Learning to bid when the auctioneer is lying about competitors

Online bidders in repeated auctions face a subtle problem: when auctioneers inject fake competing bids, the feedback becomes misleading even though the actual winning prices stay the same. The learner sees inflated competition and must figure out the real bid distributions. This team designed an algorithm that uses two approaches in parallel—one robust branch ignores suspicious reports to learn at the standard auction rate, while an optimistic branch debiases the feedback when conditions are favorable. The result nearly matches the theoretical lower bound, showing that feedback manipulation genuinely makes learning harder, but remains beatable with the right strategy.