RevengeBench

About RevengeBench

RevengeBench is a benchmark framework designed to evaluate the ability of learners to reconstruct runnable programs from the behavioral traces of opaque target agents. It frames this as an inverse problem in code-space, where a learner observes a hidden target playing against sampled opponents in various programming-game arenas. The benchmark includes 75 LLM-generated, Elo-calibrated target policies across five specific environments: BattleSnake, Halite, HuskyBench, RoboCode, and RobotRumble. The framework supports a protocol that combines passive observation with constrained active intervention. Learners can author "probe opponents"—runnable policies that interact with the target to induce new trajectories—without having privileged access to the target's source code or internal state. Evaluation is based on action-distance on held-out states and performance in downstream player-versus-player (PvP) tournaments. RevengeBench is built upon the CodeClash arena and execution abstractions, extending them with specific tournament, agent, and trace parsing logic.

Summarised from the official page linked above.