人工知能学会

Large language models (LLMs) have been widely utilized due to their high generalizability. However, practical applications face several challenges. For instance, large-scale models, such as those with 70B or 100B parameters, demand significant computational resources to achieve high-precision inference, while smaller models, such as 3B, often underperform. Furthermore, the inference process is highly sensitive to the examples included in the prompt. To address these issues, some studies employ ensemble-like methods that unify multiple inferences from different prompts or models for one sample. While these approaches can improve the accuracy, they often risk overfitting to the validation data, potentially compromising generalizability. In this study, we propose a robust multi-level ensemble method that dynamically calculates model weights at both the model and sample levels to enhance accuracy and generalizability. Comparative validation using an established benchmark demonstrates that the proposed approach consistently outperforms individual models.