Collaborative Autonomy of USV Swarms via Large-Small Language Model Synergy

Abstract：Objective To address the challenges of high reasoning latency, redundant logical content, and poor real-time performance when applying Chain-of-Thought (CoT) reasoning in large language models (LLMs) to multi-agent collaborative tasks, this paper proposes a collaborative autonomy algorithm for unmanned surface vehicle (USV) swarms based on the coordination between large and small language models. Methods Firstly, sentence-level masking and word-level semantic pruning are used to remove redundant information from CoT reasoning chains, resulting in lightweight keyword sequences. These sequences are then used to fine-tune lightweight models to enable them to generate essential CoT content from task descriptions. Secondly, a multi-model collaborative mechanism is introduced, in which three 8B lightweight models generate candidate decision solutions in parallel. A 32B-scale verifier model then performs confidence-based evaluation and selects the optimal solution, forming a closed-loop pipeline of “compression–generation–verification.” Results Experimental results in a dynamic obstacle avoidance scenario for unmanned surface vehicles (USVs) show that the proposed method reduces single-step reasoning latency from 3.45 seconds to 1.53 seconds, while maintaining a task success rate of 98.8%. The collaboration efficiency reaches 84.6%, outperforming traditional CoT-based acceleration approaches. Conclusions The proposed method significantly reduces inference latency without compromising output quality, offering an effective technical solution for real-time decision-making in complex maritime environments.