SiteEvo: An LLM Agent for Autonomous Discovery of Site-Adaptive Massive MIMO Detectors

Abstract

Designing a massive multiple-input multiple-output (MIMO) detector that is tailored to each specific deployment site can improve detection performance, yet most existing detectors, including deep-unfolding methods, adapt mainly within human-prescribed algorithmic structures. This letter proposes SiteEvo, a site-adaptive agentic discovery framework that extends adaptation from parameter tuning to detector-algorithm discovery. Rather than selecting or recombining predefined architectures, the large language model (LLM) agent synthesizes complete executable detector programs under a common design contract by combining its pretrained communication knowledge, site-specific wireless evidence, and experience accumulated during the search. Each proposal defines its own processing structure and trainable parameterization, and the parameters are optimized by a communication-system backend. The resulting high-performance detector is further structurally refined to reduce block-level computational cost under a bit error rate (BER)-preservation constraint and is finally frozen for deployment without online LLM inference. Simulations across three sites demonstrate BER reductions by factors of 2.6 to 6.0 over the best-performing competing detector at each site at 6 dB, while structural refinement reduces block-level computational cost by 16.7-34.1% under a validation-BER constraint.