Target allocation in multi-object spectroscopic surveys is a high-dimensional combinatorial optimization problem constrained by fiber reachability and mechanical collisions. In densely populated focal planes, local assignment decisions can induce non-local conflicts, whereby a single choice suppresses multiple neighboring opportunities, leading to reduced survey completeness and biased target selection. Existing approaches typically rely on fast heuristic or priority-driven algorithms, which, while operationally efficient, do not fully capture the global conflict structure and therefore yield suboptimal weighted target assignments in dense regimes. In this work, we reformulate the fiber assignment problem as a Maximum Weight Independent Set problem on a conflict graph, providing a unified representation of geometric and mechanical constraints. Building on this formulation, we develop a Graph Neural Network–based solver that leverages iterative message passing to model long-range dependencies across the focal plane. This approach enables the propagation of constraint information over the full conflict network, effectively mitigating the “ripple effect” induced by local collisions and allowing for globally consistent optimization. We validate the proposed method using both real observations from the Large Sky Area Multi-Object Fiber Spectroscopic Telescope and simulated dense focal-plane configurations. The results demonstrate that our approach achieves significantly higher weighted yields than traditional heuristic baselines, reaching up to 99.71% of the optimal solution while maintaining computational efficiency compatible with survey operations. The method consistently produces collision-free assignment plans and exhibits robust performance across varying target densities. Importantly, the proposed graph-based framework is independent of specific hardware geometries, making it directly applicable to a broad class of fiber-fed spectroscopic surveys, including Dark Energy Spectroscopic Instrument–like systems. This work establishes a scalable and physically motivated paradigm for fiber assignment, bridging combinatorial optimization and data-driven modeling, and provides a pathway toward improving survey completeness and mitigating selection biases in next-generation spectroscopic programs.