VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents | World Model + RL

VAGEN is a reinforcement learning (RL) framework that trains multi-turn VLM agents to build an internal world model via visual state reasoning. A 3B model scores 0.82 across five visual agent tasks, outperforming GPT-5 (0.75), Gemini 2.5 Pro (0.67), and Claude 4.5 (0.62). NeurIPS 2025.

关键词：world model, reinforcement learning, RL, world model RL, VLM agent, VLM agent RL, vision-language model agent, multi-turn RL, visual state reasoning, world model reasoning, POMDP, agent training, VAG…