World Action Models: A Survey
This survey reviews World Action Models (WAMs), which integrate world models with action generation for decision-making tasks. It covers key components, architectures, and training methods, and discusses applications in robotics, gaming, and autonomous systems, as well as open challenges.
Background
- This paper surveys "World Action Models" (WAMs), a new AI paradigm that learns a model of how the world behaves (a "world model") and then uses it to plan actions by predicting future outcomes — like having a mental simulator before acting.
- Key context: This is a step beyond the "next-token prediction" approach used by large language models (LLMs) like GPT and Claude. LLMs learn patterns from text; WAMs try to learn causal dynamics of real or simulated environments.
- Related to "model-based reinforcement learning" (MBRL) and "world models" from the 2018 paper by Ha & Schmidhuber, but WAMs emphasize joint training of the world model *and* an action policy, rather than using a separate planner.
- Important because it could lead to AI that generalizes better, requires less training data, and can reason about consequences — relevant to robotics, autonomous driving, and game AI.