Wireless Context Engineering for Efficient Mobile Agentic AI and Edge General Intelligence

TL;DR

This work shifts the focus of wireless edge intelligence from indiscriminate model scaling to principled context engineering, defining wireless context and a framework (WCCF) to adaptively manage heterogeneous context under inference-time constraints. It formalizes five design dimensions (acquisition, structuring, compression, persistence, delivery) and introduces robust performance metrics to evaluate context utility. A central contribution is the WCCF architecture, which orchestrates context via a Context Construction module, on-demand transmission guided by reinforcement learning, and a masked multimodal inference engine, demonstrated through an ISAC-enabled V2I beam-prediction case study. Results show adaptive context selection can achieve near full-context accuracy with substantially lower sensing and processing costs, highlighting a practical path to efficient edge intelligence in dynamic wireless environments.

Abstract

Future wireless networks demand increasingly powerful intelligence to support sensing, communication, and autonomous decision-making. While scaling laws suggest improving performance by enlarging model capacity, practical edge deployments are fundamentally constrained by latency, energy, and memory, making unlimited model scaling infeasible. This creates a critical need to maximize the utility of limited inference-time inputs by filtering redundant observations and focusing on high-impact data. In large language models and generative artificial intelligence (AI), context engineering has emerged as a key paradigm to guide inference by selectively structuring and injecting task-relevant information. Inspired by this success, we extend context engineering to wireless systems, providing a systematic way to enhance edge AI performance without increasing model complexity. In dynamic environments, for example, beam prediction can benefit from augmenting instantaneous channel measurements with contextual cues such as user mobility trends or environment-aware propagation priors. We formally introduce wireless context engineering and propose a Wireless Context Communication Framework (WCCF) to adaptively orchestrate wireless context under inference-time constraints. This work provides researchers with a foundational perspective and practical design dimensions to manage the wireless context of wireless edge intelligence. An ISAC-enabled beam prediction case study illustrates the effectiveness of the proposed paradigm under constrained sensing budgets.

Figures (4)

• Figure 1: Illustration of wireless context engineering for agentic intelligence in wireless networks. Part A shows multi-layer wireless context spanning physical, network, environmental, and service/task levels. Part B illustrates wireless context engineering across five dimensions, including acquisition, structuring, compression, persistence, and delivery. Part C presents context-conditioned prediction, generation, and decision-making enabled by agentic models in dynamic wireless environments.
• Figure 2: WCCF and ISAC-enabled beam prediction case study. (A) WCCF architecture with context construction, transmission, and inference, enabling closed-loop adaptive context engineering. (B) ISAC-enabled V2I beam prediction scenario with the objective of maximizing prediction accuracy while minimizing high-cost context usage. (C) Network structure of WCCF, including modality-specific encoders, a multimodal Transformer, and an RL-based policy.
• Figure 3: Top-3 beam prediction accuracy under different context configurations
• Figure 4: Evaluation reward of the RL-driven WCCF policy during training, compared with fixed context baselines