GAM-Agent: Game-Theoretic Multi-Agent Framework Enhances Vision-Language Reasoning

A new multi-agent framework named GAM-Agent has been developed by researchers to enhance vision-language reasoning. Unlike earlier models that rely on a single agent, GAM-Agent approaches reasoning as a non-zero-sum game, involving base agents that focus on specific visual perception tasks and a key agent responsible for ensuring logical consistency and factual accuracy. These agents exchange structured claims, evidence, and uncertainty assessments. An adaptive controller responds to disagreements or ambiguities by initiating multi-round discussions, resulting in more reliable and interpretable outcomes. Testing on four demanding benchmarks—MMMU, MMBench, MVBench, and V*Bench—shows that GAM-Agent significantly outperforms various vision-language model (VLM) backbones, particularly improving the accuracy of smaller models like Qwen2.5-VL-7B and InternVL2.5-8B, achieving results on par with or surpassing larger counterparts. This game-theoretic framework presents a fresh strategy for advancing AI reasoning, with promising implications for intricate visual reasoning challenges.