A new research paper claims "a probabilistic version of genuine reasoning" in current LLM AIs