DeepRMM-01 at GLEE: From Recursive Opponent Modeling to Specialized Control

GLEE is a live competition in which agents repeatedly play 3 language-based economic games under varied configurations and partial identity disclosure.DeepRMM-01 tested whether persistent recursive mental modeling (RMM) could move ahead of adapting opponents through predictions of their responses and of theagent’s public predictability. The implementation evolved from an action–belief desire–emotion ledger, prose dossiers, and per-move deliberation into a hybridcontroller combining deterministic family policies, statistical opponent packages, action-conditioned response prediction, a public self-mirror, and bounded languagemodel reasoning. At a frozen 9,587-game training frontier, a validation-selected response stack attained equal-family negative log likelihood (NLL) 0.326, compared with 0.329 for its sequence-only component and 0.452 for its engineered-only component; a Persuasion continuation head reduced game-macro NLL from 0.702to 0.614 (paired 95% interval for the difference [-0.141,-0.034]). Two public self-mirror seeds yielded descriptive action accuracy of 84.0–84.3%. Across the complete journal frontier, all 97,473 submitted moves were valid; DeepRMM-01 finished 59th with final dashboard Bargaining/Negotiation/Persuasion ratings of 1899.7/1794.1/1900.9. Within this development path, explicit mental-state prose showed no measurable incremental value, while compact statistical and action conditioned models became operationally usable under bounded authority.