GigaWorld-Policy: An Efficient Action-Centered World–Action Model

Technical Analysis is an action-centered world–action model that learns 2D pixel–action dynamics while enabling action-only decoding for efficient inference High-level Analogy: Imagine you're training a robot chef to cook. Traditional World Models are like training the chef to meticulously plan out every single step, including precisely visualizing how each ingredient will look after every cut, stir, or pour, before actually making the move. This helps them learn the consequences of their actions, but when it's time to cook for real, they still try to visualize every future detail on the fly. This makes them very slow, and if their mental visualization of a future step is slightly off, it can throw off the whole dish. GigaWorld-Policy is like a smarter chef. During training, this chef also learns by visualizing what the ingredients will look like after each action. This deep understanding of how actions change the world makes them a better learner. However, when it's time for the actual cooking (inference), they don't waste time visualizing every single future detail. Instead, they've learned the direct connection between seeing the current state of the ingredients and knowing exactly what action to take next. The 'future visualization' becomes an internal reasoning tool used for learning, not a mandatory step for real-time execution. They know the consequences without needing to explicitly 'see' them, allowing them to cook much faster and more reliably. Motivation of the…