Answers are grounded in this paper and cite their sources.
The article discusses the task of automating the generation of mission programs for autonomous underwater vehicles (AUV) using neural networks. The main focus is on choosing the optimal model for converting commands in natural language into code in a specialized underwater research programming language (URPL). LLaMA 3.1 8B-instruct, Gemma 2 9B IT, Mistral 7B-instruct, Qwen 7B-instruct, Phi-4 3.8B models are compared according to the criteria of generation accuracy, execution speed, resistance to errors in input data and compactness of the model when solving the given problem. The analysis results showed that the LLaMA 3.1 8B-instruct and Phi-4 3.8B models demonstrate the best performance. However, LLaMA 3.1 stands out for its higher processing speed (56.08 tokens per second) and the ability to correctly interpret commands, which makes it preferable for the tasks of generating AUV missions. To improve the accuracy and adaptability of the model, an approach to additional training of the neural network on an extended dataset created using more powerful language models is proposed.
No takes yet. Share an insight, caveat, or question.
Pugachev et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: