LSP-Swarm: A Learned Semantic Protocol for Resilient Multi-Hop Distributed Inference

Distributed intelligence in communication-denied environments (e.g., UAV swarms, underground sensor networks, and underwater links) faces a structural constraint: transmitting raw sensor streams or continuous intermediate activations is often infeasible, while executing large monolithic models on every node violates size-weight-power-and-cost (SWaP-C) limits. Existing approaches split between networking-grade reliability mechanisms that preserve packets but do not preserve semantics, and learned transmission schemes that compress effectively in single-hop settings but can degrade under multi-hop impairments, producing hop-wise drift and brittle downstream behavior. We introduce LSP-Swarm, a protocol-first framework for resilient multi-hop semantic inference that treats intermediate representations as a communicated language rather than brittle tensors. LSP-Swarm packetizes semantics into a hierarchical tuple 𝑇 = [𝐻 slow , 𝐵 fast ], where a low-entropy, task-relevant header 𝐻 slow , trained under an Information Bottleneck perspective, serves as a stable semantic anchor, and a higher-capacity token payload 𝐵 fast carries conditional detail. Relay nodes perform integrity-aware semantic decode-and-forward by combining lightweight hard checks (CRC/ECC) with a soft semantic consistency score 𝑆 trust that deterministically selects repair-andforward, forward-with-uncertainty, or header-only fallback. To mitigate anchored failures, LSP-Swarm supports compact header appeals that are resolved by multi-view arbitration at the sink. Because the contribution is protocol and systems design, we define a reproducible evaluation suite with auditable bitrate accounting derived directly from packet budgets, standardized topology and channel stress regimes, and integrity decision metrics (AUROC, calibration, FAR/FDR) to enable independent validation and collaboration.