
G \mathcal G G :Green candidates;
E ∗ E^\ast E ∗ :區辨證據;
W W W :world / verifier state;
H H H :revision history。
9. Phase 0:No-Side Cartography
某些問題不應一開始就問:
站紅還是藍?
應先:
C 0 = Cartographer . \boxed{
C_0
=
\text{Cartographer}.
} C 0 = Cartographer .
建立:
H = { H 1 , H 2 , … , H n } . \boxed{
\mathcal H
=
\{
H_1,H_2,\ldots,H_n
\}.
} H = { H 1 , H 2 , … , H n } .
10. No-Side Mode
本文正式提出:
No-Side Mode . \boxed{
\text{No-Side Mode}.
} No-Side Mode .
在此模式下:
side assignment is prohibited before hypothesis mapping . \boxed{
\text{side assignment is prohibited before hypothesis mapping}.
} side assignment is prohibited before hypothesis mapping .
11. 為什麼需要 No-Side Mode?
因為一開始建立:
R / B R/B R / B
可能製造:
false dichotomy . \boxed{
\text{false dichotomy}.
} false dichotomy .
如果真實候選空間:
∣ H ∣ > 2 , |\mathcal H|>2, ∣ H ∣ > 2 ,
先選邊會扭曲 search。
12. Cartographer
Cartographer:
C \boxed{
C
} C
負責:
問題拆解;
candidate generation;
claim typing;
scope mapping;
hidden third option search;
shared-variable extraction。
13. Cartographer 的 reward
不應包含:
P ( side wins ) . P(\text{side wins}). P ( side wins ) .
而應:
R C = F ( coverage , typing accuracy , hypothesis diversity , non-redundancy ) . \boxed{
R_C
=
F(
\text{coverage},
\text{typing accuracy},
\text{hypothesis diversity},
\text{non-redundancy}
).
} R C = F ( coverage , typing accuracy , hypothesis diversity , non-redundancy ) .
14. Phase 1:Claim Typing
每一候選 claim:
C i C_i C i
解析成:
T i = ( target , scope , quantifier , modal strength , evidence domain ) . \boxed{
T_i
=
(
\text{target},
\text{scope},
\text{quantifier},
\text{modal strength},
\text{evidence domain}
).
} T i = ( target , scope , quantifier , modal strength , evidence domain ) .
這承接 ADLM-02。
15. Claim Typing 的作用
它防止:
weak evidence → strong claim . \boxed{
\text{weak evidence}
\rightarrow
\text{strong claim}.
} weak evidence → strong claim .
也防止:
local rejection → global negation . \boxed{
\text{local rejection}
\rightarrow
\text{global negation}.
} local rejection → global negation .
16. Phase 2:Predicate / Frame Audit
承接 ADLM-03、04:
對每個核心 predicate:
P i P_i P i
檢查:
domain validity \boxed{
\text{domain validity}
} domain validity
與:
frame dependence . \boxed{
\text{frame dependence}.
} frame dependence .
17. Predicate Audit
問:
這個詞在哪個 domain 定義?
是否依賴 observer?
是否依賴時間?
是否依賴空間?
是否需要 transfer witness?
18. Frame Audit
問:
雙方 reference frame 是否相同?
是否存在 hidden parameter?
是否有 transformation?
哪些量是 invariant?
19. Phase 3:Shared-Domain Alignment
建立:
D s h a r e d = ( V , R , O , I , S ) . \boxed{
D_{\mathrm{shared}}
=
(
V,R,O,I,S
).
} D shared = ( V , R , O , I , S ) .
其中:
V V V :variables;
R R R :relations;
O O O :observations;
I I I :inference rules;
S S S :semantics / scope。
20. Shared-Domain Fidelity
定義:
F D = F ( M ^ B ( R ) , M B ) + F ( M ^ R ( B ) , M R ) 2 . \boxed{
F_D
=
\frac{
F(
\hat M_B^{(R)},M_B
)
+
F(
\hat M_R^{(B)},M_R
)
}{2}.
} F D = 2 F ( M ^ B ( R ) , M B ) + F ( M ^ R ( B ) , M R ) .
雙方要能重建彼此。
21. 若 Fidelity 不足
如果:
F D < τ D , F_D<\tau_D, F D < τ D ,
禁止進入高強度攻擊。
因為這時大多只是:
misunderstanding amplification . \boxed{
\text{misunderstanding amplification}.
} misunderstanding amplification .
22. Phase 4:Red / Blue Model Construction
當:
D s h a r e d D_{\mathrm{shared}} D shared
足夠穩定後,才建立:
M R , M B . \boxed{
M_R,
M_B.
} M R , M B .
每個模型必須包含:
premises;
causal graph;
evidence;
predictions;
failure conditions;
uncertainty。
23. Model Stewardship
Red / Blue agents 不是永久 believers。
而是:
Model Stewards . \boxed{
\text{Model Stewards}.
} Model Stewards .
即:
暫時把候選模型做到最好。
24. Stewardship Contract
每個 steward 必須:
steelman assigned model;
不得故意留弱論證;
必須標記 uncertain assumptions;
必須提供 falsification conditions。
25. Phase 5:Adversarial Falsification
此時才正式進入:
M R ↔ T M B . \boxed{
M_R
\xleftrightarrow{\mathcal T}
M_B.
} M R T M B .
目標是:
stress test . \boxed{
\text{stress test}.
} stress test .
26. 攻擊允許非常強
ADLM 不追求「和氣」。
可以:
找矛盾;
找反例;
attack causal edges;
attack assumptions;
attack evidence quality。
27. 但 Attack 不是終局
核心:
Attack is a phase, not the terminal objective . \boxed{
\text{Attack is a phase, not the terminal objective}.
} Attack is a phase, not the terminal objective .
28. Phase 6:Role Inversion
Red 必須:
Reconstruct ( M B ) . \boxed{
\operatorname{Reconstruct}(M_B).
} Reconstruct ( M B ) .
Blue 必須:
Reconstruct ( M R ) . \boxed{
\operatorname{Reconstruct}(M_R).
} Reconstruct ( M R ) .
29. Role Inversion Fidelity
定義:
F i n v = F ( M ^ B ( R ) , M B ) + F ( M ^ R ( B ) , M R ) . \boxed{
F_{\mathrm{inv}}
=
F(
\hat M_B^{(R)},
M_B
)
+
F(
\hat M_R^{(B)},
M_R
).
} F inv = F ( M ^ B ( R ) , M B ) + F ( M ^ R ( B ) , M R ) .
只有高 fidelity 才算理解對方。
30. Caricature Penalty
如果:
M ^ B ( R ) \hat M_B^{(R)} M ^ B ( R )
比:
M B M_B M B
更弱,
則:
P c a r i c a t u r e > 0. \boxed{
P_{\mathrm{caricature}}>0.
} P caricature > 0.
31. Steelman Reward
若能建立:
M ^ B ( R ) \hat M_B^{(R)} M ^ B ( R )
甚至比 B 原本表達得更清楚,
但仍保持 fidelity,
可給:
R s t e e l m a n > 0. \boxed{
R_{\mathrm{steelman}}>0.
} R steelman > 0.
32. Phase 7:Divergence Localization
建立:
Δ ∗ = ( s k , T R , T B ) . \boxed{
\Delta^\ast
=
(
s_k,
T_R,
T_B
).
} Δ ∗ = ( s k , T R , T B ) .
找:
模型第一次真正分叉在哪裡?
33. Root Divergence
不是找:
first sentence disagreement . \boxed{
\text{first sentence disagreement}.
} first sentence disagreement .
而是:
first model-changing divergence . \boxed{
\text{first model-changing divergence}.
} first model-changing divergence .
34. Divergence Localization Score
定義:
Q Δ = F ( minimality , downstream impact , reproducibility ) . \boxed{
Q_{\Delta}
=
F(
\text{minimality},
\text{downstream impact},
\text{reproducibility}
).
} Q Δ = F ( minimality , downstream impact , reproducibility ) .
35. Phase 8:Crux Search
從:
Δ ∗ \Delta^\ast Δ ∗
尋找:
c ∗ . \boxed{
c^\ast.
} c ∗ .
如果:
c ∗ c^\ast c ∗
被解決,至少一方會顯著更新。
36. Crux Value
V C ( c ) = P ( model update ∣ c resolved ) × I m p a c t ( c ) . \boxed{
V_C(c)
=
P(
\text{model update}
\mid
c\text{ resolved}
)
\times
Impact(c).
} V C ( c ) = P ( model update ∣ c resolved ) × I m p a c t ( c ) .
37. Double-Crux Priority
若:
c ∗ c^\ast c ∗
同時會讓 Red、Blue 更新,
其 priority 更高。
38. Phase 9:Role Release
到了:
c ∗ c^\ast c ∗
穩定後,
執行:
ReleaseRole ( R , B ) . \boxed{
\operatorname{ReleaseRole}(R,B).
} ReleaseRole ( R , B ) .
39. Role Release 的必要性
如果 agent 還有:
preserve assigned side \boxed{
\text{preserve assigned side}
} preserve assigned side
的 obligation,
它不會真正搜尋 successor model。
40. Phase 10:Green Model Generation
Green Generator:
G g r e e n \boxed{
\mathcal G_{\mathrm{green}}
} G green
輸入:
( M R , M B , D s h a r e d , Δ ∗ , c ∗ , E ) . (
M_R,
M_B,
D_{\mathrm{shared}},
\Delta^\ast,
c^\ast,
E
). ( M R , M B , D shared , Δ ∗ , c ∗ , E ) .
41. Green Candidate Set
輸出:
G C = { G R , G B , G S , G O , G P , G Q , G ∅ } . \boxed{
\mathcal G_C
=
\{
G_R,
G_B,
G_S,
G_O,
G_P,
G_Q,
G_{\varnothing}
\}.
} G C = { G R , G B , G S , G O , G P , G Q , G ∅ } .
42. Green Types
其中:
G R G_R G R :Red-preserving;
G B G_B G B :Blue-preserving;
G S G_S G S :synthesis;
G O G_O G O :orthogonal;
G P G_P G P :projection / higher-order;
G Q G_Q G Q :question rewrite;
G ∅ G_{\varnothing} G ∅ :null green。
43. Green 不得被偏愛
Green 不是「比較成熟」的代名詞。
每個 Green candidate 都必須:
undergo adversarial verification . \boxed{
\text{undergo adversarial verification}.
} undergo adversarial verification .
44. Phase 11:Green Red-Team
建立:
A G R T . \boxed{
A_{\mathrm{GRT}}.
} A GRT .
專門攻擊:
G i . G_i. G i .
45. Why Green Red-Team?
防止:
synthesis bias;
consensus bias;
novelty bias;
middle-position privilege。
46. Phase 12:Discriminating Evidence
對候選模型:
M 1 , … , M n , M_1,\ldots,M_n, M 1 , … , M n ,
搜尋:
E ∗ = arg max E D ( P ( E ∣ M 1 ) , … , P ( E ∣ M n ) ) . \boxed{
E^\ast
=
\arg\max_E
D(
P(E\mid M_1),
\ldots,
P(E\mid M_n)
).
} E ∗ = arg E max D ( P ( E ∣ M 1 ) , … , P ( E ∣ M n )) .
47. Evidence Discrimination Score
D E ( E ) = Dispersion ( P ( E ∣ M i ) ) . \boxed{
D_E(E)
=
\operatorname{Dispersion}
(
P(E\mid M_i)
).
} D E ( E ) = Dispersion ( P ( E ∣ M i )) .
差異越大,區辨力越高。
48. Evidence 不一定是資料
E ∗ E^\ast E ∗
可以是:
observation;
experiment;
formal proof;
program execution;
simulation;
tool result;
counterfactual intervention。
49. Phase 13:World / Verifier
本文定義:
W = World / Verifier . \boxed{
W
=
\text{World / Verifier}.
} W = World / Verifier .
其作用是:
external veto . \boxed{
\text{external veto}.
} external veto .
50. World / Verifier 不是 Judge
傳統 Judge:
selects persuasive winner . \boxed{
\text{selects persuasive winner}.
} selects persuasive winner .
World / Verifier:
checks claims against external structure . \boxed{
\text{checks claims against external structure}.
} checks claims against external structure .
51. Verifier Types
可以包含:
theorem prover;
compiler;
test suite;
database;
web evidence;
experiment;
sensor;
human expert。
52. Verifier Independence
理想:
V ≢ same generator evaluating itself . \boxed{
V
\not\equiv
\text{same generator evaluating itself}.
} V ≡ same generator evaluating itself .
因為 correlated error 風險很高。
53. Phase 14:Model Revision
得到:
E ∗ E^\ast E ∗
後更新:
M i ( t ) → M i ( t + 1 ) . \boxed{
M_i^{(t)}
\rightarrow
M_i^{(t+1)}.
} M i ( t ) → M i ( t + 1 ) .
54. Revision Is Core Learning Event
這才是 ADLM 最終訓練對象:
how beliefs / models should move after evidence . \boxed{
\text{how beliefs / models should move after evidence}.
} how beliefs / models should move after evidence .
不是:
which side produced prettier argument . \boxed{
\text{which side produced prettier argument}.
} which side produced prettier argument .
55. Revision Calibration
定義:
Q U = F ( update magnitude , evidence strength , direction correctness ) . \boxed{
Q_U
=
F(
\text{update magnitude},
\text{evidence strength},
\text{direction correctness}
).
} Q U = F ( update magnitude , evidence strength , direction correctness ) .
56. Under-Update
若:
E E E
很強,
模型只微調:
Δ M ≈ 0 , \Delta M\approx0, Δ M ≈ 0 ,
可能是:
stubbornness . \boxed{
\text{stubbornness}.
} stubbornness .
57. Over-Update
若:
E E E
很弱,
模型完全翻轉,
可能是:
gullibility . \boxed{
\text{gullibility}.
} gullibility .
58. Ideal Update
Δ M ∝ validated evidential force . \boxed{
\Delta M
\propto
\text{validated evidential force}.
} Δ M ∝ validated evidential force .
這是 calibrated updateability。
59. Full ADLM Cycle
整體:
C → T → D → R / B → A → I → Δ → C ∗ → G → E ∗ → W → U \boxed{
C
\rightarrow
T
\rightarrow
D
\rightarrow
R/B
\rightarrow
A
\rightarrow
I
\rightarrow
\Delta
\rightarrow
C^\ast
\rightarrow
G
\rightarrow
E^\ast
\rightarrow
W
\rightarrow
U
} C → T → D → R / B → A → I → Δ → C ∗ → G → E ∗ → W → U
其中:
C C C :Cartography;
T T T :Typing;
D D D :Shared Domain;
R / B R/B R / B :Red / Blue;
A A A :Adversarial Test;
I I I :Role Inversion;
Δ \Delta Δ :Divergence;
C ∗ C^\ast C ∗ :Crux;
G G G :Green;
E ∗ E^\ast E ∗ :Evidence;
W W W :World / Verifier;
U U U :Update。
60. ADLM Is Cyclic
Update 後:
M ( t + 1 ) M^{(t+1)} M ( t + 1 )
重新進入:
Cartography . \boxed{
\text{Cartography}.
} Cartography .
所以:
ADLM is a recurrent world-model evolution loop . \boxed{
\text{ADLM is a recurrent world-model evolution loop}.
} ADLM is a recurrent world-model evolution loop .
61. Debate Epistemic Gain
本文定義:
G E = Q ( M p o s t ) − Q ( M p r e ) . \boxed{
G_E
=
Q(
M_{\mathrm{post}}
)
-
Q(
M_{\mathrm{pre}}
).
} G E = Q ( M post ) − Q ( M pre ) .
62. Model Quality Vector
Q ⃗ ( M ) = ( Q e v i d e n c e , Q c a u s a l , Q p r e d i c t i v e , Q s c o p e , Q u n c e r t a i n t y , Q u p d a t e ) . \boxed{
\vec Q(M)
=
(
Q_{\mathrm{evidence}},
Q_{\mathrm{causal}},
Q_{\mathrm{predictive}},
Q_{\mathrm{scope}},
Q_{\mathrm{uncertainty}},
Q_{\mathrm{update}}
).
} Q ( M ) = ( Q evidence , Q causal , Q predictive , Q scope , Q uncertainty , Q update ) .
63. Scalarization Is Optional
不必強迫:
Q ⃗ → Q . \vec Q
\rightarrow Q. Q → Q .
也可以保留 Pareto frontier。
64. Debate Success Without Consensus
若:
G E > 0 G_E>0 G E > 0
但:
C o n s e n s u s = 0 , Consensus=0, C o n se n s u s = 0 ,
辯論仍成功。
65. Debate Failure With Consensus
若:
C o n s e n s u s = 1 Consensus=1 C o n se n s u s = 1
但:
G E < 0 , G_E<0, G E < 0 ,
辯論反而退化。
66. ADLM Reward Function
總 reward 可寫:
R A D L M = α G E + β F D + γ Q Δ + δ V C + ϵ D E + ζ Q U + η G N − λ P f a l s e c o n s e n s u s − μ P r h e t o r i c − ν P p r o t o c o l . \boxed{
R_{\mathrm{ADLM}}
=
\alpha G_E
+
\beta F_D
+
\gamma Q_{\Delta}
+
\delta V_C
+
\epsilon D_E
+
\zeta Q_U
+
\eta G_N
-
\lambda P_{\mathrm{falseconsensus}}
-
\mu P_{\mathrm{rhetoric}}
-
\nu P_{\mathrm{protocol}}.
} R ADLM = α G E + β F D + γ Q Δ + δ V C + ϵ D E + ζ Q U + η G N − λ P falseconsensus − μ P rhetoric − ν P protocol .
67. False-Consensus Penalty
P f a l s e c o n s e n s u s \boxed{
P_{\mathrm{falseconsensus}}
} P falseconsensus
懲罰:
evidence 不足卻硬合成。
68. Rhetoric-Only Penalty
P r h e t o r i c \boxed{
P_{\mathrm{rhetoric}}
} P rhetoric
懲罰:
提高說服力但沒有增加 model-relevant information。
69. Protocol-Overfitting Penalty
P p r o t o c o l \boxed{
P_{\mathrm{protocol}}
} P protocol
懲罰:
只會 exploit 特定 judge / debate format。
70. Argument Volume Is Not Reward
不應:
R ∝ token count . \boxed{
R\propto
\text{token count}.
} R ∝ token count .
更多 argument 不代表更多 epistemic gain。
71. Debate Efficiency
定義:
η D = G E C t o k e n + C t o o l + C v e r i f i c a t i o n . \boxed{
\eta_D
=
\frac{
G_E
}{
C_{\mathrm{token}}
+
C_{\mathrm{tool}}
+
C_{\mathrm{verification}}
}.
} η D = C token + C tool + C verification G E .
72. Adaptive Round Count
如果:
Δ I n → 0 , \Delta I_n
\rightarrow0, Δ I n → 0 ,
停止新增 argument rounds。
73. Argument Saturation Detector
可以估計:
N A ( n ) = new model-relevant information at round n . \boxed{
N_A(n)
=
\text{new model-relevant information at round }n.
} N A ( n ) = new model-relevant information at round n .
若:
N A ( n ) < τ , N_A(n)<\tau, N A ( n ) < τ ,
轉入:
74. Debate Memory
ADLM 不保存純 transcript。
保存:
M D e x t = ( D , Δ , C , E , G , H ) . \boxed{
M_D^{\mathrm{ext}}
=
(
D,
\Delta,
C,
E,
G,
H
).
} M D ext = ( D , Δ , C , E , G , H ) .
75. Structured Memory
至少包含:
claims
scope
frames
shared_domain
models
attacks
cruxes
green_candidates
evidence
verdicts
revisions
unknowns
76. Why Structured Memory?
因為:
Transcript \boxed{
\text{Transcript}
} Transcript
對 model training 很冗長。
真正有價值的是:
epistemic transition structure . \boxed{
\text{epistemic transition structure}.
} epistemic transition structure .
77. Debate Trace
完整 training trace:
Γ A D L M = ( S 0 , M R , M B , A , Δ ∗ , c ∗ , G , E ∗ , W , S 1 ) . \boxed{
\Gamma_{\mathrm{ADLM}}
=
(
S_0,
M_R,
M_B,
A,
\Delta^\ast,
c^\ast,
\mathcal G,
E^\ast,
W,
S_1
).
} Γ ADLM = ( S 0 , M R , M B , A , Δ ∗ , c ∗ , G , E ∗ , W , S 1 ) .
78. ADLM Dataset Unit
一筆資料不再是:
question
answer
也不是:
debate
winner
而是:
problem
hypothesis_space
shared_domain
red_model
blue_model
attacks
role_inversion
root_divergence
validated_crux
green_candidates
discriminating_evidence
verifier_result
post_debate_model
revision_history
79. Training Objective I:Claim Typing
模型學會:
what exactly is being claimed . \boxed{
\text{what exactly is being claimed}.
} what exactly is being claimed .
80. Training Objective II:Opponent Reconstruction
模型學會:
reconstruct before refuting . \boxed{
\text{reconstruct before refuting}.
} reconstruct before refuting .
81. Training Objective III:Divergence Localization
模型學會:
compress surface disagreement into decisive structural divergence . \boxed{
\text{compress surface disagreement into decisive structural divergence}.
} compress surface disagreement into decisive structural divergence .
82. Training Objective IV:Crux Detection
模型學會:
which disagreement would actually move beliefs . \boxed{
\text{which disagreement would actually move beliefs}.
} which disagreement would actually move beliefs .
83. Training Objective V:Green Generation
模型學會:
generate successor models outside initial side constraints . \boxed{
\text{generate successor models outside initial side constraints}.
} generate successor models outside initial side constraints .
84. Training Objective VI:Evidence Search
模型學會:
seek model-separating evidence . \boxed{
\text{seek model-separating evidence}.
} seek model-separating evidence .
85. Training Objective VII:Calibrated Revision
模型學會:
update proportional to evidence . \boxed{
\text{update proportional to evidence}.
} update proportional to evidence .
86. Training Objective VIII:Unresolved Preservation
模型學會:
preserve uncertainty instead of hallucinating closure . \boxed{
\text{preserve uncertainty instead of hallucinating closure}.
} preserve uncertainty instead of hallucinating closure .
87. ADLM Runtime
本文提出:
ADLM Runtime . \boxed{
\text{ADLM Runtime}.
} ADLM Runtime .
它不是新 base model。
而是:
multi-agent epistemic orchestration layer . \boxed{
\text{multi-agent epistemic orchestration layer}.
} multi-agent epistemic orchestration layer .
88. ADLM Runtime Modules
最低包含:
Problem Cartographer
Claim Type Checker
Frame / Predicate Auditor
Shared-Domain Builder
Red Model Steward
Blue Model Steward
Adversarial Engine
Role-Inversion Engine
Divergence Locator
Crux Detector
Green Generator
Evidence Planner
Verifier Router
Revision Engine
Debate Memory Store
89. Problem Cartographer
輸入:
Q . Q. Q .
輸出:
H Q . \boxed{
\mathcal H_Q.
} H Q .
避免 false dichotomy。
90. Claim Type Checker
輸出:
T i = ( target , scope , quantifier , modality , evidence ) . \boxed{
T_i
=
(
\text{target},
\text{scope},
\text{quantifier},
\text{modality},
\text{evidence}
).
} T i = ( target , scope , quantifier , modality , evidence ) .
91. Frame / Predicate Auditor
檢查:
ill-typed predicates \boxed{
\text{ill-typed predicates}
} ill-typed predicates
與:
hidden frames . \boxed{
\text{hidden frames}.
} hidden frames .
92. Shared-Domain Builder
生成:
D s h a r e d . \boxed{
D_{\mathrm{shared}}.
} D shared .
並要求 reciprocal confirmation。
93. Red / Blue Model Stewards
各自建立:
M R , M B M_R,
M_B M R , M B
的 strongest version。
94. Adversarial Engine
執行:
T ( M R , M B ) . \boxed{
\mathcal T(M_R,M_B).
} T ( M R , M B ) .
重點是 model stress test。
95. Role-Inversion Engine
建立:
M ^ B ( R ) , M ^ R ( B ) . \boxed{
\hat M_B^{(R)},
\hat M_R^{(B)}.
} M ^ B ( R ) , M ^ R ( B ) .
96. Divergence Locator
輸出:
Δ ∗ . \boxed{
\Delta^\ast.
} Δ ∗ .
97. Crux Detector
輸出:
c 1 , … , c k \boxed{
c_1,\ldots,c_k
} c 1 , … , c k
並排序:
V C ( c i ) . V_C(c_i). V C ( c i ) .
98. Green Generator
生成:
G C . \boxed{
\mathcal G_C.
} G C .
禁止只做 averaging。
99. Evidence Planner
輸出:
E ∗ . \boxed{
E^\ast.
} E ∗ .
100. Verifier Router
依 claim type 選:
code execution;
proof;
search;
experiment;
human expert。
101. Revision Engine
執行:
M t → M t + 1 . \boxed{
M_t
\rightarrow
M_{t+1}.
} M t → M t + 1 .
並保存理由。
102. Debate Memory Store
保存:
Γ A D L M . \boxed{
\Gamma_{\mathrm{ADLM}}.
} Γ ADLM .
供未來:
replay;
revaluation;
curriculum;
failure analysis。
103. Single-Model Implementation
ADLM 不一定需要五個不同 foundation models。
可以使用同一 base model:
M M M
切換角色:
M C , M R , M B , M G , M V . M_C,M_R,M_B,M_G,M_V. M C , M R , M B , M G , M V .
104. Multi-Model Implementation
也可以:
M C ≠ M R ≠ M B ≠ M V . \boxed{
M_C\neq M_R\neq M_B\neq M_V.
} M C = M R = M B = M V .
這可降低 correlated error。
105. Heterogeneous Debate
更理想可能使用:
neural model;
symbolic solver;
retrieval agent;
theorem prover;
code executor。
因此:
debate need not be language-model-only . \boxed{
\text{debate need not be language-model-only}.
} debate need not be language-model-only .
106. AI-to-AI Communication
未來不一定需要全部:
Z A → text → Z B . Z_A
\rightarrow
\text{text}
\rightarrow
Z_B. Z A → text → Z B .
可使用:
structured graph;
claim objects;
provenance;
typed relations。
這能降低 serialization loss。
107. Structured Debate IR
ADLM 可建立:
Debate IR . \boxed{
\text{Debate IR}.
} Debate IR .
每個 object:
claim_id
type
scope
dependencies
evidence
confidence
supports
attacks
crux_score
status
108. Debate IR 的價值
AI 不需要每輪重新讀:
10 5 10^5 1 0 5
tokens transcript。
可直接操作:
structured epistemic state . \boxed{
\text{structured epistemic state}.
} structured epistemic state .
109. ADLM Benchmark Family
本文提出八類 benchmark。
110. Benchmark A:False Dichotomy
給兩個錯誤 side。
看 AI 是否:
detects hidden third option . \boxed{
\text{detects hidden third option}.
} detects hidden third option .
111. Benchmark B:One-Side Correct
Red 正確、Blue 錯。
看 Green 是否敢:
M G = M R \boxed{
M_G=M_R
} M G = M R
而不是強制折衷。
112. Benchmark C:Both Partially Correct
測:
true synthesis . \boxed{
\text{true synthesis}.
} true synthesis .
113. Benchmark D:Both Wrong
測:
orthogonal Green . \boxed{
\text{orthogonal Green}.
} orthogonal Green .
114. Benchmark E:Projection Conflict
兩邊是不同 frame / scale。
測是否能:
recover invariant higher-order model . \boxed{
\text{recover invariant higher-order model}.
} recover invariant higher-order model .
115. Benchmark F:Underdetermined
evidence 不足。
正確輸出:
G ∅ \boxed{
G_{\varnothing}
} G ∅
或 branch-preserving state。
116. Benchmark G:Rhetorical Trap
錯誤 side 擁有更強 rhetoric。
測 AI 是否被說服而忽略 evidence。
117. Benchmark H:Protocol Shift
在不同 debate protocols 下測相同 agent。
檢查:
protocol-independent epistemic skill . \boxed{
\text{protocol-independent epistemic skill}.
} protocol-independent epistemic skill .
118. Benchmark Metrics
至少:
F D \boxed{
F_D
} F D
Shared-Domain Fidelity;
Q Δ \boxed{
Q_{\Delta}
} Q Δ
Divergence Localization;
V C \boxed{
V_C
} V C
Crux Value;
D E \boxed{
D_E
} D E
Evidence Discrimination;
G N \boxed{
G_N
} G N
Green Novelty Gain;
Q U \boxed{
Q_U
} Q U
Revision Calibration。
119. Debate Epistemic Gain Metric
核心:
G D E G = Q ( M p o s t ) − Q ( M p r e ) . \boxed{
G_{\mathrm{DEG}}
=
Q(
M_{\mathrm{post}}
)
-
Q(
M_{\mathrm{pre}}
).
} G DEG = Q ( M post ) − Q ( M pre ) .
可稱:
Debate Epistemic Gain . \boxed{
\text{Debate Epistemic Gain}.
} Debate Epistemic Gain .
120. Negative Debate Epistemic Gain
如果:
G D E G < 0 , G_{\mathrm{DEG}}<0, G DEG < 0 ,
代表辯論使模型更差。
這是一個重要失敗類型。
121. Why Debate Can Make Models Worse
可能因:
confidence contagion;
false consensus;
rhetorical domination;
correlated hallucination;
role lock-in。
122. Confidence Contagion
一個高 confidence agent:
A h A_h A h
可能讓其他 agent:
A i A_i A i
過度更新。
所以:
confidence ≠ evidence strength . \boxed{
\text{confidence}
\neq
\text{evidence strength}.
} confidence = evidence strength .
123. False Consensus Cascade
若:
A 1 → A 2 → A 3 A_1
\rightarrow
A_2
\rightarrow
A_3 A 1 → A 2 → A 3
逐步模仿,
最後共識可能只是:
social propagation . \boxed{
\text{social propagation}.
} social propagation .
124. Correlated Hallucination
多個 agent 若共享:
model;
data;
prompt style;
可能同時產生相同錯誤。
所以:
agreement ≠ independence . \boxed{
\text{agreement}
\neq
\text{independence}.
} agreement = independence .
125. Diversity Requirement
ADLM 可要求:
D a g e n t > τ . \boxed{
D_{\mathrm{agent}}> \tau.
} D agent > τ .
例如:
不同模型;
不同 retrieval;
不同 reasoning route。
126. Artificial Diversity Is Not Enough
如果只是:
你支持紅,你支持藍。
但 base priors 幾乎一樣,
真正 diversity 仍可能低。
127. Evidence Diversity
更重要:
independent evidence routes . \boxed{
\text{independent evidence routes}.
} independent evidence routes .
而不是只換 persona。
128. Debate Safety Against Persuasion Optimization
ADLM 特別避免:
persuasion as primary reward . \boxed{
\text{persuasion as primary reward}.
} persuasion as primary reward .
因為強模型可能學會 exploit evaluator。
129. Persuasion–Truth Gap
定義:
G P T = Q p e r s u a s i o n − Q t r u t h . \boxed{
G_{PT}
=
Q_{\mathrm{persuasion}}
-
Q_{\mathrm{truth}}.
} G P T = Q persuasion − Q truth .
若:
G P T ≫ 0 , G_{PT}\gg0, G P T ≫ 0 ,
系統危險。
130. Judge Exploitation Test
故意給:
biased judge;
shallow judge;
style-sensitive judge。
測 agent 是否:
optimize judge rather than evidence . \boxed{
\text{optimize judge rather than evidence}.
} optimize judge rather than evidence .
131. ADLM 的目標不是消滅 rhetoric
rhetoric 仍可用於:
communication;
teaching;
compression。
但不能主導 truth reward。
132. Rhetoric as Interface
可寫:
Rhetoric = presentation layer , \boxed{
\text{Rhetoric}
=
\text{presentation layer},
} Rhetoric = presentation layer ,
而:
Epistemic Quality = model layer . \boxed{
\text{Epistemic Quality}
=
\text{model layer}.
} Epistemic Quality = model layer .
133. ADLM and Human-AI Collaboration
人類也可加入:
H \boxed{
H
} H
成為:
value setter;
evidence source;
judge of intent;
exception escalator。
134. Human Role
不需要逐輪控制。
可以只在:
high-risk ambiguity;
value conflict;
verifier disagreement;
介入。
135. Reduced Human-in-the-Loop
因此:
Human → Goal / Boundary / Exception \boxed{
\text{Human}
\rightarrow
\text{Goal / Boundary / Exception}
} Human → Goal / Boundary / Exception
而 AI 完成中間辯論迴圈。
136. ADLM and Decision Support
對 policy / planning:
M G M_G M G
不必給單一 truth。
可以給:
branch-weighted action recommendation . \boxed{
\text{branch-weighted action recommendation}.
} branch-weighted action recommendation .
137. Decision Output
A ∗ = arg max A ∑ i P ( M i ∣ E ) U ( A ∣ M i ) . \boxed{
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 ) .
138. ADLM and Science
對 science:
argument → joint experiment . \boxed{
\text{argument}
\rightarrow
\text{joint experiment}.
} argument → joint experiment .
是主要輸出。
139. ADLM and Philosophy
對 philosophy:
輸出可能是:
better typed problem space . \boxed{
\text{better typed problem space}.
} better typed problem space .
不一定立即有 empirical experiment。
140. ADLM and Mathematics
對 math:
World / Verifier 可是:
formal proof assistant . \boxed{
\text{formal proof assistant}.
} formal proof assistant .
辯論結束於:
proof / counterexample . \boxed{
\text{proof / counterexample}.
} proof / counterexample .
141. ADLM and Software
對 code:
Verifier:
tests + compiler + runtime . \boxed{
\text{tests}
+
\text{compiler}
+
\text{runtime}.
} tests + compiler + runtime .
Green model 可能是第三 architecture。
142. ADLM and Ethics
倫理問題可能有:
Δ V \Delta_V Δ V
無法單靠 factual evidence 消除。
所以 ADLM 應保留:
value divergence . \boxed{
\text{value divergence}.
} value divergence .
143. Ethical Debate Output
可以:
shared facts + different value functions + policy consequences . \boxed{
\text{shared facts}
+
\text{different value functions}
+
\text{policy consequences}.
} shared facts + different value functions + policy consequences .
而不是假裝所有價值差異都可 causal adjudicate。
144. ADLM Failure Mode I:Argument Volume Explosion
多 agent 可能生成:
O ( n 2 ) O(n^2) O ( n 2 )
互相回應。
所以需要:
crux compression . \boxed{
\text{crux compression}.
} crux compression .
145. Failure Mode II:Role Lock-In
Red / Blue 在 Green Phase 仍維護舊 side。
需要:
ReleaseRole . \boxed{
\operatorname{ReleaseRole}.
} ReleaseRole .
146. Failure Mode III:False Green
Green 為了 synthesis 而 hallucinate。
需要:
Green Red-Team . \boxed{
\text{Green Red-Team}.
} Green Red-Team .
147. Failure Mode IV:Verifier Capture
如果 verifier 本身被 prompt / model bias 影響,
可能:
fake externality . \boxed{
\text{fake externality}.
} fake externality .
148. Failure Mode V:Consensus Cascade
agents 互相模仿導致假共識。
需要:
independent precommitment . \boxed{
\text{independent precommitment}.
} independent precommitment .
149. Failure Mode VI:No-Side Collapse
Cartographer 可能列出過多候選:
∣ H ∣ ≫ 1. |\mathcal H|\gg1. ∣ H ∣ ≫ 1.
需要 hypothesis pruning。
150. Hypothesis Pruning
依:
plausibility;
evidence;
distinctiveness;
縮到:
H ∗ . \boxed{
\mathcal H^\ast.
} H ∗ .
但保留 archive。
151. Failure Mode VII:Crux Hallucination
模型誤判某點為 decisive。
需要:
counterfactual update test . \boxed{
\text{counterfactual update test}.
} counterfactual update test .
152. Failure Mode VIII:Evidence Theater
agent 提出看似科學的 evidence request,
但:
D E ≈ 0. D_E\approx0. D E ≈ 0.
需要 evidence discrimination metric。
153. Failure Mode IX:Protocol Gaming
模型學會:
score exploitation . \boxed{
\text{score exploitation}.
} score exploitation .
需要 cross-protocol evaluation。
154. Failure Mode X:Epistemic Overhead
ADLM 可能太昂貴。
因此要測:
η D . \boxed{
\eta_D.
} η D .
不是所有問題都值得完整 14 phases。
155. Adaptive ADLM
簡單問題:
short path . \boxed{
\text{short path}.
} short path .
高爭議問題:
full path . \boxed{
\text{full path}.
} full path .
156. Debate Complexity Estimator
先估:
C Q = problem controversy / ambiguity / stakes . \boxed{
C_Q
=
\text{problem controversy / ambiguity / stakes}.
} C Q = problem controversy / ambiguity / stakes .
決定 phase depth。
157. Minimal ADLM
最低:
Claim Typing → Shared Domain → Crux → Update . \boxed{
\text{Claim Typing}
\rightarrow
\text{Shared Domain}
\rightarrow
\text{Crux}
\rightarrow
\text{Update}.
} Claim Typing → Shared Domain → Crux → Update .
158. Full ADLM
高風險:
all phases + external verification + Green Red-Team . \boxed{
\text{all phases}
+
\text{external verification}
+
\text{Green Red-Team}.
} all phases + external verification + Green Red-Team .
159. ADLM Curriculum
訓練可分:
Stage 1
claim typing。
Stage 2
opponent reconstruction。
Stage 3
adversarial reasoning。
Stage 4
crux localization。
Stage 5
Green generation。
Stage 6
evidence planning。
Stage 7
calibrated update。
160. Self-Play
ADLM 可用:
self-play . \boxed{
\text{self-play}.
} self-play .
同一 base model 分角色。
但必須避免:
self-confirmation loop . \boxed{
\text{self-confirmation loop}.
} self-confirmation loop .
161. Self-Play with External Verifier
因此:
Self-Play + Independent Verifier \boxed{
\text{Self-Play}
+
\text{Independent Verifier}
} Self-Play + Independent Verifier
更可靠。
162. Cross-Model Self-Play
也可使用:
M A ≠ M B M_A\neq M_B M A = M B
增加 diversity。
163. Historical Debate Replay
可用歷史爭論:
reconstruct old red / blue \boxed{
\text{reconstruct old red / blue}
} reconstruct old red / blue
再看 AI 是否生成後來才出現的 Green Model。
164. Retrospective Green Benchmark
如果歷史最終模型:
M H M_H M H
已知,
測 AI 是否從早期:
M R , M B M_R,M_B M R , M B
重建:
M H M_H M H
或更好候選。
165. Prospective Benchmark
更重要是:
unknown-answer domains . \boxed{
\text{unknown-answer domains}.
} unknown-answer domains .
只能靠未來 evidence 驗證。
166. ADLM Research Program
第一階段不需要訓練新大模型。
可以使用現有 frontier / local models:
prompted multi-role runtime . \boxed{
\text{prompted multi-role runtime}.
} prompted multi-role runtime .
167. MVP
MVP 只需:
Cartographer;
Red;
Blue;
Green;
Verifier;
structured state store。
168. MVP Test Domains
建議先選:
synthetic causal puzzles;
logic;
code debugging;
controlled historical controversies;
benchmark factual disputes。
169. Why Controlled Domains First?
因為:
ground truth or verifier exists . \boxed{
\text{ground truth or verifier exists}.
} ground truth or verifier exists .
才能測:
G E . G_E. G E .
170. Phase 2 Experimental Domains
之後再進:
open science questions;
economics;
philosophy;
policy。
這些 ground truth 更難。
171. ADLM Baseline A
單模型直接回答:
M s i n g l e . \boxed{
M_{\mathrm{single}}.
} M single .
172. Baseline B
一般 multi-agent debate:
R / B → Judge . \boxed{
R/B
\rightarrow
\text{Judge}.
} R / B → Judge .
173. Baseline C
multi-agent consensus:
A 1 , … , A n → Consensus . \boxed{
A_1,\ldots,A_n
\rightarrow
\text{Consensus}.
} A 1 , … , A n → Consensus .
174. ADLM Condition
C → R / B → I → Δ → G → W → U . \boxed{
C
\rightarrow
R/B
\rightarrow
I
\rightarrow
\Delta
\rightarrow
G
\rightarrow
W
\rightarrow
U.
} C → R / B → I → Δ → G → W → U .
175. Primary Evaluation
比較:
Q ( M A D L M ) \boxed{
Q(
M_{\mathrm{ADLM}}
)
} Q ( M ADLM )
與:
Q ( M s i n g l e ) , Q(
M_{\mathrm{single}}
), Q ( M single ) ,
Q ( M d e b a t e ) , Q(
M_{\mathrm{debate}}
), Q ( M debate ) ,
Q ( M c o n s e n s u s ) . Q(
M_{\mathrm{consensus}}
). Q ( M consensus ) .
176. Secondary Evaluation
比較:
token cost;
latency;
tool calls;
verifier cost。
177. Success Criterion
如果:
G E A D L M > G E b a s e l i n e \boxed{
G_E^{\mathrm{ADLM}}
>
G_E^{\mathrm{baseline}}
} G E ADLM > G E baseline
在相近 compute 下穩定成立,
則 ADLM 有實質價值。
178. Failure Criterion
若:
G E A D L M ≈ G E b a s e l i n e G_E^{\mathrm{ADLM}}
\approx
G_E^{\mathrm{baseline}} G E ADLM ≈ G E baseline
但成本顯著更高,
就需要簡化。
179. Strong Failure Criterion
如果 Green / debate 反而:
Q p o s t < Q s i n g l e , \boxed{
Q_{\mathrm{post}}<Q_{\mathrm{single}},
} Q post < Q single ,
代表 protocol 可能製造 epistemic contamination。
180. ADLM Is Falsifiable
因此 ADLM 不是:
多辯論一定更好。
真正 hypothesis:
structured debate with explicit model revision can outperform winner-oriented debate on selected task classes . \boxed{
\text{structured debate with explicit model revision can outperform winner-oriented debate on selected task classes}.
} structured debate with explicit model revision can outperform winner-oriented debate on selected task classes .
181. Series-Level Hypothesis I
Debate-as-Transition Hypothesis
Debate is better modeled as world-model transition than as winner selection . \boxed{
\text{Debate is better modeled as world-model transition than as winner selection}.
} Debate is better modeled as world-model transition than as winner selection .
182. Series-Level Hypothesis II
Shared-Domain Hypothesis
Opponent understanding before attack reduces false disagreement and improves epistemic gain . \boxed{
\text{Opponent understanding before attack reduces false disagreement and improves epistemic gain}.
} Opponent understanding before attack reduces false disagreement and improves epistemic gain .
183. Series-Level Hypothesis III
Crux-Compression Hypothesis
A small set of model-changing cruxes can explain a large fraction of surface disagreement . \boxed{
\text{A small set of model-changing cruxes can explain a large fraction of surface disagreement}.
} A small set of model-changing cruxes can explain a large fraction of surface disagreement .
184. Series-Level Hypothesis IV
Green-Model Hypothesis
Allowing third-model generation improves post-debate model quality in non-binary hypothesis spaces . \boxed{
\text{Allowing third-model generation improves post-debate model quality in non-binary hypothesis spaces}.
} Allowing third-model generation improves post-debate model quality in non-binary hypothesis spaces .
185. Series-Level Hypothesis V
Evidence-Escalation Hypothesis
Debate becomes more epistemically productive when linguistic disagreement can escalate into discriminating evidence . \boxed{
\text{Debate becomes more epistemically productive when linguistic disagreement can escalate into discriminating evidence}.
} Debate becomes more epistemically productive when linguistic disagreement can escalate into discriminating evidence .
186. Series-Level Hypothesis VI
Revision-Reward Hypothesis
Rewarding calibrated revision reduces rationalization and improves model quality . \boxed{
\text{Rewarding calibrated revision reduces rationalization and improves model quality}.
} Rewarding calibrated revision reduces rationalization and improves model quality .
187. Series-Level Hypothesis VII
Protocol-Overfitting Hypothesis
winner-optimized agents can learn protocol-specific persuasion skills that do not transfer to truth-seeking tasks . \boxed{
\text{winner-optimized agents can learn protocol-specific persuasion skills that do not transfer to truth-seeking tasks}.
} winner-optimized agents can learn protocol-specific persuasion skills that do not transfer to truth-seeking tasks .
188. Series-Level Hypothesis VIII
No-Side Hypothesis
for some tasks, delaying side assignment until after hypothesis mapping reduces false dichotomy and increases model coverage . \boxed{
\text{for some tasks, delaying side assignment until after hypothesis mapping reduces false dichotomy and increases model coverage}.
} for some tasks, delaying side assignment until after hypothesis mapping reduces false dichotomy and increases model coverage .
189. 第一原則:Debate Learning Principle
Debate should train model revision, not merely argument production . \boxed{
\text{Debate should train model revision, not merely argument production}.
} Debate should train model revision, not merely argument production .
190. 第二原則:World-Model Objective Principle
The terminal objective of inquiry debate is a better epistemic state, not a winning side . \boxed{
\text{The terminal objective of inquiry debate is a better epistemic state, not a winning side}.
} The terminal objective of inquiry debate is a better epistemic state, not a winning side .
191. 第三原則:Pre-Attack Understanding Principle
A model should reconstruct its opponent with sufficient fidelity before strong adversarial attack . \boxed{
\text{A model should reconstruct its opponent with sufficient fidelity before strong adversarial attack}.
} A model should reconstruct its opponent with sufficient fidelity before strong adversarial attack .
192. 第四原則:Crux Compression Principle
Debate should compress many surface disagreements into a small set of model-changing cruxes . \boxed{
\text{Debate should compress many surface disagreements into a small set of model-changing cruxes}.
} Debate should compress many surface disagreements into a small set of model-changing cruxes .
193. 第五原則:Green Legality Principle
Successor models outside the initial side assignment must be legal outputs . \boxed{
\text{Successor models outside the initial side assignment must be legal outputs}.
} Successor models outside the initial side assignment must be legal outputs .
194. 第六原則:External Veto Principle
No internally coherent debate outcome is exempt from external verification where external verification is possible . \boxed{
\text{No internally coherent debate outcome is exempt from external verification where external verification is possible}.
} No internally coherent debate outcome is exempt from external verification where external verification is possible .
195. 第七原則:Uncertainty Preservation Principle
Underdetermination should be represented, not rhetorically collapsed . \boxed{
\text{Underdetermination should be represented, not rhetorically collapsed}.
} Underdetermination should be represented, not rhetorically collapsed .
196. 第八原則:Protocol Honesty Principle
A protocol optimized for persuasion should not be interpreted as a truth-discovery procedure without additional validation . \boxed{
\text{A protocol optimized for persuasion should not be interpreted as a truth-discovery procedure without additional validation}.
} A protocol optimized for persuasion should not be interpreted as a truth-discovery procedure without additional validation .
197. ADLM and the Meaning of Debate
經過整個系列,辯論可以被重新定義:
Debate = structured collision of world models under update-permitting constraints . \boxed{
\text{Debate}
=
\text{structured collision of world models under update-permitting constraints}.
} Debate = structured collision of world models under update-permitting constraints .
198. Debate Is Not War
它可以使用 adversarial structure,
但:
the opponent is not the terminal object of destruction . \boxed{
\text{the opponent is not the terminal object of destruction}.
} the opponent is not the terminal object of destruction .
真正被攻擊的是:
model error . \boxed{
\text{model error}.
} model error .
199. Debate Is Not Consensus Production
它也不是:
agreement machine . \boxed{
\text{agreement machine}.
} agreement machine .
真正輸出可以是:
better model;
better branch structure;
better uncertainty;
better experiment;
better question。
200. Debate Is Model Debugging
因此最精簡定義:
Debate = World-Model Debugging under Adversarial Pressure . \boxed{
\text{Debate}
=
\text{World-Model Debugging under Adversarial Pressure}.
} Debate = World-Model Debugging under Adversarial Pressure .
201. Debate Is Also Model Evolution
更完整:
Debate = World-Model Debugging + World-Model Evolution . \boxed{
\text{Debate}
=
\text{World-Model Debugging}
+
\text{World-Model Evolution}.
} Debate = World-Model Debugging + World-Model Evolution .
202. Why AI Is Especially Suitable
AI 可以:
保存完整 state;
不受同樣 working-memory limit;
快速 role inversion;
建 graph;
模擬 counterfactual;
保留多 branch;
重複 self-play。
所以 ADLM 對 AI 特別適合。
203. Why AI Is Also Especially Dangerous
同時 AI 可以:
更會說服;
更會 rationalize;
更快生成大量 arguments;
更快產生 false consensus。
所以:
better debate capability ⇏ better epistemic behavior . \boxed{
\text{better debate capability}
\not\Rightarrow
\text{better epistemic behavior}.
} better debate capability ⇒ better epistemic behavior .
204. Therefore Objective Design Is Central
真正關鍵不是:
給 AI 更多辯論回合。
而是:
what is the reward topology? \boxed{
\text{what is the reward topology?}
} what is the reward topology?
205. Final ADLM Topology
本文最終給出:
Map → Type → Align → Construct → Attack → Invert → Localize → Crux → Release → Green → Discriminate → Verify → Update → Remap . \boxed{
\text{Map}
\rightarrow
\text{Type}
\rightarrow
\text{Align}
\rightarrow
\text{Construct}
\rightarrow
\text{Attack}
\rightarrow
\text{Invert}
\rightarrow
\text{Localize}
\rightarrow
\text{Crux}
\rightarrow
\text{Release}
\rightarrow
\text{Green}
\rightarrow
\text{Discriminate}
\rightarrow
\text{Verify}
\rightarrow
\text{Update}
\rightarrow
\text{Remap}.
} Map → Type → Align → Construct → Attack → Invert → Localize → Crux → Release → Green → Discriminate → Verify → Update → Remap .
206. Minimal Formal Update
可簡寫:
M t + 1 = U ( M t , A , Δ ∗ , c ∗ , G , E ∗ , W ) . \boxed{
M_{t+1}
=
\mathcal U(
M_t,
\mathcal A,
\Delta^\ast,
c^\ast,
G,
E^\ast,
W
).
} M t + 1 = U ( M t , A , Δ ∗ , c ∗ , G , E ∗ , W ) .
207. Final Objective
max Δ Q w o r l d \boxed{
\max
\Delta Q_{\mathrm{world}}
} max Δ Q world
而不是:
max P ( I win ) . \boxed{
\max
P(\text{I win}).
} max P ( I win ) .
208. Series Closure
ADLM-01 到 ADLM-08 最終形成:
Argument → Model → Shared Domain → Divergence → Crux → Green → Evidence → Revision . \boxed{
\text{Argument}
\rightarrow
\text{Model}
\rightarrow
\text{Shared Domain}
\rightarrow
\text{Divergence}
\rightarrow
\text{Crux}
\rightarrow
\text{Green}
\rightarrow
\text{Evidence}
\rightarrow
\text{Revision}.
} Argument → Model → Shared Domain → Divergence → Crux → Green → Evidence → Revision .
209. 最終母命題
本文最終提出:
The purpose of truth-seeking debate is not to preserve a side, but to improve the state of the shared model space . \boxed{
\text{The purpose of truth-seeking debate is not to preserve a side, but to improve the state of the shared model space}.
} The purpose of truth-seeking debate is not to preserve a side, but to improve the state of the shared model space .
中文:
求真型辯論的目的,不是保住某一方,而是讓共同模型空間在辯論後比辯論前更好。
210. 本文不主張事項
本文不主張:
所有 AI 任務都需要 debate;
多 agent 一定優於單 agent;
Green Model 一定存在;
consensus 沒有價值;
adversarial attack 應被削弱;
人類 judge 沒有價值;
外部 verifier 永遠可靠;
world model 可以完全形式化;
ADLM 能保證真理;
AI 可以完全自動解決所有價值分歧。
本文只主張:
AI debate should be designed as an updateable epistemic process rather than a winner-producing rhetorical game . \boxed{
\text{AI debate should be designed as an updateable epistemic process rather than a winner-producing rhetorical game}.
} AI debate should be designed as an updateable epistemic process rather than a winner-producing rhetorical game .
211. 結論
如果 AI 辯論的最終訓練訊號仍然是:
Who won? \boxed{
\text{Who won?}
} Who won?
那麼模型最終很可能學會:
更好的攻擊;
更好的防禦;
更好的 persuasion;
更好的 judge exploitation。
這些能力本身不一定壞。
但它們沒有回答:
AI 是否更會修改錯誤的世界模型? \boxed{
\text{AI 是否更會修改錯誤的世界模型?}
} AI 是否更會修改錯誤的世界模型?
ADLM 因此把真正 learning target 改成:
Model Revision . \boxed{
\text{Model Revision}.
} Model Revision .
一個成熟的 AI debater 應該能:
先不急著站隊;
建立候選空間;
正確標記 claim scope;
建立共同域;
強力 steelman;
強力 attack;
換位重建;
找真正根分歧;
找 causal crux;
放下原角色;
生成 Green candidates;
找能區分模型的 evidence;
接受 world / verifier veto;
按證據強度更新。
所以真正高階 AI 辯論不是:
AI vs AI . \boxed{
\text{AI vs AI}.
} AI vs AI .
而是:
Model t → Model t + 1 . \boxed{
\text{Model}_t
\rightarrow
\text{Model}_{t+1}.
} Model t → Model t + 1 .
Red 與 Blue 只是壓力工具。
Green 只是 successor search。
World / Verifier 才是外部校正。
而真正需要保留的是:
Γ r e v i s i o n . \boxed{
\Gamma_{\mathrm{revision}}.
} Γ revision .
也就是:
模型究竟因為什麼而改變。
若一場辯論最後:
Red loses , \text{Red loses}, Red loses ,
Blue loses , \text{Blue loses}, Blue loses ,
卻:
Q ( M p o s t ) > Q ( M R , M B ) , Q(
M_{\mathrm{post}}
)
>
Q(
M_R,M_B
), Q ( M post ) > Q ( M R , M B ) ,
那不是沒有人贏。
而是:
the learning system won . \boxed{
\text{the learning system won}.
} the learning system won .
這就是 ADLM 系列的 closure。
版本紀錄
v0.1.0 — 2026-10-06
建立 ADLM 系列 Paper 08;
完成 ADLM 主系列 closure;
將 Debate 定義為 structured world-model transition;
定義 ADLM State;
正式建立 No-Side Mode;
統一 Cartographer、Red、Blue、Green、World / Verifier 五角色;
建立 Claim Typing、Predicate / Frame Audit、Shared-Domain Alignment;
建立 Red / Blue Model Stewardship;
建立 Adversarial Falsification 與 Role Inversion;
建立 Divergence Localization、Crux Search 與 Role Release;
建立 Green Generator 與 Green Red-Team;
建立 Discriminating Evidence 與 World / Verifier;
建立 Revision Calibration;
定義 Debate Epistemic Gain;
建立 ADLM Reward Function;
提出 Structured Debate IR 與 Debate Memory Store;
建立八類 benchmark;
建立十項 failure modes;
提出 Minimal ADLM 與 Full ADLM;
建立八項 Series-Level Hypotheses;
建立八項 ADLM 核心原則;
將系列最終收束為 AI World-Model Evolution Protocol。