MAP: A Map-then-Act Paradigm for Long-Horizon Interactive Agent Reasoning

Researchers propose the Map-then-Act Paradigm (MAP), a plug-and-play framework for interactive LLM agents that shifts environment understanding before execution. Current agents suffer from Delayed Environmental Perception and an Epistemic Bottleneck due to reactive stepwise planning. MAP is inspired by human affordance perception and cognitive map theory, consisting of three stages: Global Exploration, Task-Specific Mapping, and Knowledge-Augmented Execution. Experiments on ARC-AGI-3 show consistent gains across benchmarks and LLMs, enabling frontier models to surpass near-zero-shot performance. The framework addresses the temporal inversion in goal-conditioned planning by acquiring environmental priors beforehand.