I don't think this is showing LLMs performing decoding. They're just using the LLM to propose possible words. The decoding is done by using another model to score how well a proposed word matches brain activity, and using that score to select a most likely sequence given the proposals from the LLM.