CITE: Communication-Efficient Inference Token Exchange for Privacy-Preserving Collaborative Wireless Edge LLMs

Deploying Large Language Models (LLMs) at the wireless edge is hampered by the “communication wall”: traditional Federated Learning (FL), an offline training paradigm, necessitates exchanging gigabyte-scale parameters, violating real-time latency constraints over fading channels and exposing gradient-based privacy vulnerabilities. For collaboration among already deployed models, we propose Communication-efficient Inference Token Exchange (CITE), a decentralized inference-time framework. CITE employs a lightweight semantic signaling protocol governed by a dual-sided Sanitization-Decomposition mechanism, which bounds sequence-level leakage of sensitive private attributes under explicit context-leakage assumptions. To mitigate the impact of unreliable peers in non-stationary wireless networks, we design a Synergistic Trust Consensus algorithm incorporating a Temperature-Scaled Weighting scheme and a Threshold-Based Pruning strategy, effectively capturing combinatorial peer synergy while reducing computational complexity. An optional Novelty-Weighted Online Adaptation module with Momentum-based updates enables nodes to internalize collaborative knowledge after interaction while filtering gradient noise derived from fluctuating peer qualities. Theoretical analysis characterizes wireless latency and outage feasibility, amortized communication tradeoffs against PEFT-based FL, sequence-level privacy leakage, and convergence under stated assumptions. Extensive experiments on standard NLU and knowledge-intensive benchmarks demonstrate that CITE reduces communication overhead while outperforming state-of-the-art federated baselines in inference accuracy and privacy resilience.