Keep the Future, Drop the Rollout: RIFT for World Action Models

2026-08-11

AI Summary

Existing World Action Models (WAMs) rely on iterative video rollouts for future prediction, resulting in high latency. This work proposes RIFT, the first method to enable non-iterative, future-conditioned modeling by learning "foresight tokens" that allow construction of a complete future key-value cache in a single forward pass while preserving the original read interface. By decoupling cache generation from usage, RIFT achieves a success rate of 98.8% on LIBERO with a 68.2%–89.1% reduction in action-chunk latency, and attains state-of-the-art performance on RoboTwin 2.0 with success rates of 92.9% and 92.6%, significantly improving inference efficiency without compromising task performance.

Abstract

World action models (WAMs) condition robot actions on predicted futures, but iterative video rollout increases deployment latency. We ask whether action generation requires the evolving rollout trajectory or only its future representation. Across four WAMs on all 40 LIBERO tasks, paired closed-loop interventions show that masking or reassigning future-cache values changes execution and reduces success, indicating sensitivity to future values and their assigned positions. For Joint and Cosmos-2, however, replaying one fixed final-clean key/value (K/V) cache nearly preserves unmodified execution, with $1.7$ to $1.9$~cm end-effector average displacement error and $97.9\%$ to $98.2\%$ success. This separates cache consumption from production: these models can reuse a fixed cache but still require iterative rollout to construct it. We therefore propose RIFT (\emph{Rollout-free Imagination via Future Tokens}), which uses learned anticipation tokens to construct a complete future K/V cache in one backbone pass while retaining the original future-read interface. On LIBERO, RIFT achieves $98.8\%$ success, close to rollout-based Joint, IDM, and LingBot-VA at $98.4\%$ to $98.6\%$, while reducing action-chunk latency by $68.2\%$ to $89.1\%$. On RoboTwin~2.0, RIFT reaches $92.9/92.6\%$ on clean/randomized scenes, the highest observed among the evaluated methods. These results support rollout-free future conditioning without iterative video generation at deployment.

Problem

• world action models
• video rollout
• deployment latency
• future conditioning
• action generation

Innovation

• Rollout-free
• Future Tokens
• World Action Models
• KV Cache Reuse
• Deployment Latency Reduction