六代（6G）網路的系統級資源管理，必須對本質上具空間性的網路狀態 進行推理：基地台部署、使用者分佈與同頻干擾，全都座落於二維地理 平面之上。傳統數值最佳化器、深度強化學習（DRL）策略，以及近期 僅以文字驅動的大型語言模型（LLM）編排器，同樣都在做出任何決策 之前，先將此狀態化約為一維向量或序列化的提示，從而丟棄了主宰細 胞間耦合與使用者移動的幾何結構。 本文提出一套 Vision-to-Action 多智能體框架，改以光線追蹤數位孿生將 網路狀態渲染為具空間結構的視覺模態，並讓一組多模態大型語言模型 （MLLM）智能體（Observer、Allocator 與 Evaluator）直接對其進行推 理。為使所產生的決策得以執行，生成過程在解碼階段受結構文法約 束，使每一個輸出的動作皆由建構本身保證語法有效且落於範圍之內， 而非仰賴事後檢查。 在光線追蹤都市數位孿生上、跨十二個流量情境的評估中，本框架達 到 83.6% 的平均使用者滿意率，超越 DRL 基線 12.7 個百分點、超越 比例式切片啟發式 7.6 個百分點，且此一排序在三個能力差異顯著的 backbone 上皆維持不變。在校準期間從未見過的流量型態下，DRL 基線 退化 17.3 個百分點，而未經任何情境校準的本框架，則取得所有比較方 法中最低的跨情境變異。 此優勢源自架構而非 backbone 本身：同一個模型，在接收相同觀測、 卻被要求於單一 context 中同時進行感知、配置與驗證時，其表現低於 它原本要取代的啟發式方法。受控消融實驗將增益追溯至主動式需求預 測，而視覺模態正是使其成為可能的關鍵：在數值歷史完全相同的條件 下，加入渲染快照可將需求預測準確度自 81.2% 提升至 92.0%。結構化 解碼則進一步消除了未受約束生成中 31.2% 無法被解析為可執行動作的 比例。;System-level resource management in sixth-generation (6G) networks requires reasoning over network states that are inherently spatial: base station deploy- ments， user distributions， and co-channel interference all reside in a two-dimensional geographic plane. Conventional numerical optimizers and deep reinforcement learning (DRL) policies， as well as recent large language model (LLM) based orchestrators， all flatten this state into a one-dimensional vector or a serialized textual prompt before any decision is made， discarding the geometric struc- ture that governs inter-cell coupling and user mobility. This paper presents a Vision-to-Action multi-agent framework that instead renders the network state into a spatially structured visual modality through a ray-tracing-based digital twin， and lets a team of foundation-model agents—an Observer， an Alloca- tor， and an Evaluator—reason over it directly. To make the resulting deci- sions executable， generation is constrained at the decoding stage by a structural grammar， so that every emitted action is syntactically valid and in-range by construction rather than by inspection.Evaluated over twelve traffic scenarios on a ray-traced urban digital twin， the framework attains an average user sat- isfaction rate of 83.6%， exceeding a DRL baseline by 12.7 percentage points and a proportional slicing heuristic by 7.6， with the ordering preserved across three backbones of markedly different capacity. Under a traffic regime never seen during calibration， the DRL baseline degrades by 17.3 percentage points， whereas the proposed framework， calibrated on no scenario， attains the lowest cross-scenario variance of all methods compared. A controlled ablation traces the gain to proactive demand prediction: given an identical numeric history， adding the rendered snapshot raises prediction accuracy from 81.2% to 92.0%.
