Improving Large Language Model preference prediction by averaging two perspectives

People often seek advice when choosing a film, a book, or another experience whose value depends on their tastes. Large language models (LLMs) provide a convenient way to obtain such advice. We test a simple method for improving their predictions, inspired by the wisdom of the inner crowd. An LLM first predicts a person’s movie rating from that person’s past ratings. It then predicts how the general public, given the same history, would predict the same person’s rating. Averaging takes the arithmetic mean of these two responses. We evaluated 12 GPT model snapshots using MovieLens 1M. A history-size study tested 1–20 ratings for 55 users and extended the comparison to 55 ratings for 32 eligible users. Averaging improved mean squared error relative to the first response even at the longest history. A subsequent evaluation used 10 target films per person with 10 fixed history ratings. Across 320 user–movie pairs, mean squared error was 1.218 for the first response, 1.147 for the public-perspective response, and 1.150 for Averaging. Additional analyses showed that Averaging reduced errors of at least 1.5 rating points, whereas the first response produced more exact matches. These advantages did not extend to every error threshold or metric. The findings connect inner-crowd research with a simple conversational method for preference prediction and show why the choice of response should reflect the prediction objective.