LLM Inference at the Edge: Mobile, NPU, and GPU Performance Efficiency Trade-offs Under Sustained Load

Collection of arXiv papers (Mar 26 2026) covering edge LLM inference, blockchain and decentralized-systems architectures, peer-discovery, multi-VM Layer‑1 designs, research-compute tooling, and HPC/GPU scalability. Security-relevant items: (1) LLM Inference at the Edge documents thermal/OS-enforced throttling and power/thermal limits that can be exploited to cause sustained-performance degradation (resource-exhaustion/DoS) on mobile on-device agents; (2) AetherWeave proposes a stake-backed, privacy-preserving peer-discovery protocol with on-chain slashing and publicly verifiable misbehavior‑eff