Maybe the model doesn't do multiple queries but just one long query guided by thought tokens.

It could just be programmed to follow up by querying itself with a prompt like "Come up with arguments that refute what you just wrote; if they seem compelling, try a different line of reasoning, otherwise continue with what you were doing." Different such self-administered prompts along the way could guide it through what seems like reasoning, but would really be just a facsimile thereof.