Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models

World Action Models (WAMs) enhance robotic manipulation by predicting how an environment evolves. However, a major trade-off exists: Joint-WAMs provide high robustness by maintaining future-aware representations during inference but are computationally expensive, while Fast-WAMs improve efficiency by discarding future modeling at inference, often at the cost of generalization. This paper investigates whether future representations are essential for inference-time robustness and proposes a method to retain them efficiently.

The authors demonstrate that future representations are critical for handling out-of-distribution (OOD) scenarios. To bridge the efficiency-robustness gap, they introduce Faster-WAM, which uses a one-pass future-conditioning framework. Instead of repeatedly running the video branch, Faster-WAM computes future representations once and caches them. It employs two key innovations:

• SparseMoT: A sparse interaction mechanism that limits video-action communication to a subset of network stages, using lightweight action-only refinement in between.
• Interval KV-Fusion: A method to aggregate multi-depth future representations into a compact format, ensuring the action pathway receives rich temporal information without increasing attention complexity.

Faster-WAM achieves a superior performance-efficiency trade-off compared to existing models. On the challenging LIBERO-Plus benchmark, which tests robustness under various distribution shifts, Faster-WAM achieved a 73.57% success rate, significantly outperforming Fast-WAM (49.14%) and demonstrating higher efficiency than Joint-WAM. The model also achieved state-of-the-art results on the LIBERO and RoboTwin 2.0 benchmarks, confirming that selective, sparse future conditioning effectively preserves the robustness benefits of temporal reasoning while maintaining real-time performance.