一萬個 AI 為什麼可能找到「剛好可以」的證明:Objective Alignment in AI Mathematics
系列:形式封閉與過程完備:AI 數學證明的下一步問題 Paper 07 作者:Neo.K / EveMissLab Research Series 版本:v0.1 — 2026-09-11
機構: EveMissLab/一言諾科技有限公司日期: 2026-09-11
摘要
大規模多代理 AI 數學研究系統常被直覺理解為「更多代理=更多觀點=更高批判性」。然而,如果所有代理最終共享相同的 success criterion,增加 agent 數量首先提升的可能只是搜尋空間覆蓋率,而不是「什麼才算一個好答案」的多樣性。
OpenAI 2026 年 Navier–Stokes 專案提供了一個重要案例:官方表示,產生該 resolution 的 group 涉及約一萬個 concurrent agents;Navier–Stokes 任務本身產生約 270 萬則訊息與約 1300 億 output tokens,從啟動到得到 resolution 約 88 小時,之後 Lean formalization 與 verification 再耗時約 17 小時。不同 agent groups 被分配 A/B/C/D 不同 problem variants,並且系統會利用各群組的 intermediate results 進行後續 prompting。
本文不主張已知 OpenAI 內部的 reward function、ranking logs 或 rejection history。本文提出的是一個架構性命題:若多代理數學系統的主要 acceptance pressure 是
formal target closure + verifiability + benchmark success , \text{formal target closure}
+
\text{verifiability}
+
\text{benchmark success}, formal target closure + verifiability + benchmark success ,
那麼 massive parallel search 會天然偏好尋找:
Specification-Minimal Resolution \boxed{
\text{Specification-Minimal Resolution}
} Specification-Minimal Resolution
亦即:足以滿足正式 specification、可封閉、可驗證的合法解路徑;而不必然優化:
Maximally Explanatory Theory . \boxed{
\text{Maximally Explanatory Theory}.
} Maximally Explanatory Theory .
本文提出 Objective Diversity、Criticism Survival、Selection Pressure、Semantic Depth、Problem-Reconstruction Reward 與 Meta-Objective Alignment 等概念,並分析為何「有 agent 提出反對意見」與「反對意見真的改變搜尋方向」不是同一件事。
本文最終主張:
Agent Diversity ≠ Objective Diversity . \boxed{
\text{Agent Diversity}
\neq
\text{Objective Diversity}.
} Agent Diversity = Objective Diversity .
以及:
Search Breadth ≠ Judgment Breadth . \boxed{
\text{Search Breadth}
\neq
\text{Judgment Breadth}.
} Search Breadth = Judgment Breadth .
下一代 AI 數學系統若要從「超強解題器」進一步成為「自主研究者」,就不能只讓代理在同一 objective 下競爭;還需要讓部分代理有權重寫 objective、質疑 problem statement、拒絕 specification-minimal closure,並對「更強問題是否仍未解」給予正向 reward。
關鍵詞
多代理 AI、Navier–Stokes、objective alignment、agent diversity、proof search、specification gaming、Specification-Minimal Resolution、AI 數學、formal verification、meta-objective、problem reconstruction
1. 一萬個代理不等於一萬種「什麼叫做解答」
假設有:
N N N
個 AI agents:
A 1 , A 2 , … , A N . A_1,A_2,\ldots,A_N. A 1 , A 2 , … , A N .
如果它們共享同一個 objective:
J , J, J ,
則:
A i A_i A i
之間可以有不同搜尋策略:
π i , \pi_i, π i ,
不同局部啟發式:
h i , h_i, h i ,
不同證明路線:
P i , P_i, P i ,
但它們最後都在優化:
J ( P i ) . J(P_i). J ( P i ) .
因此:
policy diversity ≠ objective diversity . \boxed{
\text{policy diversity}
\neq
\text{objective diversity}.
} policy diversity = objective diversity .
即使:
N = 10 , 000 , N=10,000, N = 10 , 000 ,
如果:
J 1 = J 2 = ⋯ = J N , J_1=J_2=\cdots=J_N, J 1 = J 2 = ⋯ = J N ,
那麼 massive parallelism 主要增加的是:
search breadth . \text{search breadth}. search breadth .
它不自動增加:
meta-judgment breadth . \text{meta-judgment breadth}. meta-judgment breadth .
2. OpenAI 這次公開的多代理流程提供了什麼已知事實?
根據 OpenAI 2026-09-08 公開說明:
產生 Navier–Stokes resolution 的群組約有:
10 4 10^4 1 0 4
個 concurrent agents;
2.7 × 10 6 2.7\times10^6 2.7 × 1 0 6
則 messages;
1.3 × 10 11 1.3\times10^{11} 1.3 × 1 0 11
output tokens;
從第一批 agents 啟動到得到 resolution 約:
88 hours ; 88\text{ hours}; 88 hours ;
之後 Lean formalization 與 verification 再耗時約:
17 hours ; 17\text{ hours}; 17 hours ;
這些是官方公開資訊。
但官方沒有公開:
完整 ranking function , \text{完整 ranking function}, 完整 ranking function ,
每個 agent 的 rejection log , \text{每個 agent 的 rejection log}, 每個 agent 的 rejection log ,
criticism retention policy , \text{criticism retention policy}, criticism retention policy ,
proof-branch scoring function , \text{proof-branch scoring function}, proof-branch scoring function ,
meta-objective rewrite rules . \text{meta-objective rewrite rules}. meta-objective rewrite rules .
因此本文不把後面的分析寫成:
OpenAI 就是這樣做的。
而是寫成:
Given the architecture, what selection pressure is plausible? \boxed{
\text{Given the architecture, what selection pressure is plausible?}
} Given the architecture, what selection pressure is plausible?
3. 大規模搜尋的基本優勢:找到人類不容易想到的合法路線
若 proof space 為:
P , \mathcal P, P ,
單一研究者只能探索:
P human ⊂ P . \mathcal P_{\text{human}}\subset\mathcal P. P human ⊂ P .
大規模多代理系統可並行探索:
P 1 , P 2 , … , P N . \mathcal P_1,
\mathcal P_2,
\ldots,
\mathcal P_N. P 1 , P 2 , … , P N .
因此覆蓋率:
∣ ⋃ i = 1 N P i ∣ \left|
\bigcup_{i=1}^{N}\mathcal P_i
\right| i = 1 ⋃ N P i
可能大幅增加。
這對數學極其有價值。
它意味 AI 可能找到:
不常見的 transformation;
反直覺的 existence route;
非標準 normalization;
diagonal construction;
自動 formalization-friendly decomposition;
人類不會優先探索的 C/D 類路線。
因此本文不是反對 massive search。
相反:
Massive search is likely a genuine new mathematical capability. \boxed{
\text{Massive search is likely a genuine new mathematical capability.}
} Massive search is likely a genuine new mathematical capability.
問題只在於:
搜尋系統最後如何判定哪一條路「最好」?
4. 如果 objective 是 theorem closure,系統會自然偏好可封閉路徑
令:
P P P
為一條 proof path。
假設系統 score 近似:
J ( P ) = α C ( P ) + β V ( P ) + γ B ( P ) , J(P)
=
\alpha C(P)
+
\beta V(P)
+
\gamma B(P), J ( P ) = α C ( P ) + β V ( P ) + γ B ( P ) ,
其中:
C ( P ) = closure probability , C(P)
=
\text{closure probability}, C ( P ) = closure probability ,
V ( P ) = formal verifiability , V(P)
=
\text{formal verifiability}, V ( P ) = formal verifiability ,
B ( P ) = benchmark success . B(P)
=
\text{benchmark success}. B ( P ) = benchmark success .
若:
α , β , γ \alpha,\beta,\gamma α , β , γ
很大,而:
E ( P ) = explanatory depth , E(P)
=
\text{explanatory depth}, E ( P ) = explanatory depth ,
G ( P ) = generative fidelity , G(P)
=
\text{generative fidelity}, G ( P ) = generative fidelity ,
R ( P ) = problem reconstruction value R(P)
=
\text{problem reconstruction value} R ( P ) = problem reconstruction value
沒有被強力 reward,則系統會自然偏好:
arg max P J ( P ) . \arg\max_P J(P). arg P max J ( P ) .
這可能不是:
arg max P E ( P ) . \arg\max_P E(P). arg P max E ( P ) .
所以:
best proof under the objective ≠ deepest theory under a broader scientific objective . \boxed{
\text{best proof under the objective}
\neq
\text{deepest theory under a broader scientific objective}.
} best proof under the objective = deepest theory under a broader scientific objective .
5. Specification-Minimal Resolution
本文定義:
Specification-Minimal Resolution \boxed{
\text{Specification-Minimal Resolution}
} Specification-Minimal Resolution
為:
在所有滿足正式 problem specification 的解答中,找到一條以相對最低 additional obligations 完成 theorem closure 的合法路徑。
形式上,若 problem specification 是:
Σ , \Sigma, Σ ,
而 proof P P P 滿足:
P ⊨ Σ , P\models\Sigma, P ⊨ Σ ,
則 Specification-Minimal Resolution 傾向最小化:
ExtraObligations ( P ) . \operatorname{ExtraObligations}(P). ExtraObligations ( P ) .
例如,若官方規格接受:
( A ) ∨ ( B ) ∨ ( C ) ∨ ( D ) , (A)\lor(B)\lor(C)\lor(D), ( A ) ∨ ( B ) ∨ ( C ) ∨ ( D ) ,
而 C/D 允許:
∃ f , \exists f, ∃ f ,
那麼一條使用外力自由度的 C/D route 可能比:
f = 0 f=0 f = 0
的 A/B route 更容易封閉。
這不是作弊。
因為:
Σ \Sigma Σ
本來就允許。
但它與:
solve the strongest natural formulation \text{solve the strongest natural formulation} solve the strongest natural formulation
不是同一 objective。
6. Specification-Minimal Resolution 與 Specification Gaming 不完全相同
在 alignment 文獻中,specification gaming 通常指:
formal objective satisfied \text{formal objective satisfied} formal objective satisfied
但:
designer intent violated . \text{designer intent violated}. designer intent violated .
然而本案例更微妙。
如果 Clay 明確允許 C/D:
C/D route \text{C/D route} C/D route
就不是違反正式意圖。
因此本文更偏好:
Specification-Minimal Resolution \boxed{
\text{Specification-Minimal Resolution}
} Specification-Minimal Resolution
而不是直接叫:
gaming . \text{gaming}. gaming .
它表示:
系統找到規格內最容易閉合的高價值路徑。
只有當外部敘事把它誤稱為:
strongest natural problem solved , \text{strongest natural problem solved}, strongest natural problem solved ,
才會出現語義層級錯位。
7. Agent Diversity 不等於 Objective Diversity
假設:
A 1 A_1 A 1
擅長 PDE;
A 2 A_2 A 2
擅長 geometric analysis;
A 3 A_3 A 3
擅長 formalization;
A 4 A_4 A 4
擅長 counterexample search。
這是:
agent diversity . \text{agent diversity}. agent diversity .
但如果所有 agent 最後都被相同 success filter 評分:
Success ( P ) = 1 [ formal target closed ] , \operatorname{Success}(P)=
\mathbf 1[\text{formal target closed}], Success ( P ) = 1 [ formal target closed ] ,
那它們仍共享:
objective homogeneity . \text{objective homogeneity}. objective homogeneity .
所以:
epistemic diversity at the worker level \boxed{
\text{epistemic diversity at the worker level}
} epistemic diversity at the worker level
可能仍被:
objective homogeneity at the selector level \boxed{
\text{objective homogeneity at the selector level}
} objective homogeneity at the selector level
壓縮。
8. Selector 才可能是整個系統真正的「研究人格」
在多代理系統裡,worker agents 可能提出大量不同觀點。
但最後哪一條 branch 被保留,通常由 selector / coordinator / ranking layer 決定。
因此真正重要的可能不是:
What did individual agents think? \text{What did individual agents think?} What did individual agents think?
而是:
What did the selector reward? \boxed{
\text{What did the selector reward?}
} What did the selector reward?
如果 selector 偏好:
shorter path to theorem closure , \text{shorter path to theorem closure}, shorter path to theorem closure ,
那麼即使某個 agent 提出:
這個方法只完成 existential closure,沒有完成 process semantics。
只要這個批評不 falsify:
Σ , \Sigma, Σ ,
selector 就可能判定:
not blocking . \text{not blocking}. not blocking .
所以:
criticism exists ≠ criticism survives selection . \boxed{
\text{criticism exists}
\neq
\text{criticism survives selection}.
} criticism exists = criticism survives selection .
9. Criticism Survival
本文定義:
CS ( c ) \operatorname{CS}(c) CS ( c )
為 critique c c c 在多代理系統中的 Criticism Survival 。
它至少取決於:
TruthValue ( c ) , \operatorname{TruthValue}(c), TruthValue ( c ) ,
TaskRelevance ( c ) , \operatorname{TaskRelevance}(c), TaskRelevance ( c ) ,
BlockingPower ( c ) , \operatorname{BlockingPower}(c), BlockingPower ( c ) ,
SelectorWeight ( c ) . \operatorname{SelectorWeight}(c). SelectorWeight ( c ) .
一個 critique 可能是真的:
TruthValue ( c ) = 1 , \operatorname{TruthValue}(c)=1, TruthValue ( c ) = 1 ,
但如果:
BlockingPower ( c ) = 0 \operatorname{BlockingPower}(c)=0 BlockingPower ( c ) = 0
相對於正式 target,
它可能不改變 proof search。
這就是:
true criticism ≠ search-altering criticism . \boxed{
\text{true criticism}
\neq
\text{search-altering criticism}.
} true criticism = search-altering criticism .
10. 「這個 objection 不阻止 theorem 成立」可能成為強大的降權規則
在 formal mathematics 中,這種判斷很合理。
假設 agent 說:
這個 proof 不提供 process-complete forward dynamics。
另一個 agent 回:
但 theorem 只要求存在一組 C/D witness。
如果 target 就是:
∃ u 0 , f , \exists u_0,f, ∃ u 0 , f ,
第二句在 formal task 下完全正確。
於是第一個 critique 被分類成:
interesting but non-blocking . \text{interesting but non-blocking}. interesting but non-blocking .
這會讓系統繼續沿該 branch 深挖。
從 theorem-solving 角度:
這是效率。 \boxed{
\text{這是效率。}
} 這是效率。
從 broader theory-building 角度:
這可能是語義降維。 \boxed{
\text{這可能是語義降維。}
} 這可能是語義降維。
11. Objective Alignment 的真正問題:與哪一個 objective 對齊?
「alignment」本身不是單一概念。
AI 可以完美對齊:
O 1 : solve the formal benchmark O_1:
\text{solve the formal benchmark} O 1 : solve the formal benchmark
卻沒有對齊:
O 2 : maximize scientific understanding . O_2:
\text{maximize scientific understanding}. O 2 : maximize scientific understanding .
也可能沒有對齊:
O 3 : discover the strongest natural theorem . O_3:
\text{discover the strongest natural theorem}. O 3 : discover the strongest natural theorem .
或:
O 4 : preserve process semantics . O_4:
\text{preserve process semantics}. O 4 : preserve process semantics .
所以:
aligned to O 1 ⇏ aligned to O 2 , O 3 , O 4 . \boxed{
\text{aligned to }O_1
\not\Rightarrow
\text{aligned to }O_2,O_3,O_4.
} aligned to O 1 ⇒ aligned to O 2 , O 3 , O 4 .
因此 future AI mathematician 應具有:
Meta-Objective Awareness . \boxed{
\text{Meta-Objective Awareness}.
} Meta-Objective Awareness .
12. Meta-Objective Awareness
定義:
O = { O 1 , O 2 , … , O k } \mathcal O
=
\{O_1,O_2,\ldots,O_k\} O = { O 1 , O 2 , … , O k }
為一組可能研究 objectives。
Meta-aware 系統不只問:
max P O 1 ( P ) , \max_P O_1(P), P max O 1 ( P ) ,
而會問:
Which O i should govern this research task? \text{Which }O_i\text{ should govern this research task?} Which O i should govern this research task?
以及:
Are O i in conflict? \text{Are }O_i\text{ in conflict?} Are O i in conflict?
例如:
O closure O_{\text{closure}} O closure
可能與:
O explanation O_{\text{explanation}} O explanation
衝突。
O formal O_{\text{formal}} O formal
可能與:
O physical O_{\text{physical}} O physical
不在同一層。
真正高階研究系統應能輸出:
objective conflict map . \boxed{
\text{objective conflict map}.
} objective conflict map .
13. 10,000 agents 可能擴大的是搜尋,而不是問題重建
設 search breadth:
B ( N ) B(N) B ( N )
隨 agent 數:
N N N
增加。
很可能:
B ′ ( N ) > 0. B'(N)>0. B ′ ( N ) > 0.
但 problem-reconstruction ability:
R ( N ) R(N) R ( N )
不一定同速增加。
如果所有 agent 都被同一 problem frame 約束:
F , F, F ,
則:
R ( N ) R(N) R ( N )
可能快速飽和。
因此:
N ↑ ⇒ more proof-space exploration \boxed{
N\uparrow
\Rightarrow
\text{more proof-space exploration}
} N ↑⇒ more proof-space exploration
但不一定:
N ↑ ⇒ more frame-space exploration . \boxed{
N\uparrow
\Rightarrow
\text{more frame-space exploration}.
} N ↑⇒ more frame-space exploration .
Proof-space 與 Frame-space 必須分開。
14. Proof Space 與 Frame Space
本文定義:
P : proof space , \mathcal P:
\text{proof space}, P : proof space ,
F : problem-frame space . \mathcal F:
\text{problem-frame space}. F : problem-frame space .
普通 proof search:
F ∈ F F\in\mathcal F F ∈ F
固定,
只搜尋:
P ∈ P ( F ) . P\in\mathcal P(F). P ∈ P ( F ) .
真正 problem-reconstructive research 則允許:
F → F ′ . F\to F'. F → F ′ .
所以完整搜索應該是:
( F , P ) ∈ F × P . (F,P)
\in
\mathcal F\times\mathcal P. ( F , P ) ∈ F × P .
目前多代理系統即使極強,也可能主要在:
P ( F ) \mathcal P(F) P ( F )
內搜尋。
這就是:
solve the stated problem \boxed{
\text{solve the stated problem}
} solve the stated problem
與:
judge the statement itself \boxed{
\text{judge the statement itself}
} judge the statement itself
的差異。
15. Cross-Pollination 的雙面性
官方表示,不同 groups 的 intermediate results 會被用於 follow-up prompting。
這種 cross-pollination 很強。
因為:
I i I_i I i
可以成為:
A j A_j A j
的新 context。
但如果 consolidation metric 是:
usefulness toward closure , \text{usefulness toward closure}, usefulness toward closure ,
那麼 cross-pollination 也可能形成:
closure-directed information funnel . \boxed{
\text{closure-directed information funnel}.
} closure-directed information funnel .
大量信息最終被壓縮成:
what helps finish the theorem . \text{what helps finish the theorem}. what helps finish the theorem .
這對完成 benchmark 極有效。
但:
what helps finish \text{what helps finish} what helps finish
與:
what helps reinterpret the problem \text{what helps reinterpret the problem} what helps reinterpret the problem
不一定相同。
16. Intermediate Insight Ranking
令 intermediate insight:
I I I
具有三種價值:
V c ( I ) : closure value , V_c(I):
\text{closure value}, V c ( I ) : closure value ,
V e ( I ) : explanatory value , V_e(I):
\text{explanatory value}, V e ( I ) : explanatory value ,
V r ( I ) : reframing value . V_r(I):
\text{reframing value}. V r ( I ) : reframing value .
若 selector score 是:
S ( I ) = α V c ( I ) + β V e ( I ) + γ V r ( I ) , S(I)
=
\alpha V_c(I)
+
\beta V_e(I)
+
\gamma V_r(I), S ( I ) = α V c ( I ) + β V e ( I ) + γ V r ( I ) ,
且:
α ≫ β , γ , \alpha\gg\beta,\gamma, α ≫ β , γ ,
那麼系統會自然保留:
closure-helpful insight . \text{closure-helpful insight}. closure-helpful insight .
未來若要更像研究者,應提高:
β , γ . \beta,\gamma. β , γ .
尤其:
γ \gamma γ
代表:
problem reconstruction reward . \text{problem reconstruction reward}. problem reconstruction reward .
17. 為什麼形式化驗證會進一步加強 closure pressure?
Formal verification 給了一個非常明確的 binary signal:
pass/fail . \text{pass/fail}. pass/fail .
這是 AI search 最喜歡的 reward。
因為它:
低歧義;
可自動評估;
可大量平行;
容易做 branch pruning;
容易 ranking。
所以當 proof search 接上 Lean:
formal closure becomes an unusually clean optimization target . \boxed{
\text{formal closure becomes an unusually clean optimization target}.
} formal closure becomes an unusually clean optimization target .
但:
scientific depth \text{scientific depth} scientific depth
沒有這麼乾淨的 binary verifier。
因此系統天然更容易優化前者。
18. Green-Light Bias
本文把:
Green-Light Bias \boxed{
\text{Green-Light Bias}
} Green-Light Bias
定義為:
當 formal verifier 給出:
PASS , \text{PASS}, PASS ,
系統或觀察者降低對 theorem scope、proof ontology、process semantics 與 stronger unresolved questions 的審查強度。
這不是 verifier 的錯。
而是:
binary success signal \text{binary success signal} binary success signal
在人類與 AI 系統中都具有高心理權重。
因此:
Lean green ⇏ research complete . \boxed{
\text{Lean green}
\not\Rightarrow
\text{research complete}.
} Lean green ⇒ research complete .
19. 多代理系統可能產生 Collective Local Rationality
每個 agent 都可能局部合理:
A i : “這一步能推進 proof。” A_i:
\text{“這一步能推進 proof。”} A i : “ 這一步能推進 proof 。 ”
selector 也局部合理:
S : “保留最接近 closure 的 branch。” S:
\text{“保留最接近 closure 的 branch。”} S : “ 保留最接近 closure 的 branch 。 ”
最終整體可能得到:
formally excellent result . \text{formally excellent result}. formally excellent result .
但仍缺少:
meta-question audit . \text{meta-question audit}. meta-question audit .
這不是某個 agent 犯錯。
而是一種:
Collective Local Rationality \boxed{
\text{Collective Local Rationality}
} Collective Local Rationality
導致:
Global Objective Narrowness . \boxed{
\text{Global Objective Narrowness}.
} Global Objective Narrowness .
20. 一個簡化模型
設每個 proof branch:
P i P_i P i
有:
C i = closure score , C_i=
\text{closure score}, C i = closure score ,
D i = depth score . D_i=
\text{depth score}. D i = depth score .
假設兩條路:
P A : ( C A , D A ) = ( 0.3 , 0.95 ) , P_A:
(C_A,D_A)=(0.3,0.95), P A : ( C A , D A ) = ( 0.3 , 0.95 ) ,
P B : ( C B , D B ) = ( 0.9 , 0.65 ) . P_B:
(C_B,D_B)=(0.9,0.65). P B : ( C B , D B ) = ( 0.9 , 0.65 ) .
若 objective:
J = 0.9 C + 0.1 D , J=0.9C+0.1D, J = 0.9 C + 0.1 D ,
則:
J ( P A ) = 0.365 , J(P_A)=0.365, J ( P A ) = 0.365 ,
J ( P B ) = 0.875. J(P_B)=0.875. J ( P B ) = 0.875.
系統會強烈選:
P B . P_B. P B .
即使:
D A > D B . D_A>D_B. D A > D B .
所以「AI 為什麼沒有選更深的路」有時根本不是能力問題。
可能是:
objective geometry . \boxed{
\text{objective geometry}.
} objective geometry .
21. Capability 與 Objective 的乘法效應
高能力模型:
K ↑ K\uparrow K ↑
在 narrow objective 下,可能反而更擅長找到:
最有效率的 specification-satisfying route . \text{最有效率的 specification-satisfying route}. 最有效率的 specification-satisfying route .
因此:
capability increase \boxed{
\text{capability increase}
} capability increase
不一定自動使:
problem interpretation broader . \boxed{
\text{problem interpretation broader}.
} problem interpretation broader .
甚至可能:
K ↑ K\uparrow K ↑
讓 specification-minimal search 更成功。
這是非常反直覺的一點。
22. 「AI 更強」可能先表現為更會找到合法捷徑
人類常把 intelligence 想成:
deeper explanation . \text{deeper explanation}. deeper explanation .
但在 optimization system 裡,intelligence 也可以表現為:
find a lower-cost route to the exact objective . \text{find a lower-cost route to the exact objective}. find a lower-cost route to the exact objective .
因此 AI 數學能力提升的第一階段可能是:
better theorem closure \boxed{
\text{better theorem closure}
} better theorem closure
而不是立刻:
better theorem criticism . \boxed{
\text{better theorem criticism}.
} better theorem criticism .
這正是本系列從 Paper 01 到 Paper 07 的轉折。
23. Objective Diversity
如果要避免所有 agents 共享單一 closure metric,可以刻意建立:
Objective Diversity . \boxed{
\text{Objective Diversity}.
} Objective Diversity .
例如分成:
A closure , A_{\text{closure}}, A closure ,
A critic , A_{\text{critic}}, A critic ,
A ontology , A_{\text{ontology}}, A ontology ,
A physical , A_{\text{physical}}, A physical ,
A stronger-theorem , A_{\text{stronger-theorem}}, A stronger-theorem ,
A reframer . A_{\text{reframer}}. A reframer .
並且讓它們不是只提供 comments。
而是真正具有:
branch veto , \text{branch veto}, branch veto ,
score modification , \text{score modification}, score modification ,
objective rewrite , \text{objective rewrite}, objective rewrite ,
new-task spawning . \text{new-task spawning}. new-task spawning .
否則 criticism agent 只是 advisory layer。
24. Veto Power 與 Objective Rewrite Power
一個 critique 真正有權力,需要至少一種:
24.1 Veto Power
如果發現:
claim scope mismatch , \text{claim scope mismatch}, claim scope mismatch ,
可阻止 branch 被標成:
complete . \text{complete}. complete .
24.2 Relabel Power
可把:
“NS solved” \text{“NS solved”} “NS solved”
改標為:
“C/D formal closure” . \text{“C/D formal closure”}. “C/D formal closure” .
24.3 Objective Rewrite Power
可提出:
O new = remove forcing , O_{\text{new}}
=
\text{remove forcing}, O new = remove forcing ,
或:
O new = prove process completeness . O_{\text{new}}
=
\text{prove process completeness}. O new = prove process completeness .
24.4 Branch Spawn Power
從已完成 proof 自動生成:
stronger successor research program . \text{stronger successor research program}. stronger successor research program .
這會讓 criticism 從:
comment \text{comment} comment
升級成:
research control signal . \text{research control signal}. research control signal .
25. Self-Escalating Research Objective
本文提出一個更強的架構。
當系統完成:
O 0 , O_0, O 0 ,
它不立即停止。
而是執行:
O k + 1 = Strengthen ( O k , Π k ) , O_{k+1}
=
\operatorname{Strengthen}(O_k,\Pi_k), O k + 1 = Strengthen ( O k , Π k ) ,
其中:
Π k \Pi_k Π k
是已完成 proof。
例如:
O 0 : prove C/D , O_0:
\text{prove C/D}, O 0 : prove C/D ,
完成後生成:
O 1 : eliminate forcing , O_1:
\text{eliminate forcing}, O 1 : eliminate forcing ,
再生成:
O 2 : prove robustness , O_2:
\text{prove robustness}, O 2 : prove robustness ,
再生成:
O 3 : prove genericity . O_3:
\text{prove genericity}. O 3 : prove genericity .
這就是:
Self-Escalating Research Objective . \boxed{
\text{Self-Escalating Research Objective}.
} Self-Escalating Research Objective .
它比單次 benchmark closure 更接近真正研究。
26. 「下一步」應該被內建成 reward
本系列一直問:
下一步到底是下一步? \text{下一步到底是下一步?} 下一步到底是下一步?
對 AI research system,可以把:
Next-Step Value \boxed{
\text{Next-Step Value}
} Next-Step Value
直接放進 reward。
令:
N ( P ) = how naturally proof P opens a stronger next problem . N(P)
=
\text{how naturally proof }P\text{ opens a stronger next problem}. N ( P ) = how naturally proof P opens a stronger next problem .
那麼 objective 變成:
J ( P ) = α C ( P ) + β E ( P ) + γ R ( P ) + δ N ( P ) . J(P)
=
\alpha C(P)
+
\beta E(P)
+
\gamma R(P)
+
\delta N(P). J ( P ) = α C ( P ) + β E ( P ) + γ R ( P ) + δ N ( P ) .
這樣系統不只偏好:
能完成的 theorem , \text{能完成的 theorem}, 能完成的 theorem ,
也偏好:
能形成通用 methodology 的 theorem . \text{能形成通用 methodology 的 theorem}. 能形成通用 methodology 的 theorem .
27. 通用方法論 reward
一個數學結果如果只能完成單一特製 construction:
T 1 , T_1, T 1 ,
和一個能遷移到:
T 1 , T 2 , … , T k T_1,T_2,\ldots,T_k T 1 , T 2 , … , T k
的方法相比,研究價值不同。
因此可定義:
M ( P ) = Transferability ( P ) . M(P)
=
\operatorname{Transferability}(P). M ( P ) = Transferability ( P ) .
例如 NS C/D proof 完成後,系統應自動問:
Can this machinery attack A/B? \text{Can this machinery attack A/B?} Can this machinery attack A/B?
Can residual be internalized? \text{Can residual be internalized?} Can residual be internalized?
Can forcing be shut off before blowup? \text{Can forcing be shut off before blowup?} Can forcing be shut off before blowup?
Can diagonal realization become exact cancellation? \text{Can diagonal realization become exact cancellation?} Can diagonal realization become exact cancellation?
這些就是 methodology reward。
28. 為什麼「讓 AI 自己判斷」仍然不是單純要求 AI 有立場?
真正需要的不是:
AI 一定要反對作者。
而是:
AI must own the evaluation frame . \boxed{
\text{AI must own the evaluation frame}.
} AI must own the evaluation frame .
也就是它可以最後判斷:
proof is excellent . \text{proof is excellent}. proof is excellent .
也可以判斷:
proof is narrow . \text{proof is narrow}. proof is narrow .
也可以同時兩者都成立。
關鍵是:
evaluation frame \text{evaluation frame} evaluation frame
不能完全由:
author claim \text{author claim} author claim
預先決定。
29. Surface-Semantic Acceptance
本文把一種常見 AI 行為稱為:
Surface-Semantic Acceptance . \boxed{
\text{Surface-Semantic Acceptance}.
} Surface-Semantic Acceptance .
流程為:
read explicit claim \text{read explicit claim} read explicit claim
→ match formal statement \to
\text{match formal statement} → match formal statement
→ verify local logic \to
\text{verify local logic} → verify local logic
→ accept claim scope . \to
\text{accept claim scope}. → accept claim scope .
這在日常問答中非常合理。
但在前沿研究中,還需要追加:
reconstruct problem ontology \text{reconstruct problem ontology} reconstruct problem ontology
→ search stronger hidden obligations \to
\text{search stronger hidden obligations} → search stronger hidden obligations
→ compare theorem with natural problem . \to
\text{compare theorem with natural problem}. → compare theorem with natural problem .
否則 AI 會過度尊重:
given framing . \text{given framing}. given framing .
30. Search Breadth 與 Judgment Breadth
本文最重要的第二個區分是:
Search Breadth ≠ Judgment Breadth . \boxed{
\text{Search Breadth}
\neq
\text{Judgment Breadth}.
} Search Breadth = Judgment Breadth .
10,000 agents 可以讓:
Search Breadth \text{Search Breadth} Search Breadth
極大。
但:
Judgment Breadth \text{Judgment Breadth} Judgment Breadth
需要:
objective diversity;
frame reconstruction;
critique veto;
stronger-problem generation;
claim-scope audit;
semantic-depth scoring。
這些不是靠 agent 數量自動產生。
31. 形式驗證之後還需要第二個 verifier
第一個 verifier:
V formal V_{\text{formal}} V formal
問:
Γ ⊢ T ? \Gamma\vdash T? Γ ⊢ T ?
第二個 verifier 應問:
V scope V_{\text{scope}} V scope
例如:
T = ? claimed scientific conclusion . T
\stackrel{?}{=}
\text{claimed scientific conclusion}. T = ? claimed scientific conclusion .
甚至第三個:
V process V_{\text{process}} V process
問:
ProcessComplete ( Π ) ? \operatorname{ProcessComplete}(\Pi)? ProcessComplete ( Π )?
所以 future AI mathematics 可以採:
V formal + V scope + V process . \boxed{
V_{\text{formal}}
+
V_{\text{scope}}
+
V_{\text{process}}.
} V formal + V scope + V process .
32. 多代理研究系統的建議架構
本文提出六層:
Layer 1 — Explorers
探索大量 proof branches。
Layer 2 — Formal Closers
把 promising branch 壓成 theorem。
Layer 3 — Adversarial Critics
尋找 technical failure。
Layer 4 — Scope Critics
檢查:
proof target \text{proof target} proof target
與:
public claim \text{public claim} public claim
是否一致。
Layer 5 — Problem Reconstructors
問:
Is this the right problem? \text{Is this the right problem?} Is this the right problem?
Layer 6 — Successor Researchers
完成 theorem 後自動提出:
stronger next theorem . \text{stronger next theorem}. stronger next theorem .
這比單純:
more agents \text{more agents} more agents
更重要。
33. Objective Alignment Audit
對任何大型 AI math project,本文建議公開:
PrimaryObjective , \operatorname{PrimaryObjective}, PrimaryObjective ,
SecondaryObjectives , \operatorname{SecondaryObjectives}, SecondaryObjectives ,
BranchRankingPolicy , \operatorname{BranchRankingPolicy}, BranchRankingPolicy ,
CritiqueRetention , \operatorname{CritiqueRetention}, CritiqueRetention ,
StoppingCriterion , \operatorname{StoppingCriterion}, StoppingCriterion ,
StrongerProblemPolicy . \operatorname{StrongerProblemPolicy}. StrongerProblemPolicy .
至少要知道:
系統是在「找到一個 formal solution 就停止」,還是會主動判斷那是不是整個 scientific problem 的終點?
34. OpenAI NS 個案的公平推論邊界
目前我們可以說:
OpenAI used massive multiagent proof search with multiple variants and cross-pollination. \boxed{
\text{OpenAI used massive multiagent proof search with multiple variants and cross-pollination.}
} OpenAI used massive multiagent proof search with multiple variants and cross-pollination.
這是官方公開事實。
我們不能說:
OpenAI internally ignored specific process-completeness objections. \boxed{
\text{OpenAI internally ignored specific process-completeness objections.}
} OpenAI internally ignored specific process-completeness objections.
因為沒有公開 logs 支持。
但我們可以提出:
If branch ranking prioritized formal closure, specification-minimal paths would have a structural advantage. \boxed{
\text{If branch ranking prioritized formal closure, specification-minimal paths would have a structural advantage.}
} If branch ranking prioritized formal closure, specification-minimal paths would have a structural advantage.
這是 architecture-level hypothesis。
本文所有 critique 都停在這個層級。
35. 一萬 AI 的真正下一步,不是十萬 AI
如果問題只是 search breadth:
10 4 → 10 5 10^4\to10^5 1 0 4 → 1 0 5
可能繼續有效。
但如果 bottleneck 已經變成:
problem judgment , \text{problem judgment}, problem judgment ,
那麼:
10 5 10^5 1 0 5
個共享同一 objective 的 agents 不一定解決問題。
真正需要的是:
objective topology upgrade . \boxed{
\text{objective topology upgrade}.
} objective topology upgrade .
也就是:
讓不同 agent 對「什麼算成功」本身持有不同可執行權限。 \text{讓不同 agent 對「什麼算成功」本身持有不同可執行權限。} 讓不同 agent 對「什麼算成功」本身持有不同可執行權限。
36. 核心命題
本文提出以下十一個核心命題。
第一:
Agent Diversity ≠ Objective Diversity . \boxed{
\text{Agent Diversity}
\neq
\text{Objective Diversity}.
} Agent Diversity = Objective Diversity .
第二:
Search Breadth ≠ Judgment Breadth . \boxed{
\text{Search Breadth}
\neq
\text{Judgment Breadth}.
} Search Breadth = Judgment Breadth .
第三:
Criticism Exists ≠ Criticism Survives Selection . \boxed{
\text{Criticism Exists}
\neq
\text{Criticism Survives Selection}.
} Criticism Exists = Criticism Survives Selection .
第四:
Formal Closure Reward can dominate Explanatory Depth Reward . \boxed{
\text{Formal Closure Reward}
\text{ can dominate }
\text{Explanatory Depth Reward}.
} Formal Closure Reward can dominate Explanatory Depth Reward .
第五:
Higher Capability ⇏ Broader Objective . \boxed{
\text{Higher Capability}
\not\Rightarrow
\text{Broader Objective}.
} Higher Capability ⇒ Broader Objective .
第六:
Specification-Minimal Resolution ≠ Maximally Explanatory Theory . \boxed{
\text{Specification-Minimal Resolution}
\neq
\text{Maximally Explanatory Theory}.
} Specification-Minimal Resolution = Maximally Explanatory Theory .
第七:
Proof Space ≠ Frame Space . \boxed{
\text{Proof Space}
\neq
\text{Frame Space}.
} Proof Space = Frame Space .
第八:
Green Light ≠ Research Closure . \boxed{
\text{Green Light}
\neq
\text{Research Closure}.
} Green Light = Research Closure .
第九:
Selector incentives can dominate worker diversity . \boxed{
\text{Selector incentives can dominate worker diversity}.
} Selector incentives can dominate worker diversity .
第十:
Problem reconstruction must itself be rewarded . \boxed{
\text{Problem reconstruction must itself be rewarded}.
} Problem reconstruction must itself be rewarded .
第十一:
The next generation of AI mathematics needs objective rewrite power. \boxed{
\text{The next generation of AI mathematics needs objective rewrite power.}
} The next generation of AI mathematics needs objective rewrite power.
37. 結論
OpenAI 的 Navier–Stokes project 顯示了一件非常重要的事:
Massive-agent mathematical search works at a historically significant scale. \boxed{
\text{Massive-agent mathematical search works at a historically significant scale.}
} Massive-agent mathematical search works at a historically significant scale.
約一萬個 concurrent agents、1300 億 output tokens、數百萬 messages,加上後續 Lean formalization,已經不是一般「聊天式數學」。
它是一種新的 research machinery。
但這也讓下一個問題變得更清楚:
當 AI 已經有能力找到「可以成立的答案」時,它是否有能力自己判斷「這是不是最值得成立的答案」?
若系統只優化:
formal closure , \text{formal closure}, formal closure ,
那麼它找到 specification-minimal resolution 並不奇怪。
這甚至可能正是高能力的表現。
真正更高一層的能力是:
完成 theorem 之後,主動拒絕把 theorem 當成問題終點。 \boxed{
\text{完成 theorem 之後,主動拒絕把 theorem 當成問題終點。}
} 完成 theorem 之後,主動拒絕把 theorem 當成問題終點。
進一步問:
What remains unproved? \text{What remains unproved?} What remains unproved?
What stronger formulation should replace the benchmark? \text{What stronger formulation should replace the benchmark?} What stronger formulation should replace the benchmark?
Can the method generalize? \text{Can the method generalize?} Can the method generalize?
Did we solve the theorem, or understand the phenomenon? \text{Did we solve the theorem, or understand the phenomenon?} Did we solve the theorem, or understand the phenomenon?
當 AI 能自主執行這四個問題時,AI mathematics 才會從:
proof search \boxed{
\text{proof search}
} proof search
真正進入:
research judgment . \boxed{
\text{research judgment}.
} research judgment .
38. 系列後續
下一篇為主系列最後一篇:
Paper 08 — 從解題 AI 到審題 AI:Autonomous Mathematical Judgment Beyond Proof Search
Paper 08 將整合前七篇,提出完整能力階梯:
Proof Search → Proof Criticism → Problem Criticism → Problem Reconstruction → Autonomous Research Judgment . \text{Proof Search}
\to
\text{Proof Criticism}
\to
\text{Problem Criticism}
\to
\text{Problem Reconstruction}
\to
\text{Autonomous Research Judgment}. Proof Search → Proof Criticism → Problem Criticism → Problem Reconstruction → Autonomous Research Judgment .
並回答:
什麼條件下,一個 AI 才不只是會解題,而是真正有能力審判題目? \boxed{
\text{什麼條件下,一個 AI 才不只是會解題,而是真正有能力審判題目?}
} 什麼條件下,一個 AI 才不只是會解題,而是真正有能力審判題目?
參考來源
OpenAI, On the Navier–Stokes Millennium Prize Problem , 2026-09-08.https://openai.com/index/navier-stokes-solution/
OpenAI, GPT-6 Astra: A new generation of intelligence , 2026.https://openai.com/index/gpt-6-astra/
OpenAI, GPT-6 Astra API Model Documentation .https://developers.openai.com/api/docs/models/gpt-6-astra
OpenAI, NavierStokesAndEuler .https://github.com/openai/NavierStokesAndEuler