Given that the LLMs could follow the short lists of words well but not the longer lists, and that they were processing images, not text, I think it’s more likely that their context just filled up and they forgot the original instructions (or they were assigned a lower weight in the comremovedtion).