LLM Test-Time Ensemble

A technique that generates multiple answers during the inference stage of a large language model and aggregates them by weight to improve performance. It requires no increase in the number of model parameters, and the weights can be learned from training data.