Stanford allowed GPT-6 Astra to directly modify robot parameters: only moving 65 lines of word embeddings, fixing a type of command failure

Stanford allowed GPT-6 Astra to modify robot parameters, increasing the success rate from 74.3% to 87.5%; only modifying 65 lines of word embeddings, fixing a problem where one type of robot 'failed when rephrased'. The task was to open the stove, and the researchers gave the robot the command 'activate the cooking applia...'