Quo Vadis, World Modeling?
The paper argues that world models should serve agents as feedback proxies, not just predictors of future physical states.
The authors define “Agent-Centric Interactive World Proxies” as systems that return agent-usable information: outcomes, memories, skills, or verification signals. They group these proxies into six forms: dynamics, spatial, execution, memory/experience, skill, and reward/verification. The paper also lays out three use levels, from inference-time guidance to training-time optimization to agent-proxy co-evolution. HF Daily Papers' note
The authors define “Agent-Centric Interactive World Proxies” as systems that return agent-usable information: outcomes, memories, skills, or verification signals. They group these proxies into six forms: dynamics, spatial, execution, memory/experience, skill, and reward/verification. The paper also lays out three use levels, from inference-time guidance to training-time optimization to agent-proxy co-evolution. HF Daily Papers' note
score 4