Researchers Use Mortal Kombat II to Test Multimodal AI Models in Real-Time Competition

Evaluating large multimodal models (LMMs) on static benchmarks often misses their ability to perform under pressure in dynamic, real-world scenarios. Researchers have introduced a fresh approach: LM Fight Arena, which tests these AI systems by having them compete in Mortal Kombat II, the iconic 1990s arcade fighter.

A Game-Based Testing Framework#

The benchmark places six leading large multimodal models—both open-source and proprietary—into the fighting game as competing agents. Each model controls the same character to ensure fair comparison, tasked with interpreting individual game frames and performance metrics to decide what action to take next. Rather than answering multiple-choice questions or describing static images, the models must reason rapidly and adapt their strategy across multiple rounds in a tournament setting.

Why This Matters#

Traditional benchmarks rely on fixed tasks that don't mirror real-time, adversarial conditions. LM Fight Arena eliminates subjectivity by automating the entire evaluation process, making results reproducible and objectively measurable. The fighting game environment demands both visual perception—recognizing opponent movements and game states—and sequential tactical reasoning, as each decision affects future gameplay.

Bridging AI and Entertainment#

By grounding model evaluation in an engaging competitive format, this work opens a pathway between rigorous AI assessment and interactive gaming. The researchers note that their framework is "fully automated, reproducible, and objective," offering a more naturalistic test of how LMMs perform when decision-making must happen under pressure.