“A research paper by Google DeepMind's Rohan Shah establishes upper bounds on the reasoning capabilities of different model architectures, such as standard transformers, without using a chain of thought.”