LLaDA is a novel large language model (LLM) based on diffusion models instead of the traditional autoregressive approach. LLaDA employs a masking process, predicting masked tokens using a Transformer network. It demonstrates comparable performance to established LLMs like LLaMA3 in various language tasks. The model showcases strong scalability, instruction-following, and excels in reversal reasoning. This challenges the notion that autoregressive modeling is the exclusive path to achieving LLM capabilities, suggesting diffusion models offer a viable alternative. Further research directions include scaling LLaDA and exploring multimodal applications.#artificialintelligence #llm #llada Hosted on Acast. See acast.com/privacy for more information.

Ver todos los episodios