Language models reportedly show some human-like bidding patterns in auction tests

A researcher involved in the work says language models could generate synthetic bidding data to help test auction designs. The models produced bid distributions that matched human ones ordinally and showed behaviors such as the winner’s curse and last-minute bidding in eBay-style auctions. They also overbid more in second-price auctions and seemed more loss-averse. The researcher says newer models can give textbook answers, making those patterns harder to draw out.