Planning with Language and Generative Models: Toward General Reward-Guided Wireless Network Design

TL;DR

This work tackles indoor AP deployment planning under complex geometries and non-convex propagation by reframing the task as reward-guided generative inference with a unified reward function. It contrasts an agentic LLM reasoning loop with verifier-based feedback against diffusion-based sampling under a Gibbs form $p(\mathbf{P}|\mathcal{D}) \propto e^{\beta r(\mathbf{P},\mathcal{D})}$, showing diffusion yields more robust global optimization by implicit score smoothing. A large-scale real-world dataset requiring over $50k$ CPU hours to train general reward functions enables strong in-distribution and out-of-distribution generalization. The work also provides theoretical insights that reward estimation quality is the primary bottleneck and demonstrates practical scalability for domain-agnostic indoor AP planning.

Abstract

Intelligent access point (AP) deployment remains challenging in next-generation wireless networks due to complex indoor geometries and signal propagation. We firstly benchmark general-purpose large language models (LLMs) as agentic optimizers for AP planning and find that, despite strong wireless domain knowledge, their dependence on external verifiers results in high computational costs and limited scalability. Motivated by these limitations, we study generative inference models guided by a unified reward function capturing core AP deployment objectives across diverse floorplans. We show that diffusion samplers consistently outperform alternative generative approaches. The diffusion process progressively improves sampling by smoothing and sharpening the reward landscape, rather than relying on iterative refinement, which is effective for non-convex and fragmented objectives. Finally, we introduce a large-scale real-world dataset for indoor AP deployment, requiring over $50k$ CPU hours to train general reward functions, and evaluate in- and out-of-distribution generalization and robustness. Our results suggest that diffusion-based generative inference with a unified reward function provides a scalable and domain-agnostic foundation for indoor AP deployment planning.

Figures (11)

• Figure 1: The LLM operates as a a agentic planner that takes the environment and task context as input, proposes candidate AP deployments, and iteratively refines them based on feedback from a physical verifier. Deployment actions are evaluated using 3D ray-tracing simulator (verifier), producing quantitative scores and visual coverage maps. Evaluation results are written to a deployment history module, which tracks the best solution and provides reference feedback for subsequent iterations, forming a closed-loop optimization.
• Figure 2: Non-convex and fragmented reward landscape over a given floor plan. The surface shows the reward value associated with each candidate access point (AP) location, where warmer colors indicate higher rewards. The shaded region projected onto the ground plane corresponds to the underlying floor plan, illustrating how spatial constraints shape the reward distribution.
• Figure 3: AP placement planning results on two indoor floorplans. The background colormap represents the reward heatmap, where warmer colors indicate higher expected deployment rewards. The star marker denotes the selected AP location, while the trajectories illustrate the intermediate solution refinement processes of different methods. From left to right, we compare the proposed agentic method with convex optimization, diffusion sampling, and weighted sampling. The results reveal distinct planning behaviors across methods and demonstrate the effectiveness (less iterative steps) of the diffusion sampling in identifying high-reward AP placements under complex indoor layouts.
• Figure 4: AP planning based on diffusion sampling over a multi-level building. The left panels show the 3D building geometry from Level 1 to Level 4. The right panel visualizes the final deployment produced by diffusion-based planning, where the heat map indicates the pointwise satisfication for building and blue nodes denote the AP locations. See visualization of intermediate reasoning and iterative optimization states in Figures \ref{['fig:reasoning_level_1']}, \ref{['fig:reasoning_level_2']}, \ref{['fig:reasoning_level_3']}, \ref{['fig:reasoning_level_4']} in Appendix \ref{['append:Visualization of Intermediate Reasoning and Iterative Optimization States']}.
• Figure 5: Visualization of reward-guided diffusion trajectories. The shaded regions on the ground plane depict the indoor floor plan, while the colored surfaces represent the reward landscape over AP deployment space. The red trajectories illustrate reverse-time SDE sampling paths guided by the reward-induced score. Due to Gaussian mollification under the VE SDE, the score function computed with a coarse-to-fine structure: large noise levels smooth non-convex and fragmented reward regions, enabling globally coherent motion, while progressively smaller noise refines the trajectory toward high-reward optima. In contrast to convex optimization in Figure \ref{['fig:AP_iternative_refinment']}, which is prone to local traps, the injected VE noise allows diffusion sampling to overcome disconnected modes and navigate complex reward landscapes effectively.

Theorems & Definitions (2)

• Lemma 4.1: Girsanov Theorem for Drift Perturbation shiryaev2016probability
• Lemma 4.2