自我校準與漂移感知的認知編排器:從推理路由到 Meta-Orchestration
Self-Calibrating and Drift-Aware Cognitive Orchestration: From Reasoning Routing to Meta-Orchestration
作者: Neo.K機構: EveMissLab/一言諾科技有限公司日期: 2026-08-11
摘要
當 AI 系統由單一模型推理逐步演化為多模型、多 Agent、多工具與多專家角色協同時,一個新的系統層問題開始出現:上層 orchestrator 如何知道哪個 specialist 現在最值得被呼叫,而且如何知道自己對 specialist 的認識是否已經過期?
前一篇工作以 residual doubt 與 Value of Information(VOI)為核心,將 proof-search 前的推理資源配置建模為 specialist routing 問題。然而,若 specialist cost、能力與 runtime behavior 被視為固定常數,則 routing policy 只能在靜態世界中成立。實際部署中,模型版本、prompt、tool、hardware、API pricing、latency、retrieval quality 與 provider reliability 都可能改變,使:
P t ( C , R ∣ M ) ≠ P t + 1 ( C , R ∣ M ) , P_t(C,R\mid M)
\neq
P_{t+1}(C,R\mid M), P t ( C , R ∣ M ) = P t + 1 ( C , R ∣ M ) ,
其中 C C C 表示 specialist cost, R R R 表示其對 residual doubt 的 reduction capability。
本文提出 Self-Calibrating and Drift-Aware Cognitive Orchestration(SDCO) 。系統透過 runtime telemetry 持續估計:
C ^ t ( M i ) , R ^ t ( M i ∣ X , D ) , \widehat C_t(M_i),
\qquad
\widehat R_t(M_i\mid X,D), C t ( M i ) , R t ( M i ∣ X , D ) ,
並更新:
VOI ^ t ( M i ∣ D t ) = E ^ [ Δ V ∣ M i , D t ] C ^ t ( M i ) . \widehat{\operatorname{VOI}}_t(M_i\mid D_t)
=
\frac{
\widehat{\mathbb E}[\Delta V\mid M_i,D_t]
}{
\widehat C_t(M_i)
}. VOI t ( M i ∣ D t ) = C t ( M i ) E [ Δ V ∣ M i , D t ] .
由於只利用 exploitation 會造成 selection bias 與 cold-start lock-in,本文加入 uncertainty bonus:
Score t ( M i ) = VOI ^ t ( M i ) + β U t ( M i ) , \operatorname{Score}_t(M_i)
=
\widehat{\operatorname{VOI}}_t(M_i)
+
\beta U_t(M_i), Score t ( M i ) = VOI t ( M i ) + β U t ( M i ) ,
形成 exploration–exploitation tension。進一步,當 specialist behavior 發生 non-stationary drift 時,引入 staleness state A t ( M ) A_t(M) A t ( M ) 、forgetting factor、change-point detection 與 local reset / re-exploration。
本文將上層 specialist state 表示為:
S M ( t ) = ( C ^ t , R ^ t , U t , A t , V t ) , \mathcal S_M(t)
=
(
\widehat C_t,
\widehat R_t,
U_t,
A_t,
V_t
), S M ( t ) = ( C t , R t , U t , A t , V t ) ,
其中 V t V_t V t 記錄 model / prompt / runtime version。透過 EXP-0008~EXP-0010 的 runtime telemetry contract、self-calibration plumbing test 與 non-stationary synthetic simulation,我們驗證了從 raw telemetry 到 posterior update、VOI routing、shift detection 與 recalibration 的工程閉環可以被實作。所有實驗均為 synthetic / proxy simulation,本文不宣稱已證明真實模型上的效率提升。
本文的核心主張是:成熟的 AI orchestrator 不只應管理 specialist,也應管理自己對 specialist 的認識 。因此 meta-orchestration 的研究問題不再只是:
Who should think next? \text{Who should think next?} Who should think next?
而是:
How should the system learn, forget, and revise its beliefs about who should think next? \boxed{
\text{How should the system learn, forget, and revise its beliefs about who should think next?}
} How should the system learn, forget, and revise its beliefs about who should think next?
關鍵詞: Meta-Orchestration、Self-Calibration、Agent Routing、Concept Drift、Runtime Telemetry、Contextual Bandit、Exploration–Exploitation、Staleness、Change-Point Detection、Mathematical AI
1. 引言
多模型與多 Agent 系統的基本 routing 問題可以寫成:
x → M ∗ ( x ) , x
\rightarrow
M^\ast(x), x → M ∗ ( x ) ,
即對輸入 x x x 選擇最適合的模型或 Agent。
近年的 LLM routing 研究已逐步從 static router 走向 adaptive router。FlyRoute 透過實際流量累積 success evidence、自動重建 agent capability profiles,並針對 under-profiled agents 進行 targeted exploration;OrcaRouter 使用 hybrid offline–online learning 與 contextual bandit,在部署時只更新被選中模型的 arm;ParetoBandit 則直接把 production routing 建模為 non-stationary quality–cost problem,使用 forgetting mechanism、budget pacing 與 runtime hot-swap 來處理價格、品質與模型集合的變動。
這些研究顯示:
Router itself must be adaptive. \boxed{
\text{Router}
\text{ itself must be adaptive.}
} Router itself must be adaptive.
本文接受這個方向,但研究對象不同。
一般 model router 問:
哪個模型回答這個 query 的 quality–cost trade-off 最好?
本文研究:
哪個認知 specialist 對當前 residual doubt 最有價值?
例如:
MPF specialist;
Source-Fidelity specialist;
Bridge auditor;
NLU interpreter;
premise retriever;
Lean prover;
counterexample agent。
因此 routing target 不只是 output quality,而是:
expected reduction of proof-relevant uncertainty . \boxed{
\text{expected reduction of proof-relevant uncertainty}.
} expected reduction of proof-relevant uncertainty .
這使 orchestrator 本身變成一個需要持續學習的 epistemic control system。
2. 從 Adaptive Routing 到 Meta-Orchestration
前一篇 DDRA 架構定義:
D t = ( D F , D S , D B , … ) D_t
=
(D_F,D_S,D_B,\ldots) D t = ( D F , D S , D B , … )
與:
VOI ( M i ∣ D t ) = E [ Δ V ∣ M i , D t ] C ( M i ) . \operatorname{VOI}(M_i\mid D_t)
=
\frac{
\mathbb E[\Delta V\mid M_i,D_t]
}{
C(M_i)
}. VOI ( M i ∣ D t ) = C ( M i ) E [ Δ V ∣ M i , D t ] .
如果:
C ( M i ) C(M_i) C ( M i )
與:
R i ( D ) = E [ D t − D t + 1 ∣ M i , D t ] R_i(D)
=
\mathbb E[D_t-D_{t+1}\mid M_i,D_t] R i ( D ) = E [ D t − D t + 1 ∣ M i , D t ]
都是人工常數,
則 scheduler 本質上仍是 static policy。
本文改為:
C ( M i ) → C ^ t ( M i ) , \boxed{
C(M_i)
\rightarrow
\widehat C_t(M_i),
} C ( M i ) → C t ( M i ) ,
以及:
R i ( D ) → R ^ t ( M i ∣ X t , D t , H t ) , \boxed{
R_i(D)
\rightarrow
\widehat R_t(M_i\mid X_t,D_t,\mathcal H_t),
} R i ( D ) → R t ( M i ∣ X t , D t , H t ) ,
其中:
X t X_t X t :problem / runtime context;
H t \mathcal H_t H t :歷史 telemetry。
因此:
Reasoning Routing → Learning-to-Route Reasoning . \boxed{
\text{Reasoning Routing}
\rightarrow
\text{Learning-to-Route Reasoning}.
} Reasoning Routing → Learning-to-Route Reasoning .
本文稱此層為:
Meta-Orchestration . \boxed{
\text{Meta-Orchestration}.
} Meta-Orchestration .
3. Runtime Telemetry 是上層學習的基礎
如果沒有 observability,orchestrator 只能依賴人工印象。
因此需要保留每次 specialist invocation 的 raw telemetry。
令事件:
e t = ( M i , X t , D t , D t + 1 , U t , L t , P t , Q t , … ) , e_t
=
(
M_i,
X_t,
D_t,
D_{t+1},
U_t,
L_t,
P_t,
Q_t,
\ldots
), e t = ( M i , X t , D t , D t + 1 , U t , L t , P t , Q t , … ) ,
至少包括:
Model usage
T p r o m p t , T c o m p l e t i o n , T r e a s o n i n g , T c a c h e d . T_{\mathrm{prompt}},
\quad
T_{\mathrm{completion}},
\quad
T_{\mathrm{reasoning}},
\quad
T_{\mathrm{cached}}. T prompt , T completion , T reasoning , T cached .
Runtime latency
L w a l l , L T T F T , L q u e u e . L_{\mathrm{wall}},
\quad
L_{\mathrm{TTFT}},
\quad
L_{\mathrm{queue}}. L wall , L TTFT , L queue .
Tool usage
N t o o l , N w e b , N f i l e , N p y t h o n . N_{\mathrm{tool}},
\quad
N_{\mathrm{web}},
\quad
N_{\mathrm{file}},
\quad
N_{\mathrm{python}}. N tool , N web , N file , N python .
Proof telemetry
N p r e m i s e , N b r a n c h , N n o d e , N t a c t i c , N e l a b − f a i l , T p r o v e r . N_{\mathrm{premise}},
\quad
N_{\mathrm{branch}},
\quad
N_{\mathrm{node}},
\quad
N_{\mathrm{tactic}},
\quad
N_{\mathrm{elab-fail}},
\quad
T_{\mathrm{prover}}. N premise , N branch , N node , N tactic , N elab − fail , T prover .
Epistemic telemetry
D t , D t + 1 , Freeze / Reject / Prove state . D_t,
\qquad
D_{t+1},
\qquad
\text{Freeze / Reject / Prove state}. D t , D t + 1 , Freeze / Reject / Prove state .
4. Raw Telemetry 不應被單一 Cost Scalar 取代
可以定義:
C t ( M ) = w 1 T p r o m p t + w 2 T c o m p l e t i o n + w 3 L w a l l + w 4 N t o o l + w 5 N p r o o f − n o d e + ⋯ C_t(M)
=
w_1T_{\mathrm{prompt}}
+
w_2T_{\mathrm{completion}}
+
w_3L_{\mathrm{wall}}
+
w_4N_{\mathrm{tool}}
+
w_5N_{\mathrm{proof-node}}
+\cdots C t ( M ) = w 1 T prompt + w 2 T completion + w 3 L wall + w 4 N tool + w 5 N proof − node + ⋯
但本文強調:
Raw Telemetry ≠ Composite Cost . \boxed{
\text{Raw Telemetry}
\neq
\text{Composite Cost}.
} Raw Telemetry = Composite Cost .
原因包括:
API pricing 會變;
模型 token accounting 會變;
hardware latency 會變;
proof node 與 token 的相對價值依任務而變;
不同部署場景可以有不同 objective。
因此正確流程是:
Raw telemetry → calibrated view → C t . \boxed{
\text{Raw telemetry}
\rightarrow
\text{calibrated view}
\rightarrow
C_t.
} Raw telemetry → calibrated view → C t .
而不是:
Raw telemetry → discard → single scalar . \text{Raw telemetry}
\rightarrow
\text{discard}
\rightarrow
\text{single scalar}. Raw telemetry → discard → single scalar .
5. Specialist Capability Estimation
設 specialist:
M i . M_i. M i .
每次呼叫得到:
D t → D t + 1 . D_t
\rightarrow
D_{t+1}. D t → D t + 1 .
可觀測 reduction:
Δ D t = D t − D t + 1 . \Delta D_t
=
D_t-D_{t+1}. Δ D t = D t − D t + 1 .
因此可以估:
R ^ t ( M i ) = E [ Δ D ∣ M i , H t ] . \widehat R_t(M_i)
=
\mathbb E[
\Delta D
\mid
M_i,\mathcal H_t
]. R t ( M i ) = E [ Δ D ∣ M i , H t ] .
更一般:
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 可以包含:
domain;
theorem form;
source type;
corruption signature;
model version;
prompt version;
tool access;
runtime environment。
這使 capability model 由:
“MPF is good at quantifiers” \boxed{
\text{“MPF is good at quantifiers”}
} “MPF is good at quantifiers”
變成:
“MPF-v3 under prompt P7 reduces formal doubt by this distribution on this task class.” \boxed{
\text{“MPF-v3 under prompt P7 reduces formal doubt by this distribution on this task class.”}
} “MPF-v3 under prompt P7 reduces formal doubt by this distribution on this task class.”
6. Shrinkage Estimation 與 Cold Start
新 specialist 的 observation 可能很少。
若直接使用 sample mean:
C ^ = 1 n ∑ i = 1 n C i \widehat C
=
\frac1n
\sum_{i=1}^{n}C_i C = n 1 i = 1 ∑ n C i
在:
n ≪ 1 n\ll1 n ≪ 1
時極不穩定。
因此可使用 prior:
C 0 ( M ) , R 0 ( M ) , C_0(M),
\qquad
R_0(M), C 0 ( M ) , R 0 ( M ) ,
與 pseudo-count:
α 0 . \alpha_0. α 0 .
posterior-like shrinkage estimate:
C ^ t ( M ) = α 0 C 0 ( M ) + ∑ τ ≤ t : M τ = M C τ α 0 + n t ( M ) . \widehat C_t(M)
=
\frac{
\alpha_0C_0(M)+\sum_{\tau\le t:M_\tau=M}C_\tau
}{
\alpha_0+n_t(M)
}. C t ( M ) = α 0 + n t ( M ) α 0 C 0 ( M ) + ∑ τ ≤ t : M τ = M C τ .
同理:
R ^ t ( M ) = α 0 R 0 ( M ) + ∑ Δ D τ α 0 + n t ( M ) . \widehat R_t(M)
=
\frac{
\alpha_0R_0(M)+\sum \Delta D_\tau
}{
\alpha_0+n_t(M)
}. R t ( M ) = α 0 + n t ( M ) α 0 R 0 ( M ) + ∑ Δ D τ .
這不是要求 Bayesian interpretation,只是一個簡單、可工程化的 shrinkage estimator。
7. Selection Bias
然而,只利用歷史被選中的 specialist 會有 fundamental selection bias。
Orchestrator 看不到:
what would have happened if an unselected specialist had been called . \text{what would have happened if an unselected specialist had been called}. what would have happened if an unselected specialist had been called .
即 bandit feedback。
這與 online LLM routing 研究中的 partial-feedback 問題一致:部署時通常只知道被選中模型的 reward,而不是所有候選模型的 counterfactual reward。OrcaRouter 與其他 bandit routing 工作正是針對這類限制進行 online update。
對 cognitive orchestrator:
Only selected cognition produces telemetry . \boxed{
\text{Only selected cognition produces telemetry}.
} Only selected cognition produces telemetry .
若一個 specialist 因早期資料差而被減少調用:
few calls → few observations → high uncertainty → possibly fewer calls . \text{few calls}
\rightarrow
\text{few observations}
\rightarrow
\text{high uncertainty}
\rightarrow
\text{possibly fewer calls}. few calls → few observations → high uncertainty → possibly fewer calls .
可能形成 self-confirming lock-in。
8. Exploration–Exploitation Tension
因此不能只選:
arg max i VOI ^ t ( M i ) . \arg\max_i
\widehat{\operatorname{VOI}}_t(M_i). arg i max VOI t ( M i ) .
我們加入:
U t ( M i ) U_t(M_i) U t ( M i )
表示 epistemic uncertainty。
定義:
Score t ( M i ) = VOI ^ t ( M i ) + β U t ( M i ) . \boxed{
\operatorname{Score}_t(M_i)
=
\widehat{\operatorname{VOI}}_t(M_i)
+
\beta U_t(M_i).
} Score t ( M i ) = VOI t ( M i ) + β U t ( M i ) .
其中:
β ≥ 0. \beta\ge0. β ≥ 0.
這使:
exploitation:選目前估計最好;
exploration:給 under-observed specialist 機會。
因此:
Reasoning allocation itself has a tolerance tension . \boxed{
\text{Reasoning allocation itself has a tolerance tension}.
} Reasoning allocation itself has a tolerance tension .
9. Targeted Exploration 而非 Blind Exploration
如果 uncertainty 高就無條件探索:
U ( M ) ↑ ⇒ P ( M ) ↑ , U(M)\uparrow
\Rightarrow
P(M)\uparrow, U ( M ) ↑⇒ P ( M ) ↑ ,
會退化為浪費。
FlyRoute 提出的 targeted exploration 給出一個相似直覺:應優先探索 under-profiled agent,但只在 query 與 agent capability plausibly relevant 時進行,而非為了補資料隨機派發不相關任務。
因此本文建議:
ExploreScore ( M i ) = U i ⋅ Relevance ( M i , D t ) . \operatorname{ExploreScore}(M_i)
=
U_i
\cdot
\operatorname{Relevance}(M_i,D_t). ExploreScore ( M i ) = U i ⋅ Relevance ( M i , D t ) .
最後:
Score = VOI ^ + β U Relevance . \operatorname{Score}
=
\widehat{\operatorname{VOI}}
+
\beta
U
\operatorname{Relevance}. Score = VOI + β U Relevance .
如此 Bridge specialist 不會因為沒資料,就被拿去處理純代數量詞問題。
10. EXP-0008:Telemetry Contract
EXP-0008 建立一個 runtime event schema。
每個事件保存:
( run id , sample id , module , target state , D b e f o r e , D a f t e r , usage , latency , proof telemetry , quality ) . (
\text{run id},
\text{sample id},
\text{module},
\text{target state},
D_{\mathrm{before}},
D_{\mathrm{after}},
\text{usage},
\text{latency},
\text{proof telemetry},
\text{quality}
). ( run id , sample id , module , target state , D before , D after , usage , latency , proof telemetry , quality ) .
事件類型包括:
run start;
surface audit;
specialist selection;
specialist start/end;
target freeze/reject;
proof start/attempt/end;
error。
此實驗以 synthetic telemetry 驗證:
event emission → JSONL → aggregation → per-run metrics \text{event emission}
\rightarrow
\text{JSONL}
\rightarrow
\text{aggregation}
\rightarrow
\text{per-run metrics} event emission → JSONL → aggregation → per-run metrics
流程可以運作。
但 synthetic events 被明確標記,不能與 measured benchmark 混合。
11. 從 Telemetry 到 Self-Calibration
有了 history:
H t = { e 1 , … , e t } , \mathcal H_t
=
\{e_1,\ldots,e_t\}, H t = { e 1 , … , e t } ,
Orchestrator 更新:
C ^ t ( M ) \widehat C_t(M) C t ( M )
與:
R ^ t ( M ) . \widehat R_t(M). R t ( M ) .
因此:
Telemetry → Estimate → VOI → Routing → New Telemetry . \boxed{
\text{Telemetry}
\rightarrow
\text{Estimate}
\rightarrow
\text{VOI}
\rightarrow
\text{Routing}
\rightarrow
\text{New Telemetry}.
} Telemetry → Estimate → VOI → Routing → New Telemetry .
這形成 closed loop。
12. EXP-0009:Self-Calibration Plumbing Test
EXP-0009 使用 EXP-0008 synthetic telemetry 作為初始 history。
起始資料不平衡:
MPF 有少量 observations;
Source Fidelity 幾乎沒有;
Bridge 沒有。
因此:
U M P F < U S o u r c e ≈ U B r i d g e . U_{\mathrm{MPF}}
<
U_{\mathrm{Source}}
\approx
U_{\mathrm{Bridge}}. U MPF < U Source ≈ U Bridge .
再透過三輪 synthetic online environment 產生新 observations。
結果顯示:
MPF posterior cost 向 synthetic true cost 移動;
Source Fidelity posterior cost 同樣移動;
observation 數增加時 uncertainty 下降;
Bridge 因沒有被相關 task 觸發,uncertainty 保持高值。
這個最後現象尤其重要:
No observation ⇒ remain uncertain . \boxed{
\text{No observation}
\Rightarrow
\text{remain uncertain}.
} No observation ⇒ remain uncertain .
而不是:
No observation ⇒ assume learned . \text{No observation}
\Rightarrow
\text{assume learned}. No observation ⇒ assume learned .
13. Meta-Orchestration 與一般 Model Training 的區別
本文的 self-calibration 並不要求重新訓練 foundation model。
底層:
M 1 , … , M k M_1,\ldots,M_k M 1 , … , M k
可以全部固定。
改變的是上層:
π t : ( D , X , H ) → a . \pi_t
:
(D,X,\mathcal H)
\rightarrow
a. π t : ( D , X , H ) → a .
即 routing / orchestration policy。
因此學習對象是:
how to allocate cognition , \boxed{
\text{how to allocate cognition},
} how to allocate cognition ,
不是:
how to perform cognition internally . \boxed{
\text{how to perform cognition internally}.
} how to perform cognition internally .
這使 meta-orchestration 可以比 foundation-model retraining 更輕量、更頻繁地更新。
14. Non-Stationarity
靜態 estimator 假設:
P ( C , R ∣ M ) P(C,R\mid M) P ( C , R ∣ M )
固定。
實際部署更合理:
P t ( C , R ∣ M ) \boxed{
P_t(C,R\mid M)
} P t ( C , R ∣ M )
隨時間變動。
原因包括:
Model drift
provider 更新 model;
silent regression;
quantization change;
context-window change。
Prompt drift
P t ≠ P t + 1 . P_t
\neq
P_{t+1}. P t = P t + 1 .
Tool drift
retriever、search API、Lean library、CAS 版本改變。
Cost drift
API price、rate limit、latency、hardware utilization 改變。
Task-distribution drift
輸入 domain distribution 改變。
因此:
Historical competence is not permanent competence . \boxed{
\text{Historical competence is not permanent competence}.
} Historical competence is not permanent competence .
15. 與 Non-Stationary LLM Routing 的關係
ParetoBandit 特別把 production LLM routing 描述為 non-stationary quality–cost environment:
provider pricing 改變;
model quality 可能 silent regression;
new models runtime onboarding。
其方法使用 forgetting、online budget control 與 new-model exploration,使 router 在條件改變時重新學習 quality–cost niche。
這與本文有重要共同點:
old routing evidence must sometimes decay . \boxed{
\text{old routing evidence must sometimes decay}.
} old routing evidence must sometimes decay .
但本文 tracking 的不是只有:
( quality , price ) , (\text{quality},\text{price}), ( quality , price ) ,
而是:
( doubt-reduction capability , runtime cost , target-state effect ) . \boxed{
(
\text{doubt-reduction capability},
\text{runtime cost},
\text{target-state effect}
).
} ( doubt-reduction capability , runtime cost , target-state effect ) .
16. Staleness State
因此每個 specialist 的 meta-state 不應只有:
( C ^ , R ^ , U ) . (\widehat C,\widehat R,U). ( C , R , U ) .
我們加入:
A t ( M ) A_t(M) A t ( M )
表示 age / staleness。
最小 specialist state:
S M ( t ) = ( C ^ t , R ^ t , U t , A t , V t ) . \boxed{
\mathcal S_M(t)
=
(
\widehat C_t,
\widehat R_t,
U_t,
A_t,
V_t
).
} S M ( t ) = ( C t , R t , U t , A t , V t ) .
其中:
C ^ t \widehat C_t C t :估計成本;
R ^ t \widehat R_t R t :估計能力;
U t U_t U t :uncertainty;
A t A_t A t :staleness;
V t V_t V t :version metadata。
17. Staleness Score
可以定義:
S t ( M ) = f ( Δ t l a s t , E t , Δ V t , Δ P t , Δ H t ) . S_t(M)
=
f(
\Delta t_{\mathrm{last}},
E_t,
\Delta V_t,
\Delta P_t,
\Delta H_t
). S t ( M ) = f ( Δ t last , E t , Δ V t , Δ P t , Δ H t ) .
其中:
Δ t l a s t \Delta t_{\mathrm{last}} Δ t last :距上次 observation 的時間;
E t E_t E t :prediction error;
Δ V t \Delta V_t Δ V t :model version change;
Δ P t \Delta P_t Δ P t :prompt version change;
Δ H t \Delta H_t Δ H t :runtime / hardware change。
當:
S t ( M ) ↑ , S_t(M)\uparrow, S t ( M ) ↑ ,
可以:
w h i s t o r i c a l ↓ , w_{\mathrm{historical}}\downarrow, w historical ↓ ,
或:
U t ( M ) ↑ , U_t(M)\uparrow, U t ( M ) ↑ ,
或:
P ( re-explore M ) ↑ . P(\text{re-explore }M)\uparrow. P ( re-explore M ) ↑ .
18. Forgetting Factor
最簡單的 non-stationary estimator 為 exponential forgetting。
對 observation:
x t , x_t, x t ,
更新:
x ^ t = ( 1 − α ) x ^ t − 1 + α x t . \widehat x_t
=
(1-\alpha)\widehat x_{t-1}
+
\alpha x_t. x t = ( 1 − α ) x t − 1 + α x t .
其中:
0 < α ≤ 1. 0<\alpha\le1. 0 < α ≤ 1.
大 α \alpha α :
adaptation 快;
variance 大。
小 α \alpha α :
所以 forgetting 本身也有 bias–variance tension。
19. Change-Point Detection
若 specialist behavior abrupt change:
θ t − ≠ θ t + , \theta_{t^-}
\neq
\theta_{t^+}, θ t − = θ t + ,
持續 EWMA 可能適應太慢。
定義 innovation:
I t = g ( C t − C ^ t − , R t − R ^ t − ) . I_t
=
g(
C_t-\widehat C_{t^-},
R_t-\widehat R_{t^-}
). I t = g ( C t − C t − , R t − R t − ) .
若:
I t I_t I t
顯著超出近期 residual distribution,
觸發:
ShiftSuspected ( M ) = 1. \boxed{
\operatorname{ShiftSuspected}(M)=1.
} ShiftSuspected ( M ) = 1.
然後執行:
local reset or aggressive forgetting . \text{local reset}
\quad\text{or}\quad
\text{aggressive forgetting}. local reset or aggressive forgetting .
20. Partial Observability of Drift
一個重要限制:
如果 specialist 沒被呼叫,
就看不到它變了。
因此 shift detection latency 至少分為:
L s h i f t = L o p p o r t u n i t y + L s t a t i s t i c a l . \boxed{
L_{\mathrm{shift}}
=
L_{\mathrm{opportunity}}
+
L_{\mathrm{statistical}}.
} L shift = L opportunity + L statistical .
其中:
L o p p o r t u n i t y L_{\mathrm{opportunity}} L opportunity :到下一次 specialist 被觀測的時間;
L s t a t i s t i c a l L_{\mathrm{statistical}} L statistical :取得足夠證據的時間。
這再次說明:
exploration is required not only for learning capability, but also for detecting capability drift . \boxed{
\text{exploration is required not only for learning capability, but also for detecting capability drift}.
} exploration is required not only for learning capability, but also for detecting capability drift .
21. EXP-0010:Non-Stationary Simulation
EXP-0010 建立 180 個 synthetic tasks,並在:
t = 61 t=61 t = 61
及:
t = 121 t=121 t = 121
改變 hidden specialist environment。
Phase 2:
MPF cost 上升;
MPF formal reduction 下降;
Source Fidelity cost 下降;
Source capability 提升。
Phase 3:
Bridge cost 大幅下降;
Bridge capability 提升;
MPF 部分恢復。
比較:
stationary mean;
EWMA;
change-point + reset。
22. Choice Accuracy 不足以衡量 Adaptation
初步 scheduler-level choice accuracy 可能很快恢復。
但這可能只是因 task mix 剛好讓 stale estimate 沒造成太多錯選。
因此 EXP-0010 增加 parameter-level recovery:
要求:
∣ C ^ − C ∗ ∣ C ∗ ≤ 0.2 \frac{
|\widehat C-C^\ast|
}{
C^\ast
}
\le0.2 C ∗ ∣ C − C ∗ ∣ ≤ 0.2
且 specialist 主能力:
∣ R ^ − R ∗ ∣ ≤ 0.1. |\widehat R-R^\ast|
\le0.1. ∣ R − R ∗ ∣ ≤ 0.1.
才算重新學到新環境。
此結果顯示,stationary estimator 對某些大幅 shift 在整個 horizon 內都沒有真正恢復,而 change-point policy 在 abrupt shift 下可以大幅縮短 parameter recovery delay。
23. Abrupt Shift 與 Gradual Drift 不同
但 change-point policy 在 partial recovery / smaller drift 上可能漏報。
因此:
Abrupt Change Detection ≠ General Drift Tracking . \boxed{
\text{Abrupt Change Detection}
\neq
\text{General Drift Tracking}.
} Abrupt Change Detection = General Drift Tracking .
未來需要至少區分:
Abrupt Shift , Gradual Drift , Recurring Drift , Context-Conditional Shift . \begin{aligned}
&\text{Abrupt Shift},\\
&\text{Gradual Drift},\\
&\text{Recurring Drift},\\
&\text{Context-Conditional Shift}.
\end{aligned} Abrupt Shift , Gradual Drift , Recurring Drift , Context-Conditional Shift .
這意味著單一 z-score detector 不足以構成完整 drift-aware orchestrator。
24. False Alarm 也是成本
change-point detector 越敏感:
detection delay ↓ , \text{detection delay}\downarrow, detection delay ↓ ,
但:
false alarm rate ↑ . \text{false alarm rate}\uparrow. false alarm rate ↑ .
每次 false alarm 可能導致:
estimator reset;
forced re-exploration;
temporary routing instability;
unnecessary specialist calls。
所以真正目標是:
min ( L r e c o v e r y , R r e g r e t , N f a l s e − a l a r m , C r e e x p l o r e ) . \boxed{
\min
\left(
L_{\mathrm{recovery}},
R_{\mathrm{regret}},
N_{\mathrm{false-alarm}},
C_{\mathrm{reexplore}}
\right).
} min ( L recovery , R regret , N false − alarm , C reexplore ) .
這是一個 multi-objective control problem。
25. 三個 Feedback Loops
整體架構現在具有三個回饋迴路。
25.1 Epistemic Loop
D t → M t → D t + 1 . D_t
\rightarrow
M_t
\rightarrow
D_{t+1}. D t → M t → D t + 1 .
回答:
哪個 specialist 能消除當前問題疑點?
25.2 Operational Loop
Telemetry t → ( C ^ t , R ^ t ) → Routing → Telemetry t + 1 . \text{Telemetry}_t
\rightarrow
(\widehat C_t,\widehat R_t)
\rightarrow
\text{Routing}
\rightarrow
\text{Telemetry}_{t+1}. Telemetry t → ( C t , R t ) → Routing → Telemetry t + 1 .
回答:
我們目前對 specialist 的能力與成本估計是什麼?
25.3 Drift Loop
Prediction Error → Shift Suspicion → Forget / Reset / Re-explore → Recalibrate . \text{Prediction Error}
\rightarrow
\text{Shift Suspicion}
\rightarrow
\text{Forget / Reset / Re-explore}
\rightarrow
\text{Recalibrate}. Prediction Error → Shift Suspicion → Forget / Reset / Re-explore → Recalibrate .
回答:
我們以前學到的 specialist model 是否仍然有效?
26. 從 Orchestrator 到 Epistemic Governor
傳統 router:
x → M . x\rightarrow M. x → M .
Adaptive router:
( x , H ) → M . (x,\mathcal H)\rightarrow M. ( x , H ) → M .
本文的 cognitive orchestrator:
( D , X , H , V , A ) → { M i , Freeze } . (D,X,\mathcal H,V,A)
\rightarrow
\{
M_i,\operatorname{Freeze}
\}. ( D , X , H , V , A ) → { M i , Freeze } .
因此分層可以寫為:
L 0 : Worker / Solver L 1 : Specialist Reasoner L 2 : Adaptive Orchestrator L 3 : Epistemic Governor \begin{array}{ll}
L_0 &: \text{Worker / Solver}\\
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 : Worker / Solver : Specialist Reasoner : Adaptive Orchestrator : Epistemic Governor
L 3 L_3 L 3 不只決定誰工作。
它還管理:
哪些 specialist belief 已 stale;
哪些能力模型需要重新估;
哪些 uncertainty 需要探索;
哪些 history 應該忘記;
哪些 target 可以 Freeze;
哪些 operational shifts 必須觸發 re-routing。
27. Self-Evolving Agent Profiles 與本文的關係
FlyRoute 使用成功案例與 traffic data 持續更新 agent profiles,顯示 static capability description 在 agent 持續變動的環境中會失效。
本文採取更數值化的角色模型:
Profile ( M ) = ( C ^ , R ^ , U , A , V ) . \operatorname{Profile}(M)
=
(
\widehat C,
\widehat R,
U,
A,
V
). Profile ( M ) = ( C , R , U , A , V ) .
兩者可以互補:
textual capability profile:
用於 semantic matching;
quantitative telemetry profile:
用於 VOI / cost / uncertainty routing。
因此未來可以:
Semantic Profile + Empirical Performance Profile . \boxed{
\text{Semantic Profile}
+
\text{Empirical Performance Profile}.
} Semantic Profile + Empirical Performance Profile .
28. Contextual Bandit Interpretation
若每一輪只能看到被選 specialist 的結果,
meta-orchestration 可被視為 contextual bandit:
Context:
X t = ( D t , problem features , runtime state ) . X_t
=
(D_t,\text{problem features},\text{runtime state}). X t = ( D t , problem features , runtime state ) .
Arm:
M i . M_i. M i .
Reward:
r t = Δ V t − λ C t . r_t
=
\Delta V_t
-
\lambda C_t. r t = Δ V t − λ C t .
或:
r t = Δ V t C t . r_t
=
\frac{\Delta V_t}{C_t}. r t = C t Δ V t .
但本文比一般 bandit 多一個特殊 action:
Freeze . \boxed{
\operatorname{Freeze}.
} Freeze .
而且 freeze 受到 hard target-safety constraints。
所以更精確是:
constrained contextual bandit / sequential control with stopping . \boxed{
\text{constrained contextual bandit / sequential control with stopping}.
} constrained contextual bandit / sequential control with stopping .
29. Routing Regret
若 hindsight oracle:
M t ∗ = arg max i VOI t t r u e ( M i ) , M_t^\ast
=
\arg\max_i
\operatorname{VOI}^{\mathrm{true}}_t(M_i), M t ∗ = arg i max VOI t true ( M i ) ,
則:
R T = ∑ t = 1 T [ VOI t t r u e ( M t ∗ ) − VOI t t r u e ( M t ) ] . R_T
=
\sum_{t=1}^{T}
[
\operatorname{VOI}^{\mathrm{true}}_t(M_t^\ast)
-
\operatorname{VOI}^{\mathrm{true}}_t(M_t)
]. R T = t = 1 ∑ T [ VOI t true ( M t ∗ ) − VOI t true ( M t )] .
在真實世界:
VOI t r u e \operatorname{VOI}^{\mathrm{true}} VOI true
通常不可直接觀測。
因此 regret 只能透過:
controlled benchmark;
full-information offline evaluation;
counterfactual estimator;
synthetic environment;
近似。
所以本文中的 EXP-0010 regret 明確只是一個 synthetic diagnostic。
30. Specialist Versioning
當:
M ( v 1 ) → M ( v 2 ) , M^{(v1)}
\rightarrow
M^{(v2)}, M ( v 1 ) → M ( v 2 ) ,
歷史 telemetry 是否可直接繼承?
不一定。
如果只看 role:
Role ( M v 1 ) = Role ( M v 2 ) = M P F , \operatorname{Role}(M^{v1})
=
\operatorname{Role}(M^{v2})
=
\mathrm{MPF}, Role ( M v 1 ) = Role ( M v 2 ) = MPF ,
但 capability distribution:
P ( R , C ∣ M v 1 ) ≠ P ( R , C ∣ M v 2 ) , P(R,C\mid M^{v1})
\neq
P(R,C\mid M^{v2}), P ( R , C ∣ M v 1 ) = P ( R , C ∣ M v 2 ) ,
則完全繼承可能錯。
因此:
Specialist Role ≠ Specialist Instance . \boxed{
\text{Specialist Role}
\neq
\text{Specialist Instance}.
} Specialist Role = Specialist Instance .
這是下一階段 model replacement / identity continuity 的核心問題。
31. Capability Transfer
可以定義兩 specialist instances 的 transfer coefficient:
ρ ( M v 1 , M v 2 ) ∈ [ 0 , 1 ] . \rho
(
M^{v1},
M^{v2}
)
\in[0,1]. ρ ( M v 1 , M v 2 ) ∈ [ 0 , 1 ] .
若:
ρ = 1 , \rho=1, ρ = 1 ,
完全繼承。
若:
ρ = 0 , \rho=0, ρ = 0 ,
cold start。
一般:
0 < ρ < 1. 0<\rho<1. 0 < ρ < 1.
new prior:
θ 0 v 2 = ρ θ ^ v 1 + ( 1 − ρ ) θ r o l e − p r i o r . \theta_0^{v2}
=
\rho\widehat\theta^{v1}
+
(1-\rho)\theta_{\mathrm{role-prior}}. θ 0 v 2 = ρ θ v 1 + ( 1 − ρ ) θ role − prior .
這讓 model hot-swap 不必永遠從零開始,也不必盲目繼承全部 history。
32. Runtime Identity 與 Functional Identity
因此需要區分:
Runtime identity
具體:
model ID;
model version;
prompt;
tool set;
hardware;
endpoint。
Functional identity
例如:
M P F S p e c i a l i s t . \mathrm{MPF\ Specialist}. MPF Specialist .
同一 functional role:
R R R
可以由多個 runtime substrate 實現。
因此:
Functional Role ≠ Runtime Substrate . \boxed{
\text{Functional Role}
\neq
\text{Runtime Substrate}.
} Functional Role = Runtime Substrate .
這也是 scalable agent runtime 的重要抽象。
33. 評測指標
真正部署時至少需要:
Calibration
∣ C ^ − C ∣ , ∣ R ^ − R ∣ . |\widehat C-C|,
\qquad
|\widehat R-R|. ∣ C − C ∣ , ∣ R − R ∣.
Routing
selection accuracy , regret . \text{selection accuracy},
\quad
\text{regret}. selection accuracy , regret .
Exploration
N e x p l o r e , C e x p l o r e . N_{\mathrm{explore}},
\quad
C_{\mathrm{explore}}. N explore , C explore .
Drift
L d e t e c t , L r e c o v e r , false alarm rate . L_{\mathrm{detect}},
\quad
L_{\mathrm{recover}},
\quad
\text{false alarm rate}. L detect , L recover , false alarm rate .
Safety
false freeze rate , corruption miss rate . \text{false freeze rate},
\quad
\text{corruption miss rate}. false freeze rate , corruption miss rate .
Downstream impact
proof success , proof nodes , tokens , latency . \text{proof success},
\quad
\text{proof nodes},
\quad
\text{tokens},
\quad
\text{latency}. proof success , proof nodes , tokens , latency .
34. EXP-0008~0010 的研究地位
必須明確區分三個層級。
Level 1:Schema / Plumbing
已完成:
telemetry event schema;
recorder;
aggregator;
cost interface;
self-calibration learner;
shift monitor。
Level 2:Synthetic Simulation
已完成:
posterior movement;
exploration–exploitation;
non-stationary shift;
change-point reset;
parameter recovery。
Level 3:Measured Empirical Validation
尚未完成。
需要:
fresh independent agents;
real token usage;
real latency;
real tool logs;
real Lean/prover logs;
fixed model versions;
blinded evaluation。
因此:
engineering feasibility ≠ empirical effectiveness . \boxed{
\text{engineering feasibility}
\neq
\text{empirical effectiveness}.
} engineering feasibility = empirical effectiveness .
35. 可檢驗假說
H1:Self-Calibration
真實 telemetry 可降低:
∣ C ^ − C ∣ |\widehat C-C| ∣ C − C ∣
與:
∣ R ^ − R ∣ |\widehat R-R| ∣ R − R ∣
相較固定人工 proxy。
H2:Exploration
適度 exploration 應降低 cold-start / selection-lock-in regret。
H3:Forgetting
在 non-stationary environment:
R T f o r g e t t i n g < R T s t a t i o n a r y R_T^{\mathrm{forgetting}}
<
R_T^{\mathrm{stationary}} R T forgetting < R T stationary
至少在某些 shift regimes 成立。
H4:Change-Point Reset
對 abrupt shift:
L r e c o v e r r e s e t < L r e c o v e r E W M A . L_{\mathrm{recover}}^{\mathrm{reset}}
<
L_{\mathrm{recover}}^{\mathrm{EWMA}}. L recover reset < L recover EWMA .
H5:Drift Cost
過度敏感的 change detector 會提高:
N f a l s e − a l a r m + C r e e x p l o r e . N_{\mathrm{false-alarm}}
+
C_{\mathrm{reexplore}}. N false − alarm + C reexplore .
因此不存在單純「越敏感越好」。
36. 失敗模式
36.1 Feedback Poisoning
錯誤 quality label 會污染 capability model。
36.2 Selection Bias
只有被選 specialist 有 observation。
36.3 Survivorship Bias
失敗 Agent 若被移除,其歷史可能不再進入比較。
36.4 Cross-Task Generalization Error
某 specialist 在一個 domain 很好,不代表另一 domain 相同。
36.5 Version Leakage
不同 model version telemetry 被錯誤混合。
36.6 Cost Miscalibration
token cost、latency 與 proof-node cost 被錯誤壓成不合理 scalar。
36.7 Change Detector Overreaction
noise 被誤判成 shift。
36.8 Under-Exploration
新 specialist 永遠沒有足夠資料。
36.9 Over-Exploration
系統退化為 Always-Dual。
36.10 Governance Failure
上層 orchestrator 本身的 objective 若錯,底層 specialists 再強也無法保證整體正確。
37. 三層「知道」
一個 solver 需要知道:
how to solve . \boxed{
\text{how to solve}.
} how to solve .
一個 orchestrator 需要知道:
who should solve . \boxed{
\text{who should solve}.
} who should solve .
一個 meta-orchestrator 需要知道:
how confident it should be in its belief about who should solve . \boxed{
\text{how confident it should be in its belief about who should solve}.
} how confident it should be in its belief about who should solve .
因此:
knowledge → knowledge allocation → knowledge about allocation quality . \boxed{
\text{knowledge}
\rightarrow
\text{knowledge allocation}
\rightarrow
\text{knowledge about allocation quality}.
} knowledge → knowledge allocation → knowledge about allocation quality .
這是本文認為 meta-orchestration 的核心認知層次。
38. 從 AI Router 到 AI Organization
當 specialist 數量增加:
∣ M ∣ ≫ 1 , |\mathcal M|
\gg1, ∣ M ∣ ≫ 1 ,
整個系統越來越像一個 organization。
每個 specialist:
有 role;
有成本;
有歷史績效;
有 uncertainty;
會更新;
可能退化;
可能被替換。
Orchestrator:
分配工作;
監測結果;
更新 profile;
調整信任;
重新探索;
停止無效工作。
因此可以把它視為:
Cognitive Organization Controller . \boxed{
\text{Cognitive Organization Controller}.
} Cognitive Organization Controller .
但本文不賦予「organization」任何人格或法律主體含義;這只是一個 functional systems analogy。
39. Problem-Space Governance 的完整閉環
整個系列到本文可以寫為:
C s o u r c e → C c a n d i d a t e → C f r e e z e ∗ → D t → M t → D t + 1 \boxed{
C_{\mathrm{source}}
\rightarrow
\mathcal C_{\mathrm{candidate}}
\rightarrow
C^\ast_{\mathrm{freeze}}
\rightarrow
D_t
\rightarrow
M_t
\rightarrow
D_{t+1}
} C source → C candidate → C freeze ∗ → D t → M t → D t + 1
同時:
Telemetry → C ^ t , R ^ t , U t , A t → routing policy update . \boxed{
\text{Telemetry}
\rightarrow
\widehat C_t,
\widehat R_t,
U_t,
A_t
\rightarrow
\text{routing policy update}.
} Telemetry → C t , R t , U t , A t → routing policy update .
如果 drift:
Prediction Error → Shift Detection → Forget / Reset / Re-explore . \boxed{
\text{Prediction Error}
\rightarrow
\text{Shift Detection}
\rightarrow
\text{Forget / Reset / Re-explore}.
} Prediction Error → Shift Detection → Forget / Reset / Re-explore .
最後 target freeze 後:
C f r e e z e ∗ → M C D M → P r o o f R o u t e r → F o r m a l P r o o f . C^\ast_{\mathrm{freeze}}
\rightarrow
\mathrm{MCDM}
\rightarrow
\mathrm{ProofRouter}
\rightarrow
\mathrm{FormalProof}. C freeze ∗ → MCDM → ProofRouter → FormalProof .
40. 討論
本文的核心不是讓 AI「自我修改」到不可控。
相反地,meta-orchestration 的更新對象可以非常有限:
routing beliefs and allocation policy . \boxed{
\text{routing beliefs and allocation policy}.
} routing beliefs and allocation policy .
底層:
model weights;
proof kernel;
tool permissions;
可以完全固定。
這提供一個重要工程優勢:
系統可以適應環境,而不必讓所有能力層都同步可變。
因此 adaptive orchestration 可以成為:
a controlled adaptation layer . \boxed{
\text{a controlled adaptation layer}.
} a controlled adaptation layer .
41. 結論
本文提出 Self-Calibrating and Drift-Aware Cognitive Orchestration。
其最小 specialist state 為:
S M ( t ) = ( C ^ t , R ^ t , U t , A t , V t ) . \boxed{
\mathcal S_M(t)
=
(
\widehat C_t,
\widehat R_t,
U_t,
A_t,
V_t
).
} S M ( t ) = ( C t , R t , U t , A t , V t ) .
其中:
C ^ t \widehat C_t C t :cost estimate;
R ^ t \widehat R_t R t :capability / doubt-reduction estimate;
U t U_t U t :uncertainty;
A t A_t A t :staleness;
V t V_t V t :runtime identity/version。
Routing 使用:
Score t ( M ) = VOI ^ t ( M ) + β U t ( M ) \boxed{
\operatorname{Score}_t(M)
=
\widehat{\operatorname{VOI}}_t(M)
+
\beta U_t(M)
} Score t ( M ) = VOI t ( M ) + β U t ( M )
並在 non-stationary setting 下加入:
forgetting , change-point detection , re-exploration . \text{forgetting},
\quad
\text{change-point detection},
\quad
\text{re-exploration}. forgetting , change-point detection , re-exploration .
因此系統形成三個 feedback loops:
Epistemic Loop \boxed{
\text{Epistemic Loop}
} Epistemic Loop
Operational Learning Loop \boxed{
\text{Operational Learning Loop}
} Operational Learning Loop
Drift Adaptation Loop . \boxed{
\text{Drift Adaptation Loop}.
} Drift Adaptation Loop .
數學 AI 的上層控制問題因此由:
Who should think next? \text{Who should think next?} Who should think next?
進一步變成:
How should the system update its belief about who should think next? \boxed{
\text{How should the system update its belief about who should think next?}
} How should the system update its belief about who should think next?
本文中的 EXP-0008~0010 只證明該閉環可以被資料化、實作與 synthetic stress-test;真正的性能主張仍必須由 measured runtime telemetry、獨立 Agent 與 formal prover benchmark 驗證。
但就架構而言,數學 AI 已不再只能被描述為 solver 的集合。它可以被更完整地描述為:
a self-monitoring cognitive organization with controlled epistemic resource allocation . \boxed{
\text{a self-monitoring cognitive organization with controlled epistemic resource allocation}.
} a self-monitoring cognitive organization with controlled epistemic resource allocation .
參考文獻
Li, R., Zhou, Z., Wu, Y. FlyRoute: Self-Evolving Agent Profiling via Data Flywheel for Adaptive Task Routing. arXiv:2605.22057, 2026.
Bao, Z., Tian, F., Zhang, C., Chen, Z., Ma, X., Shi, Y. OrcaRouter: A Production-Oriented LLM Router with Hybrid Offline-Online Learning. arXiv:2605.30736, 2026.
Taberner-Miller, A. ParetoBandit: Budget-Paced Adaptive Routing for Non-Stationary LLM Serving. arXiv:2604.00136, 2026.
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.
Wei, W., Yang, T., Chen, H., Zhao, Y., Dernoncourt, F., Rossi, R. A., Eldardiry, H. Learning to Route LLMs from Bandit Feedback: One Policy, Many Trade-offs. arXiv:2510.07429, 2025.
Li, Y. LLM Bandit: Cost-Efficient LLM Generation via Preference-Conditioned Dynamic Routing. arXiv:2502.02743, 2025.
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.
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 Self-Calibrating Cognitive Orchestration、specialist state S M ( t ) \mathcal S_M(t) S M ( t ) 、doubt-reduction capability model、staleness state、drift loop 及其與 theorem-target governance 的整合,屬於本文提出的研究框架。
EXP-0008~EXP-0010 使用的 telemetry、cost、capability、task distribution、change point 與 regret 均包含 synthetic / proxy components。本文不將任何 synthetic numerical result 解讀為真實 LLM routing accuracy、token savings、latency improvement 或 formal theorem proving speedup。
真正 empirical validation 需要固定 model / prompt / runtime version,使用獨立 Agent,保存 raw measured telemetry,並在 blinded benchmark 與 formal prover logs 上重現。