Researchers at the University of Washington propose an innovative agent tuning method that efficiently optimizes large language models without requiring access to model weights. This method guides the predictions of the base model toward the tuned model by comparing the predictions of a small tuned model and an untuned model, thereby improving model performance and better retaining training knowledge. This breakthrough technology was verified in fine-tuning experiments on the original 13B and 70B models of LLAMA-2, demonstrating its significant efficiency advantages.
Webmaster Home reported that the University of Washington launched a proxy tuning method that can achieve efficient tuning of large models without touching the model weights by comparing the prediction results of small adjusted models and unadjusted models. This method can better retain training knowledge during decoding and improve tuning efficiency. The agent tuning performance was verified by researchers fine-tuning the 13B and 70B original models of LlAMA-2. This method compares the output prediction distributions of the basic model M and the tuning model M+, and guides the predictions of the basic model to move in the direction of the tuning model. It is an innovative tuning method. The agent tuning method provides a solution for efficient tuning of large models while better retaining training knowledge during decoding, which is expected to bring new enlightenment to the AI field.This new method provides a more convenient and efficient way for large model tuning, reduces the need for direct manipulation of model weights, and effectively improves model performance and knowledge retention capabilities, bringing new possibilities to the development of the field of artificial intelligence. sex. In the future, this method is expected to be applied in more large-scale language model tuning, further promoting the progress of AI technology.