Large multimodal models are increasingly used as agents that must perceive, reason, and act in interactive multi-agent environments. However, according to a new arXiv paper, most existing benchmarks for such agents are turn-based, which does not reflect the continuous decision-making required in real-time settings.
The paper introduces ROAR, a benchmark designed to evaluate LMM agents under real-time constraints in multiplayer action-shooter games. By shifting from turn-based evaluation to a real-time multiplayer environment, ROAR aims to capture challenges that prior benchmarks miss, such as rapid reactions and simultaneous interactions.
Because this is a single source, there are no conflicting findings to compare. The announcement provides only the abstract, so details on the benchmark's tasks, metrics, and results are not yet available in this source.