疑點驅動的推理資源配置:Freeze Controller、Value of Information 與自適應 Specialist Routing
Doubt-Driven Reasoning Allocation: Freeze Control, Value of Information, and Adaptive Specialist Routing
作者: Neo.K機構: EveMissLab/一言諾科技有限公司日期: 2026-08-11
摘要
隨著大型語言模型與形式定理證明系統的 inference-time compute 持續擴張,數學 AI 面臨的問題已不再只是「能否投入更多推理資源」,而是「何時值得投入更多資源,以及資源應投入哪一種推理活動」。現有 adaptive test-time compute 研究主要依據問題難度、不確定性或預期正確率收益,決定對單一輸入增加多少 sampling、search 或 verification;多 Agent 形式數學系統則開始將自然語言推理、形式驗證、修復與 orchestration 分離。然而,對研究級數學問題而言,尚存在一個更上游的資源配置問題:當 theorem target 尚未完全穩定時,額外計算究竟應該用於量詞與定義域分析、來源版本核對、自然語言解釋、跨框架 bridge audit,還是應停止擴張並凍結問題?
本文提出 Doubt-Driven Reasoning Allocation(DDRA) ,將 proof-search 前的 residual uncertainty 表示為多維 residual doubt vector :
D t = ( D F , D S , D B , D P , … ) , D_t=(D_F,D_S,D_B,D_P,\ldots), D t = ( D F , D S , D B , D P , … ) ,
其中各維分別代表 formal-structural、source-fidelity、framework-bridge、perspective/pragmatic 等未決風險。系統維持一組認知 specialist:
M = { M M P F , M S o u r c e , M B r i d g e , M N L U , … } , \mathcal M
=
\{
M_{\mathrm{MPF}},
M_{\mathrm{Source}},
M_{\mathrm{Bridge}},
M_{\mathrm{NLU}},
\ldots
\}, M = { M MPF , M Source , M Bridge , M NLU , … } ,
並估計每個 specialist 的成本 C ( M i ) C(M_i) C ( M i ) 與對 residual doubt 的預期消除量。由此定義:
VOI ( M i ∣ D t ) = E [ V ( D t ) − V ( D t + 1 ) ∣ M i ] C ( M i ) , \operatorname{VOI}(M_i\mid D_t)
=
\frac{
\mathbb E[
V(D_t)-V(D_{t+1})
\mid M_i
]
}{
C(M_i)
}, VOI ( M i ∣ D t ) = C ( M i ) E [ V ( D t ) − V ( D t + 1 ) ∣ M i ] ,
其中 V ( D ) V(D) V ( D ) 是 unresolved doubt 的加權風險函數。
本文將推理控制分為兩個問題:
Should I think more? \boxed{\text{Should I think more?}} Should I think more?
以及:
What should I think about next? \boxed{\text{What should I think about next?}} What should I think about next?
前者由 Freeze Controller 決定;後者由 VOI Specialist Scheduler 決定。當 residual doubt 已低於 freeze threshold,或所有剩餘 specialist 的邊際 VOI 低於停止門檻時,系統停止語義擴張並形成 frozen theorem target;反之,選擇最高 VOI specialist 進行下一步審計。
本文整理一組由先前 error-injection 與 holdout-style experiments 所衍生的結構代理實驗。結果顯示,在既定人工標註與 proxy cost model 下,adaptive routing 可以在保持 corruption coverage 的同時顯著減少不必要 specialist calls;但本文明確不將這些結果解讀為實際 token、latency 或 theorem-prover speedup。本文的主要貢獻是提出一個可與真實 telemetry 接軌的推理資源治理模型,使數學 AI 從「固定流程的 solver」轉向「依問題狀態動態分配認知資源的 orchestrator」。
關鍵詞: Reasoning Allocation、Value of Information、Test-Time Compute、Specialist Routing、Freeze Controller、Residual Doubt、Mathematical AI、Agent Orchestration、Inference-Time Compute、Formal Theorem Proving
1. 引言
大型語言模型的 reasoning 能力可以透過增加 inference-time compute 改善。常見方法包括:
repeated sampling;
best-of- N N N ;
self-consistency;
search;
verifier reranking;
extended reasoning;
multi-agent decomposition。
然而,若所有輸入都使用相同計算量,系統會出現明顯浪費:
easy cases are over-computed, hard cases may remain under-computed. \boxed{
\text{easy cases are over-computed,}
\qquad
\text{hard cases may remain under-computed.}
} easy cases are over-computed, hard cases may remain under-computed.
2026 年的 adaptive test-time compute 研究已明確把問題寫成有限平均 budget 下的資源分配:哪些問題值得更多 sampling / reasoning,哪些問題可以低成本完成。Uncertainty-Aware Budget Allocation 以 per-question uncertainty 配置額外 sampling;Adaptive Test-Time Compute Allocation via Constrained Policy Optimization 則將 accuracy–cost trade-off 寫成 constrained optimization,再學習輸入到 compute action 的 allocation policy。ThinkBooster 更進一步將 test-time scaling strategy、scorer、quality–cost benchmark 與 deployable proxy service 整合成 runtime framework。
這些工作處理的是:
How much compute should this input receive? \boxed{
\text{How much compute should this input receive?}
} How much compute should this input receive?
本文研究另一個相鄰但不同的問題:
What kind of reasoning should receive the next unit of compute? \boxed{
\text{What kind of reasoning should receive the next unit of compute?}
} What kind of reasoning should receive the next unit of compute?
在研究級數學問題中,一個輸入可能同時存在:
theorem statement 本身的 formal uncertainty;
source formulation 是否選對的問題;
hidden quantifier / witness dependency;
framework bridge;
representation provenance;
natural-language perspective ambiguity。
因此,「多算一次」不是唯一 action。
下一個計算單位可能應該交給:
M M P F , M_{\mathrm{MPF}}, M MPF ,
也可能交給:
M S o u r c e F i d e l i t y , M_{\mathrm{SourceFidelity}}, M SourceFidelity ,
或:
M B r i d g e , M_{\mathrm{Bridge}}, M Bridge ,
甚至:
誰都不要叫,直接 Freeze。 \boxed{
\text{誰都不要叫,直接 Freeze。}
} 誰都不要叫,直接 Freeze 。
因此本文提出:
Doubt-Driven Reasoning Allocation . \boxed{
\text{Doubt-Driven Reasoning Allocation}.
} Doubt-Driven Reasoning Allocation .
2. 從 Compute Scaling 到 Cognitive Allocation
傳統 test-time compute allocation 可以抽象成:
x → b ( x ) , x
\rightarrow
b(x), x → b ( x ) ,
其中 b ( x ) b(x) b ( x ) 是輸入 x x x 的 computation budget。
若 action 僅為增加 sample 數:
b ( x ) ∈ { 1 , 2 , … , B } . b(x)\in\{1,2,\ldots,B\}. b ( x ) ∈ { 1 , 2 , … , B } .
本文將 action space 擴張為:
a t ∈ { Freeze , M 1 , M 2 , … , M k } . a_t
\in
\{
\operatorname{Freeze},
M_1,
M_2,
\ldots,
M_k
\}. a t ∈ { Freeze , M 1 , M 2 , … , M k } .
其中 M i M_i M i 不是單純不同 LLM,也可以是不同認知功能:
formal-structure auditor;
source-fidelity auditor;
NLU interpreter;
bridge auditor;
counterexample searcher;
premise retriever;
Lean prover;
CAS;
numerical experiment agent。
因此:
Compute Quantity Allocation ⊂ Cognitive Resource Allocation . \boxed{
\text{Compute Quantity Allocation}
\subset
\text{Cognitive Resource Allocation}.
} Compute Quantity Allocation ⊂ Cognitive Resource Allocation .
本文的核心不是假定每個 specialist 必須由不同模型執行。
同一個 foundation model 也可以透過:
不同 system prompt;
不同 tool access;
不同 context;
不同 objective;
不同 verifier;
扮演不同 specialist。
因此 specialist 是 functional role,而不是必然等於 model identity。
3. 與多 Agent 形式數學的關係
形式數學系統已開始自然出現角色分離。
MA-LoT 使用多 Agent 結構,把自然語言長鏈推理與 Lean formal-language verification 結合。LeanMarathon 則建立一個 multi-agent harness,讓不同 contract-scoped agents 構建、審計、證明與修復共享 blueprint,並由 orchestrator 先穩定 target fidelity,再處理 proof DAG。
這些系統顯示:
Mathematical reasoning ≠ one monolithic model call . \boxed{
\text{Mathematical reasoning}
\neq
\text{one monolithic model call}.
} Mathematical reasoning = one monolithic model call .
本文的延伸在於:
如果已經存在多種 Agent / cognitive roles,則上層系統還需要決定「何時叫誰」。
也就是:
Multi-Agent Architecture ≠ Adaptive Agent Allocation . \boxed{
\text{Multi-Agent Architecture}
\neq
\text{Adaptive Agent Allocation}.
} Multi-Agent Architecture = Adaptive Agent Allocation .
僅僅擁有五個 agents,不等於每題都應該呼叫五個 agents。
4. Residual Doubt Vector
令問題在時間 t t t 的狀態為:
S t . S_t. S t .
我們定義 residual doubt:
D t = ( D F , D S , D B , D P , D R , D C , … ) . D_t
=
(
D_F,
D_S,
D_B,
D_P,
D_R,
D_C,
\ldots
). D t = ( D F , D S , D B , D P , D R , D C , … ) .
可包含:
D F D_F D F :Formal-Structural Doubt;
D S D_S D S :Source-Fidelity Doubt;
D B D_B D B :Framework-Bridge Doubt;
D P D_P D P :Perspective / Pragmatic Doubt;
D R D_R D R :Representation Doubt;
D C D_C D C :Certificate / Closure Doubt。
其中:
D j ≥ 0. D_j\ge0. D j ≥ 0.
D j = 0 D_j=0 D j = 0 不必解讀為「哲學上完全無疑問」,而是:
在目前 theorem identity 與 proof obligation 下,此維度沒有剩餘的 load-bearing unresolved issue。
因此 residual doubt 是 operational state,而不是主觀信念的完整哲學模型。
5. Doubt Value Function
不同 doubt dimensions 的危險性不一定相同。
例如:
一個 minor notation ambiguity;
一個量詞 swap;
一個錯 source formulation;
其後果不同。
因此定義:
V ( D ) = ∑ j λ j D j . V(D)
=
\sum_j
\lambda_jD_j. V ( D ) = j ∑ λ j D j .
其中:
λ j ≥ 0 \lambda_j\ge0 λ j ≥ 0
是 risk weight。
更一般可以:
V : R ≥ 0 m → R ≥ 0 . V:
\mathbb R_{\ge0}^m
\rightarrow
\mathbb R_{\ge0}. V : R ≥ 0 m → R ≥ 0 .
而不要求線性。
例如若 source-fidelity doubt 與 formal doubt 同時高時風險超線性增加,可使用:
V ( D ) = ∑ j λ j D j + ∑ i < j γ i j D i D j . V(D)
=
\sum_j\lambda_jD_j
+
\sum_{i<j}\gamma_{ij}D_iD_j. V ( D ) = j ∑ λ j D j + i < j ∑ γ ij D i D j .
這反映:
doubt dimensions may interact . \boxed{
\text{doubt dimensions may interact}.
} doubt dimensions may interact .
6. Specialist Model
令 specialist set:
M = { M 1 , … , M k } . \mathcal M
=
\{M_1,\ldots,M_k\}. M = { M 1 , … , M k } .
每個 specialist 至少具有兩種估計。
6.1 Cost
C ( M i ∣ S t ) > 0. C(M_i\mid S_t)>0. C ( M i ∣ S t ) > 0.
成本可以是:
tokens;
wall-clock latency;
API cost;
tool calls;
proof nodes;
human supervision;
energy;
mixed operational cost。
初期可以使用 proxy cost,未來應改用 runtime telemetry。
6.2 Expected doubt transition
呼叫:
M i M_i M i
使:
D t → D t + 1 . D_t
\rightarrow
D_{t+1}. D t → D t + 1 .
定義:
R i ( D t ) = E [ D t − D t + 1 ∣ M i , S t ] . R_i(D_t)
=
\mathbb E[
D_t-D_{t+1}
\mid
M_i,S_t
]. R i ( D t ) = E [ D t − D t + 1 ∣ M i , S t ] .
這是一個向量。
例如 MPF:
R M P F ≈ ( r F , r S , r B , … ) R_{\mathrm{MPF}}
\approx
(
r_F,
r_S,
r_B,\ldots
) R MPF ≈ ( r F , r S , r B , … )
一般期望:
r F r_F r F
較高,
但不必假定:
r S = r B = 0. r_S=r_B=0. r S = r B = 0.
一個 specialist 可能順帶消除多種 doubt。
7. Value of Information
本文定義 specialist 的基本 VOI:
VOI ( M i ∣ D t ) = E [ V ( D t ) − V ( D t + 1 ) ∣ M i ] C ( M i ∣ S t ) . \boxed{
\operatorname{VOI}(M_i\mid D_t)
=
\frac{
\mathbb E[
V(D_t)-V(D_{t+1})
\mid M_i
]
}{
C(M_i\mid S_t)
}.
} VOI ( M i ∣ D t ) = C ( M i ∣ S t ) E [ V ( D t ) − V ( D t + 1 ) ∣ M i ] .
若使用線性 V V V :
VOI ( M i ∣ D t ) = ∑ j λ j E [ Δ D j ∣ M i ] C ( M i ) . \operatorname{VOI}(M_i\mid D_t)
=
\frac{
\sum_j
\lambda_j
\mathbb E[\Delta D_j\mid M_i]
}{
C(M_i)
}. VOI ( M i ∣ D t ) = C ( M i ) ∑ j λ j E [ Δ D j ∣ M i ] .
這表示:
每一單位認知成本,預期消掉多少 proof-relevant unresolved risk。
因此 routing:
M t ∗ = arg max M i VOI ( M i ∣ D t ) . \boxed{
M_t^\ast
=
\arg\max_{M_i}
\operatorname{VOI}(M_i\mid D_t).
} M t ∗ = arg M i max VOI ( M i ∣ D t ) .
8. VOI 不等於 Accuracy Predictor
這裡必須區分。
一般 inference allocation 可能估:
P ( correct ∣ x , b ) . P(\text{correct}\mid x,b). P ( correct ∣ x , b ) .
而本文主要估:
E [ Δ D ∣ M i , D ] . \mathbb E[\Delta D\mid M_i,D]. E [ Δ D ∣ M i , D ] .
兩者可能相關,但不相同。
例如:
Case A
MPF specialist 不直接提高 theorem proof probability,
但可以發現:
∀ x ∃ y \forall x\exists y ∀ x ∃ y
被誤寫成:
∃ y ∀ x . \exists y\forall x. ∃ y ∀ x .
它的價值是阻止 proof search 進入錯 target。
Case B
Source Fidelity specialist 完全不會做代數證明,
但可以發現:
AI 正在證 source 中較強的另一個 variant。
因此:
Value of a reasoning module ≠ its standalone solving accuracy . \boxed{
\text{Value of a reasoning module}
\neq
\text{its standalone solving accuracy}.
} Value of a reasoning module = its standalone solving accuracy .
這是 specialist orchestration 的重要原則。
9. Freeze Controller:Should I Think More?
只會選 specialist 還不夠。
如果每輪都必須選一個 M i M_i M i :
M t ∗ = arg max i VOI , M_t^\ast
=
\arg\max_i\operatorname{VOI}, M t ∗ = arg i max VOI ,
系統仍可能永遠推理。
因此 action set 必須包含:
Freeze . \boxed{
\operatorname{Freeze}.
} Freeze .
我們提出兩種基本停止條件。
9.1 State-Based Freeze
若:
V ( D t ) ≤ θ D , V(D_t)
\le
\theta_D, V ( D t ) ≤ θ D ,
則:
Freeze ( S t ) = 1. \operatorname{Freeze}(S_t)=1. Freeze ( S t ) = 1.
即 residual doubt 已低於可接受 threshold。
9.2 Marginal-Value Freeze
即使:
V ( D t ) > θ D , V(D_t)>\theta_D, V ( D t ) > θ D ,
剩餘疑點也可能成本極高、影響極低。
如果:
max i VOI ( M i ∣ D t ) ≤ θ V O I , \max_i
\operatorname{VOI}(M_i\mid D_t)
\le
\theta_{\mathrm{VOI}}, i max VOI ( M i ∣ D t ) ≤ θ VOI ,
則可以 conditional freeze。
因此:
Freeze ⟺ [ V ( D t ) ≤ θ D ] ∨ [ max i VOI i ≤ θ V O I ] , \boxed{
\operatorname{Freeze}
\iff
\left[
V(D_t)\le\theta_D
\right]
\lor
\left[
\max_i\operatorname{VOI}_i
\le\theta_{\mathrm{VOI}}
\right],
} Freeze ⟺ [ V ( D t ) ≤ θ D ] ∨ [ i max VOI i ≤ θ VOI ] ,
但第二種 freeze 必須記錄 remaining debt。
10. Hard Constraints:不是所有風險都可用低 VOI 忽略
純 VOI 最大化可能出現危險。
例如一個 quantifier corruption 的發現成本很高,但若漏掉,其 consequence 非常大。
因此定義 load-bearing set:
L ( D t ) . \mathcal L(D_t). L ( D t ) .
若:
D j ∈ L D_j\in\mathcal L D j ∈ L
且:
D j > θ j h a r d , D_j>\theta_j^{\mathrm{hard}}, D j > θ j hard ,
則禁止 freeze。
即:
FreezeEligible = SoftStop ∧ HardClosure . \boxed{
\operatorname{FreezeEligible}
=
\operatorname{SoftStop}
\land
\operatorname{HardClosure}.
} FreezeEligible = SoftStop ∧ HardClosure .
其中:
HardClosure = ⋀ j ∈ L [ D j ≤ θ j h a r d ] . \operatorname{HardClosure}
=
\bigwedge_{j\in\mathcal L}
[
D_j\le\theta_j^{\mathrm{hard}}
]. HardClosure = j ∈ L ⋀ [ D j ≤ θ j hard ] .
這使系統不會因成本太高就跳過 theorem identity 的核心風險。
11. Risk-Aware VOI
更一般地,可以把 catastrophic target error 直接放入價值函數。
若 specialist M i M_i M i 有機率發現 fatal corruption:
p i f a t a l , p_i^{\mathrm{fatal}}, p i fatal ,
漏掉此 corruption 的 expected loss:
L f a t a l , L_{\mathrm{fatal}}, L fatal ,
則:
VOI r i s k ( M i ) = E [ Δ V ] + p i f a t a l L f a t a l C ( M i ) . \operatorname{VOI}_{\mathrm{risk}}(M_i)
=
\frac{
\mathbb E[\Delta V]
+
p_i^{\mathrm{fatal}}L_{\mathrm{fatal}}
}{
C(M_i)
}. VOI risk ( M i ) = C ( M i ) E [ Δ V ] + p i fatal L fatal .
如此,「平常不太有用,但偶爾能抓到致命 target error」的 specialist 不會被簡單平均效益淘汰。
12. Adaptive Routing Algorithm
基本 runtime:
INPUT: source problem / candidate target
1. SURFACE AUDIT
2. estimate residual doubt D_t
3. check hard closure
4. if freeze eligible:
FREEZE
else:
estimate VOI(M_i | D_t)
choose argmax VOI
run specialist
update state and D_t
repeat
5. output frozen target
6. hand off to MCDM / Proof Router
形式化:
S 0 → D 0 → M 0 ∗ → S 1 → D 1 → ⋯ → C f r e e z e ∗ . S_0
\rightarrow
D_0
\rightarrow
M_0^\ast
\rightarrow
S_1
\rightarrow
D_1
\rightarrow
\cdots
\rightarrow
C^\ast_{\mathrm{freeze}}. S 0 → D 0 → M 0 ∗ → S 1 → D 1 → ⋯ → C freeze ∗ .
13. Specialist Roles
本文至少區分四個 proof-search 前 specialist。
13.1 Surface Auditor
成本最低。
檢查:
明顯 domain omission;
source 與 target 的 literal mismatch;
顯式 boundary;
已充分指定的 clean target。
若 Surface 已足夠:
do not call deeper specialists . \boxed{
\text{do not call deeper specialists}.
} do not call deeper specialists .
13.2 MPF Specialist
檢查:
domain;
quantifier;
hidden/transitive quantifier;
witness dependency;
uniformity;
asymptotic semantics;
certificate tree。
13.3 Source-Fidelity Specialist
檢查:
source formulation family;
target-strength relation;
definition provenance;
representation equivalence;
benchmark target identity。
13.4 Bridge Specialist
檢查:
probability → \rightarrow → decision;
simulation → \rightarrow → theorem;
empirical observation → \rightarrow → universal mathematical claim;
one model/framework → \rightarrow → another。
14. 為何不是 Always-Dual?
一個直覺安全策略是:
每題全部跑 Surface + MPF + NLU + Source + Bridge。
稱為:
AlwaysDual . \operatorname{AlwaysDual}. AlwaysDual .
其優勢是高 coverage。
但成本:
C A l w a y s = ∑ i = 1 k C ( M i ) C_{\mathrm{Always}}
=
\sum_{i=1}^{k}C(M_i) C Always = i = 1 ∑ k C ( M i )
每題都支付。
若大量問題其實 Surface 就可 freeze:
Always-Dual creates systematic over-expansion . \boxed{
\text{Always-Dual creates systematic over-expansion}.
} Always-Dual creates systematic over-expansion .
所以真正目標不是:
max analysis depth , \max\text{analysis depth}, max analysis depth ,
而是:
min C a u d i t \boxed{
\min C_{\mathrm{audit}}
} min C audit
subject to:
P ( accept corrupted target ) ≤ ϵ . \boxed{
P(\text{accept corrupted target})
\le\epsilon.
} P ( accept corrupted target ) ≤ ϵ .
15. 與 Adaptive Test-Time Compute 的關係
Uncertainty-aware allocation 已顯示,固定 sampling budget 平均分配到所有問題並非理想策略;依 uncertainty 配置 budget 可以改善有限計算資源下的 reasoning performance。
Constrained Policy Optimization 類方法則可寫成:
max π E [ Accuracy ] \max_\pi
\mathbb E[\operatorname{Accuracy}] π max E [ Accuracy ]
subject to:
E [ C ] ≤ B . \mathbb E[C]\le B. E [ C ] ≤ B .
本文可類比寫為:
min π E [ R t a r g e t + α C ] \min_\pi
\mathbb E[
\mathcal R_{\mathrm{target}}
+
\alpha C
] π min E [ R target + α C ]
或:
max π E [ TargetFidelity ] \max_\pi
\mathbb E[
\operatorname{TargetFidelity}
] π max E [ TargetFidelity ]
subject to:
E [ C a u d i t ] ≤ B . \mathbb E[C_{\mathrm{audit}}]
\le B. E [ C audit ] ≤ B .
差異是 action 不只是 budget level:
a ≠ more samples . a\neq\text{more samples}. a = more samples .
而是:
a ∈ { Freeze , M M P F , M S o u r c e , M B r i d g e , … } . a\in\{
\operatorname{Freeze},
M_{\mathrm{MPF}},
M_{\mathrm{Source}},
M_{\mathrm{Bridge}},
\ldots
\}. a ∈ { Freeze , M MPF , M Source , M Bridge , … } .
因此 DDRA 可以理解為:
structured test-time compute allocation over cognitive roles . \boxed{
\text{structured test-time compute allocation over cognitive roles}.
} structured test-time compute allocation over cognitive roles .
16. 與 Selective Verification 的關係
近期 selective verification 工作指出,在 search 中平均驗證所有 intermediate hypotheses 會浪費 verifier calls;應優先驗證較具資訊價值的狀態。
本文的精神相似:
not every unresolved branch deserves verification . \boxed{
\text{not every unresolved branch deserves verification}.
} not every unresolved branch deserves verification .
但我們把 allocation point 放得更上游:
selective verification:選擇哪個 reasoning state 要驗證;
DDRA:選擇哪個 problem-understanding / formalization specialist 值得先投入。
兩者未來可以組合:
Problem-Space VOI → Proof-State VOI . \text{Problem-Space VOI}
\rightarrow
\text{Proof-State VOI}. Problem-Space VOI → Proof-State VOI .
17. EXP-0005:Freeze Controller 的結構代理結果
在先前開發集的結構模擬中,我們比較:
Surface-only;
Always-MPF;
Always-Dual;
Adaptive Controller。
以每個 audit module invocation 為一個 proxy cost unit。
結果顯示,在該人工建立 development set 上:
Surface-only 成本最低,但漏掉大量深層 target corruption;
Always-MPF 能抓 formal-structure corruption,但漏掉 source-fidelity 與 bridge 類;
Always-Dual coverage 完整,但存在大量 unnecessary module calls;
Adaptive Controller 在相同人工 coverage 下,用顯著更少 module invocations 完成 audit。
其中 Adaptive 相較 Always-Dual 的 module-call proxy 約降低:
57.7 % . 57.7\%. 57.7%.
但:
57.7 % is not measured compute saving . \boxed{
57.7\%
\text{ is not measured compute saving}.
} 57.7% is not measured compute saving .
它只表示在人工 annotation 與 unit-cost assumption 下,啟動的分析模組較少。
18. EXP-0006:Holdout-Style Generalization
為避免控制器只記住 development cases,後續換用新的 Formal Conjectures 問題類型:
Erdős 12;
Erdős 26;
Erdős 56;
Open Quantum Problem 13;
Open Quantum Problem 35;
OEIS A6697。
這些案例新增:
optimization-target replacement;
representation-coverage failure;
definition substitution / provenance gap。
固定 controller policy 後的 holdout-style 結構測試顯示,routing taxonomy 可以自然把:
quantifier / uniformity → \rightarrow → MPF;
source identity / provenance → \rightarrow → Source Fidelity;
explicit boundary → \rightarrow → Surface;
representation coverage → \rightarrow → MPF / Source Fidelity。
但也暴露 rule-based controller 的保守性:
一些 case 會同時啟動兩個 specialist,即使一個 specialist 已可能足夠。
這直接導向 VOI scheduler。
19. EXP-0007:VOI Specialist Scheduler
在下一步 simulation 中,為 specialist 指定 proxy costs:
C S u r f a c e = 1 , C_{\mathrm{Surface}}=1, C Surface = 1 ,
C M P F = 2 , C_{\mathrm{MPF}}=2, C MPF = 2 ,
C S o u r c e = 1.3 , C_{\mathrm{Source}}=1.3, C Source = 1.3 ,
C B r i d g e = 1.2. C_{\mathrm{Bridge}}=1.2. C Bridge = 1.2.
並以人工 residual-doubt profiles 估計各 specialist 的 expected doubt reduction。
得到:
VOI ( M i ) = E [ Δ D ] C i . \operatorname{VOI}(M_i)
=
\frac{\mathbb E[\Delta D]}{C_i}. VOI ( M i ) = C i E [ Δ D ] .
在相同 holdout-style sample 上,VOI routing 能避免部分 rule-based 的保守 double-routing。
代理結果:
Always-Dual:48 module calls;
Rule-Based Adaptive:26;
VOI Scheduler:24。
VOI scheduler 的 weighted proxy cost 相對 Always-Dual 約降低:
51.8 % . 51.8\%. 51.8%.
相對 Rule-Based Adaptive 約降低:
11.2 % . 11.2\%. 11.2%.
與利用 gold information 才能定義的 oracle lower bound 相比,proxy overhead 約:
4.3 % . 4.3\%. 4.3%.
再次強調:
這些全部是 controller simulation,不是實際 LLM / Lean 性能結果。 \boxed{
\text{這些全部是 controller simulation,}
\text{不是實際 LLM / Lean 性能結果。}
} 這些全部是 controller simulation , 不是實際 LLM / Lean 性能結果。
20. 一個重要結果:Routing 可以是序列,而不是一次分類
有些問題:
D F ≫ 0 , D S ≫ 0. D_F\gg0,
\qquad
D_S\gg0. D F ≫ 0 , D S ≫ 0.
此時可能:
M S o u r c e → M M P F → Freeze . M_{\mathrm{Source}}
\rightarrow
M_{\mathrm{MPF}}
\rightarrow
\operatorname{Freeze}. M Source → M MPF → Freeze .
例如 representation fidelity 問題:
Source specialist 先判定某 representation 是否可能忠實;
MPF 再檢查 symmetry / quantifier coverage 是否完整。
因此:
Routing ≠ one-shot classification . \boxed{
\text{Routing}
\neq
\text{one-shot classification}.
} Routing = one-shot classification .
更像:
D t → M t → D t + 1 → M t + 1 . D_t
\rightarrow
M_t
\rightarrow
D_{t+1}
\rightarrow
M_{t+1}. D t → M t → D t + 1 → M t + 1 .
也就是 closed-loop reasoning control。
21. Problem-Level Scheduler 與 Proof-Level Scheduler
本文主要處理 proof search 前。
但完整數學 AI 可以有兩級 scheduler。
Level A:Problem-Space Scheduler
決定:
target 是否穩定;
該叫哪個 semantic/formal specialist;
是否 Freeze。
Level B:Proof-Space Scheduler
Target freeze 後決定:
Lean search;
premise retrieval;
counterexample;
CAS;
numerical experiment;
natural-language lemma planning;
ATP。
因此:
Problem-Space Governance → Proof-Space Governance . \boxed{
\text{Problem-Space Governance}
\rightarrow
\text{Proof-Space Governance}.
} Problem-Space Governance → Proof-Space Governance .
這避免 proof-level agent 被要求替上游 target corruption 買單。
22. 與 MCDM 的接口
MCDM 類 difficulty router 用來評估 frozen problem 的研究障礙向量。
因此順序應是:
C f r e e z e ∗ → M C D M → R e s e a r c h R o u t e r . C^\ast_{\mathrm{freeze}}
\rightarrow
\mathrm{MCDM}
\rightarrow
\mathrm{ResearchRouter}. C freeze ∗ → MCDM → ResearchRouter .
而不是:
C r a w → M C D M . C_{\mathrm{raw}}
\rightarrow
\mathrm{MCDM}. C raw → MCDM .
因為如果 candidate target 本身不穩定:
M C D M ( C 1 ) ≠ M C D M ( C 2 ) . \mathrm{MCDM}(C_1)
\neq
\mathrm{MCDM}(C_2). MCDM ( C 1 ) = MCDM ( C 2 ) .
甚至:
Var [ M C D M ( C i ) ] ≫ 0 \operatorname{Var}
[
\mathrm{MCDM}(C_i)
]
\gg0 Var [ MCDM ( C i )] ≫ 0
可以成為 target-instability diagnostic。
因此 DDRA 位於 MCDM 的上游。
23. 可學習的 Routing Policy
目前的 C ( M i ) C(M_i) C ( M i ) 與 R i ( D ) R_i(D) R i ( D ) 可以人工指定。
但真正 Runtime 應由 telemetry 學習:
C ^ t ( M i ) \widehat C_t(M_i) C t ( M i )
以及:
R ^ t ( M i ∣ X , D ) . \widehat R_t(M_i\mid X,D). R t ( M i ∣ X , D ) .
其中 X X X 包含:
problem domain;
theorem form;
source type;
model version;
prompt version;
tool availability。
於是:
VOI ^ t ( M i ) = E ^ [ Δ V ] C ^ t ( M i ) . \widehat{\operatorname{VOI}}_t(M_i)
=
\frac{
\widehat{\mathbb E}[\Delta V]
}{
\widehat C_t(M_i)
}. VOI t ( M i ) = C t ( M i ) E [ Δ V ] .
這將是本系列下一篇 self-calibrating orchestration 的核心。
24. 評測指標
未來真實 benchmark 至少應保存:
24.1 Target Safety
Corruption Recall , \text{Corruption Recall}, Corruption Recall ,
False Freeze Rate , \text{False Freeze Rate}, False Freeze Rate ,
Clean Freeze Rate . \text{Clean Freeze Rate}. Clean Freeze Rate .
24.2 Routing Efficiency
N m o d u l e c a l l s , N_{\mathrm{module\ calls}}, N module calls ,
tokens , \text{tokens}, tokens ,
latency , \text{latency}, latency ,
tool calls . \text{tool calls}. tool calls .
24.3 Proof-Downstream Effect
proof branches , \text{proof branches}, proof branches ,
premise retrieval count , \text{premise retrieval count}, premise retrieval count ,
Lean elaboration failures , \text{Lean elaboration failures}, Lean elaboration failures ,
proof search nodes , \text{proof search nodes}, proof search nodes ,
proof success . \text{proof success}. proof success .
24.4 Regret
若存在 oracle 或 hindsight estimate:
R T = ∑ t = 1 T [ V ( a t ∗ ) − V ( a t ) ] . R_T
=
\sum_{t=1}^{T}
[
V(a_t^\ast)-V(a_t)
]. R T = t = 1 ∑ T [ V ( a t ∗ ) − V ( a t )] .
24.5 Over-Expansion
OER = unnecessary specialist calls all specialist calls . \operatorname{OER}
=
\frac{
\text{unnecessary specialist calls}
}{
\text{all specialist calls}
}. OER = all specialist calls unnecessary specialist calls .
25. 三個核心 Hypotheses
H1:Adaptive Freeze
相較 Always-Dual:
E [ C a u d i t ] \mathbb E[C_{\mathrm{audit}}] E [ C audit ]
應下降,而 target-corruption miss rate 不顯著上升。
H2:VOI Routing
相較固定 rule-based routing:
VOI Scheduler \operatorname{VOI\ Scheduler} VOI Scheduler
應降低 over-expansion。
H3:Downstream Search
若上游 target audit 避免了 wrong target / wrong certificate route:
E [ C p r o o f ∣ D D R A ] < E [ C p r o o f ∣ D i r e c t P r o o f ] . \mathbb E[
C_{\mathrm{proof}}
\mid
\mathrm{DDRA}
]
<
\mathbb E[
C_{\mathrm{proof}}
\mid
\mathrm{DirectProof}
]. E [ C proof ∣ DDRA ] < E [ C proof ∣ DirectProof ] .
但這是 empirical hypothesis,不是本文已證結果。
26. 失敗模式
26.1 Wrong Doubt Estimation
若:
D t D_t D t
估錯,
則 VOI 全部跟著錯。
26.2 Wrong Specialist Capability Model
若:
R i ( D ) R_i(D) R i ( D )
估得太樂觀,會過度調用某 specialist。
26.3 Cheap-but-Useless Bias
若成本權重過重:
arg max VOI \arg\max\operatorname{VOI} arg max VOI
可能偏向便宜但只能消除小疑點的 specialist。
因此需 hard closure / fatal-risk term。
26.4 Endless Audit
若:
θ D \theta_D θ D
過低,
系統可能永遠不 Freeze。
26.5 Premature Freeze
若:
θ D \theta_D θ D
過高,
則重新產生 Paper 02 的 premature closure。
26.6 Correlated Specialists
兩個 specialists 可能高度冗餘:
R i ≈ R j . R_i\approx R_j. R i ≈ R j .
若 scheduler 不考慮 conditional marginal value,可能重複付費。
更完整應估:
VOI ( M j ∣ D , M i already run ) . \operatorname{VOI}
(
M_j
\mid
D,
M_i\text{ already run}
). VOI ( M j ∣ D , M i already run ) .
27. 從 Solver 到 Orchestrator
傳統 AI solver:
Problem → Answer . \boxed{
\text{Problem}
\rightarrow
\text{Answer}.
} Problem → Answer .
Agent system:
Problem → Agents → Answer . \boxed{
\text{Problem}
\rightarrow
\text{Agents}
\rightarrow
\text{Answer}.
} Problem → Agents → Answer .
本文提出的 adaptive orchestrator:
Problem → D t → Choose Specialist → D t + 1 → ⋯ → Freeze . \boxed{
\text{Problem}
\rightarrow
D_t
\rightarrow
\text{Choose Specialist}
\rightarrow
D_{t+1}
\rightarrow
\cdots
\rightarrow
\operatorname{Freeze}.
} Problem → D t → Choose Specialist → D t + 1 → ⋯ → Freeze .
因此上層 AI 真正需要回答:
Who should think? \boxed{
\text{Who should think?}
} Who should think?
What should they think about? \boxed{
\text{What should they think about?}
} What should they think about?
以及:
When should everyone stop thinking? \boxed{
\text{When should everyone stop thinking?}
} When should everyone stop thinking?
28. 認知資源治理
這使推理系統可以分層:
L 0 : Solver / Worker L 1 : Specialist Reasoner L 2 : Adaptive Orchestrator L 3 : Epistemic Governor \begin{array}{ll}
L_0 &: \text{Solver / Worker}\\
L_1 &: \text{Specialist Reasoner}\\
L_2 &: \text{Adaptive Orchestrator}\\
L_3 &: \text{Epistemic Governor}
\end{array} L 0 L 1 L 2 L 3 : Solver / Worker : Specialist Reasoner : Adaptive Orchestrator : Epistemic Governor
L 0 L_0 L 0 解局部問題。
L 1 L_1 L 1 專門處理某種認知功能。
L 2 L_2 L 2 決定 specialist allocation。
L 3 L_3 L 3 管理:
target 是否值得 Freeze;
哪些 unresolved risks 不可忽略;
推理預算怎麼分;
哪些認知活動目前沒有邊際價值。
因此:
Tool Routing ⊂ Reasoning Routing ⊂ Problem-Space Governance . \boxed{
\text{Tool Routing}
\subset
\text{Reasoning Routing}
\subset
\text{Problem-Space Governance}.
} Tool Routing ⊂ Reasoning Routing ⊂ Problem-Space Governance .
本文主要建立中間層:
Reasoning Routing . \boxed{
\text{Reasoning Routing}.
} Reasoning Routing .
29. 討論
Adaptive test-time compute 已經提出一個重要觀念:
inference compute 應是可分配資源,而不是固定常數。
本文再推一步:
reasoning type 本身也應是可分配資源。
對數學 AI 而言,額外的一千 tokens 可以拿來:
繼續同一路推 proof;
回頭檢查量詞;
找 source;
檢查 provenance;
反例搜尋;
形式驗證。
這些 action 的 information gain 完全不同。
因此真正的 inference-time intelligence 不只是:
think longer . \boxed{
\text{think longer}.
} think longer .
更是:
think in the right place . \boxed{
\text{think in the right place}.
} think in the right place .
而更成熟時,還包括:
know when not to think further . \boxed{
\text{know when not to think further}.
} know when not to think further .
30. 結論
本文提出 Doubt-Driven Reasoning Allocation,將數學 AI 的 proof-search 前資源治理建模為 residual doubt 上的 sequential decision problem。
核心狀態:
D t . D_t. D t .
核心 action:
a t ∈ { Freeze , M 1 , … , M k } . a_t
\in
\{
\operatorname{Freeze},
M_1,\ldots,M_k
\}. a t ∈ { Freeze , M 1 , … , M k } .
核心選擇準則:
VOI ( M i ∣ D t ) = E [ V ( D t ) − V ( D t + 1 ) ] C ( M i ) . \boxed{
\operatorname{VOI}(M_i\mid D_t)
=
\frac{
\mathbb E[
V(D_t)-V(D_{t+1})
]
}{
C(M_i)
}.
} VOI ( M i ∣ D t ) = C ( M i ) E [ V ( D t ) − V ( D t + 1 )] .
核心停止原則:
Freeze when residual risk is sufficiently closed, \boxed{
\text{Freeze when residual risk is sufficiently closed,}
} Freeze when residual risk is sufficiently closed,
或:
when no remaining reasoning action has adequate marginal value, \boxed{
\text{when no remaining reasoning action has adequate marginal value,}
} when no remaining reasoning action has adequate marginal value,
但 load-bearing target risks 必須滿足 hard closure。
因此:
Should I think more? \boxed{
\text{Should I think more?}
} Should I think more?
由 Freeze Controller 回答,
而:
What should I think about next? \boxed{
\text{What should I think about next?}
} What should I think about next?
由 VOI Scheduler 回答。
這使數學 AI 從固定 pipeline:
Formalize → Prove \text{Formalize}
\rightarrow
\text{Prove} Formalize → Prove
轉向:
Observe → Estimate Doubt → Allocate Reasoning → Update → Freeze → Prove . \boxed{
\text{Observe}
\rightarrow
\text{Estimate Doubt}
\rightarrow
\text{Allocate Reasoning}
\rightarrow
\text{Update}
\rightarrow
\text{Freeze}
\rightarrow
\text{Prove}.
} Observe → Estimate Doubt → Allocate Reasoning → Update → Freeze → Prove .
本文的結構代理實驗只證明該控制流程可以被工程化與量測,而不證明它已在真實模型上降低 inference 或 formal-proof 成本。下一篇將進一步研究如何從 runtime telemetry 自動學習 specialist cost、能力與 routing policy,使 orchestrator 從固定規則提升為 self-calibrating meta-orchestrator。
參考文獻
Nguyen, M., Gupta, S., Le, H. Uncertainty-Aware Budget Allocation for Adaptive Test-Time Reasoning. arXiv:2605.26849, 2026.
Zhai, Z., Li, B., Xiao, B., Li, M., Wang, X. Adaptive Test-Time Compute Allocation for Reasoning LLMs via Constrained Policy Optimization. arXiv:2604.14853, 2026.
Qu, S. Adaptive Test-Time Compute Allocation via Learned Heuristics over Categorical Structure. arXiv:2602.03975, 2026.
Smirnov, V., Nguyen, C., Senichev, S., et al. ThinkBooster: A Unified Framework for Seamless Test-Time Scaling of LLM Reasoning. arXiv:2606.06915, 2026.
Zhang, Y., Sun, Y., Suzuki, T., Lee, J. D., Liu, F. LeanMarathon: Toward Reliable AI Co-Mathematicians through Long-Horizon Lean Autoformalization. arXiv:2606.05400, 2026.
Wang, R., Pan, R., Li, Y., et al. MA-LoT: Multi-Agent Lean-based Long Chain-of-Thought Reasoning Enhances Formal Theorem Proving. arXiv:2503.03205, 2025.
Chen, J., Chen, W., Du, J., et al. Seed-Prover 1.5: Mastering Undergraduate-Level Theorem Proving via Learning from Experience. arXiv:2512.17260, 2025.
Firsching, M., Lezeau, P., Mercuri, S., et al. Formal Conjectures: An Open and Evolving Benchmark for Verified Discovery in Mathematics. arXiv:2605.13171, 2026.
研究狀態聲明
本文提出的 Doubt-Driven Reasoning Allocation、Residual Doubt Vector、Freeze Controller、Specialist VOI、Hard Closure 與相關 routing 架構,屬於本文提出的研究框架。
本文所述 EXP-0005~EXP-0007 數值來自人工 annotation、結構代理成本與 controller simulation;它們不應被引用為實際 LLM token savings、wall-clock speedup、Lean proof-node reduction 或通用性能提升。這些 empirical claims 必須由獨立 Agent、固定模型版本、真實 runtime telemetry 與 formal prover logs 重新驗證。