Researchers release OceanCoder model for automated modeling in ocean engineering
A research team published the OceanCoder model on arXiv, a fine-tuned version of Qwen3-8B for automated creation of simulation models in ocean engineering. Pass@1 rose from 0% for the base model to 51.43% after fine-tuning and RL optimization.
A research team published a preprint on arXiv describing the OceanCoder model — a fine-tuned version of the Qwen3-8B model designed for automated creation of simulation models in ocean engineering. According to the authors, creating such models typically requires expert knowledge, and mere syntactic correctness of the generated code does not by itself mean that the resulting model and its outputs meet engineering requirements.
The model was created by fine-tuning on 12 000 verified supervised fine-tuning (SFT) examples across 12 categories of engineering tasks, created using a custom data pipeline. Part of the solution is a three-layer verification framework that checks the model's inputs against engineering parameters and task constraints, verifies completion of the required simulation analyses, and assesses the engineering plausibility of the computed outputs. The results of this verification serve as a graded reward for optimization using group relative policy optimization (GRPO).
The authors also created the first benchmark for automated modeling in ocean engineering, which includes out-of-training-distribution tasks and real-world engineering cases. According to the experimental results, Pass@1 rose from 0% for the base model to 46.77% after SFT and to 51.43% after GRPO optimization, with the model, according to the authors, outperforming both the base models and the compared closed-source model.
Why it matters
The paper shows how combining a verification framework with reinforcement learning (GRPO) can improve the ability of language models to generate engineering-valid simulation models in a specialized domain with a shortage of training data. For the field of automated modeling, the "engineering verification as training signal" approach is particularly relevant, as is the newly created benchmark, which makes it possible to compare similar solutions on both real-world and out-of-distribution ocean engineering tasks.
Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.