Fool You Twice? Reciprocal Opponent Shaping in Multi-Round Persuasion Games

Unlike one-shot persuasion, multi-round persuasion games have interesting characteristics that enable an easier exploitation for both sides of the game, while also posing unique challenges at the same time, mostly regarding the improvement of the other side as the game progresses, as well as providing an opportunity to understand, and potentially exploit the opponent's strategy. In this paper, we introduce BizarreBazaar, a bidirectional agent for repeated persuasion games, made for the 2026 GLEE competition. As a seller, BizarreBazaar combines persistent belief modeling over the game-local buyer's hidden valuation, which it calculates and attempts to shift, using a trust-capital mechanism that can be spent with low-quality recommendations and gained via high-quality ones. As a buyer, it performs opponent shaping through strategic refusal, sometimes rejecting even positive-EV recommendations when doing so is predicted to induce more informative future signaling. The final iteration of the Agent was run on 1,044 games (530 as a buyer, 514 as a seller). Ultimately, the agent finished 4th in the persuasion competition. These findings show us that through utilizing and estimating the methods shown in this paper, an agent can optimize multi-round persuasion games in an efficient manner.