Building a marine reasoning large model: a method based on structured chain-of-thought fine-tuning and knowledge graph

Abstract To address the long-standing professional knowledge bottlenecks in scientific marine research and aquaculture, this paper proposes a marine reasoning large language model construction framework based on structured reasoning chain-of-thought (SRCoT) fine-tuning and a knowledge graph (KG). To implement the framework, an indent-driven article heuristic search method is first adopted to construct a marine-domain-specific dataset, followed by the development of a sliding window and weight-matrix-based strategy for dataset deduplication. Subsequently, a marine-domain KG is constructed, and an entity entailment method based on pointwise mutual information vectors is designed. Finally, a model post-training approach integrating SRCoT and three-stage direct preference optimization (DPO) is proposed. The base model is fine-tuned on the marine-domain SRCoT dataset and post-trained using the three-stage DPO strategy. During deployment, the custom-built marine-domain KG is used as an external reference to enhance the model responses. The experimental results demonstrate that the model trained with the proposed framework achieves performance improvements in marine-domain complex reasoning tasks and is effective in mitigating over-reasoning and refining model responses.