大型語言模型在複雜推理上表現卓越，但主流方法有兩項核心缺陷：思維鏈（Chain of Thought）的線性推理因自迴歸不可逆，早期的幻覺前提會沿推理鏈向下傳播；傳統結果獎勵模型（Outcome Reward Model）僅對最終答案給予稀疏回饋，無法對中間步驟給予細粒度評分，導致邏輯錯誤難以精確定位與修正。 為此，本研究提出具過程驗證之多代理人編排框架（MOPV）。在結構層面，系統以共享白板協調三種異質代理人——思維圖譜（Graph of Thought）負責非線性多路徑探索、思維分層（Layer of Thought）負責層級化分解、思維逆轉（Reversal of Thought）負責逆向驗證推理前提之正確性——迫使代理人以顯性辯證解決矛盾，而非被動投票聚合；各代理人的推理軌跡經白板整合為結構化推理流（ThoughtFlow）。在驗證層面，實作基於 LLM 的過程獎勵模型（PRM），對每個中間步驟即時評分，若信心低於閾值（τ = 0.9）即觸發遞迴修正，最多三次。 以 20B 參數的 gpt-oss-20b 為骨幹，本研究於兩項基準進行驗證。在評估知識廣度的 MMLU 上，MOPV 準確率達 90.64%，超越 GPT-4（86.4%）與基線模型（86.71%），並於 57 個子類別中 28 類超越當前排行榜最佳成績；在評估多步軟性推理的 MuSR 上，MOPV 平均準確率達 66.5%，與 GPT-4（69.9%）相當，並於 Team Allocation 子任務以 72.0% 超越 GPT-4（68.4%）。消融實驗進一步證實移除任一核心元件均導致顯著下降。此外，本研究以 MOPV 產生之推理流（ThoughtFlow）作為訓練資料，透過 CasCoD 級聯目標增強小型學生模型 Llama-3.1-8B-Instruct；其表現與以相同骨幹進行之單次 CoT 監督相當，顯示 MOPV 之 ThoughtFlow 所傳遞的是推理過程層級的監督訊號，而非僅止於答案標籤，可作為一種獨特且可用於下游的推理監督來源。;Large Language Models (LLMs) excel at complex reasoning， yet prevailing methods suffer from two core limitations: Chain-of-Thought (CoT) linear inference is irreversible under autoregressive generation， allowing hallucinated premises to propagate unchecked; and Outcome Reward Models (ORMs) provide only sparse terminal feedback， lacking the finegrained intermediate scoring needed to localize logical errors. We propose MOPV (Multi-agent Orchestration with Process Verification). A Shared Whiteboard coordinates three heterogeneous agents: Graph of Thoughts (GoT) for non-linear exploration， Layer of Thoughts (LoT) for hierarchical decomposition， and Reversal of Thought (RoT) for backward premise verification. These agents resolve conflicts through explicit dialectic rather than voting， integrating their trajectories into structured ThoughtFlows. An LLM-based Process Reward Model (PRM) scores each intermediate step， triggering recursive refinement， capped at three iterations， when confidence falls below a threshold (τ = 0.9). On a 20B-parameter backbone， MOPV attains 90.64% on MMLU (surpassing GPT-4 at 86.4% and the base model at 86.71%， with new best scores in 28 of 57 subcategories) and 66.5% on MuSR， competitive with GPT-4 (69.9%) and exceeding it on Team Allocation (72.0% vs. 68.4%). Ablations confirm every core component contributes significantly. Finally， augmenting a Llama-3.1-8B-Instruct student with MOPV ThoughtFlows via the CasCoD objective matches single-pass CoT supervision from the same backbone， showing that ThoughtFlows transfer process-level reasoning supervision—not just answer labels—as a distinct， downstream-usable signal.
