Using LLMs to Directly Guess Conditional Expectations Can Improve Efficiency in Causal Estimation