Proves LLMs possess predictive partial-world models via task-agnostic affordances for intents. Introduces distribution-robust affordances for multi-task efficiency. Reduces search branching in robotics, outperforming full world models.
Key Points
- 1.Formal proof links affordances to partial modeling
- 2.Improves search efficiency in tabletop robotics tasks
Impact Analysis
Makes LLM world modeling practical by focusing on intent-linked states. Enhances robotics rewards via reduced branching factors.
Technical Details
Extracts partial models from affordance-linked state-actions. Uses distribution-robustness for multi-task generalization.
