AgentGarten: Code Worlds for Evolving Agents
AgentGarten claims agents learned in 4 rounds where a conventional reinforcement-learning baseline needed millions.
The paper presents a framework for building interactive code-defined worlds with persistent state, programmed rules, and real-time visual observations. Its renderer adapts a pretrained video model to structured geometry conditions, with “Adversarial Forcing” used to improve how prior observations shape later predictions. Agents use each round of experience to produce playbooks that later agents inherit and refine. The authors frame this as a way to scale training environments in both number and difficulty. Source: HF Daily Papers' note.
The paper presents a framework for building interactive code-defined worlds with persistent state, programmed rules, and real-time visual observations. Its renderer adapts a pretrained video model to structured geometry conditions, with “Adversarial Forcing” used to improve how prior observations shape later predictions. Agents use each round of experience to produce playbooks that later agents inherit and refine. The authors frame this as a way to scale training environments in both number and difficulty. Source: HF Daily Papers' note.
score 5