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

The evolution of robot control has increasingly integrated deep learning-based future prediction, moving from classical model-based planning (Schrittwieser et al., 2020) to end-to-end video-conditioned policies (Li et al., 2026). Recent advances like vision-language-action models (Zitkovich et al., 2023) have demonstrated the potential of direct observation-to-action mappings, reducing inference latency. However, these approaches often depend on iterative video generation or multiple forward passes, which significantly increase deployment latency, limiting real-time applicability. Efforts such as Fast-WAM and PFD aim to reduce this overhead by removing explicit future representations during inference, but they still rely on training-time future conditioning or distillation, leaving a gap in achieving both efficiency and high success. The challenge remains to develop models that can perform future-conditioned control with minimal latency while maintaining robustness and accuracy.

The core problem addressed is how to eliminate the latency introduced by multi-step video rollouts in world action models without sacrificing the ability to condition actions on future states. Existing methods depend on iterative generation, which is computationally expensive and unsuitable for real-time deployment. The question is whether a fixed, complete future cache can be constructed in a single pass, enabling immediate access to future information during control. Additionally, understanding the dependency of action policies on the evolving versus static future representations is critical. Overcoming these issues requires innovative methods that can prefill future caches reliably, preserve the future read interface, and ensure high task success in complex environments.

The key innovations include: 1) the RIFT framework, which uses learned anticipation tokens to fill the entire future K/V cache in a single backbone pass, replacing iterative rollout; 2) a causal intervention protocol that manipulates future cache content to analyze its influence on execution; 3) integration of conditional flow matching to shape anticipation states, enabling accurate one-pass cache construction; 4) maintaining the original future-read interface, ensuring compatibility with existing WAM architectures; and 5) extensive validation demonstrating near-equivalent success rates with significantly reduced latency. These innovations collectively redefine how future information is represented and utilized in robotic control.

• �� Construct anticipation tokens: learn a set of tokens E that encode future spatial-temporal information for each layer.
• �� Cache prefill: map the first-frame observation and anticipation tokens to full future K/V caches using a shared backbone, ensuring alignment with the original future-read interface.
• �� Causal intervention: mask, reassign, or fix the cache content during evaluation to analyze the dependency of actions on future values.
• �� Conditional flow matching: train anticipation tokens by modeling the distribution of future states, enabling accurate prediction without multi-step generation.
• �� Training process: combine native video supervision, action flow matching, and auxiliary probe losses to optimize anticipation tokens.
• �� Evaluation: measure success rate and end-effector displacement errors across various interventions to validate the importance of future cache content.
• �� Deployment: perform a single cache prefill per action chunk, then execute actions solely based on the fixed cache, avoiding iterative video generation.

Experiments involve evaluating RIFT on LIBERO's 40 tasks and RoboTwin-2.0's diverse scenes, comparing success rates and latency against baseline models like Joint, IDM, and Fast-WAM. Each model undergoes 2000 trials per task, with success rate and end-effector displacement as primary metrics. Ablation studies test the impact of cache interventions, anticipation token dimensions, and training objectives. Hyperparameters such as cache size, number of anticipation tokens, and training curriculum are optimized for performance. Results demonstrate that RIFT achieves high success with substantially lower latency, validating the hypothesis that a fixed, complete future cache suffices for effective control.

RIFT attains 98.8% success on LIBERO, matching or exceeding rollout-based methods like Joint (98.4%) and IDM (98.6%), while reducing action chunk latency by 68.2% to 89.1%. On RoboTwin-2.0, success rates reach 92.9% in clean scenes and 92.6% in randomized scenarios, outperforming all other methods. Intervention experiments show that masking or reassigning future cache values significantly degrades performance, confirming the reliance of action policies on future content. Replaying a fixed final-clean cache nearly preserves original execution, indicating that a static cache can be sufficient. These findings demonstrate that one-pass cache construction can match the effectiveness of multi-step rollout, with major gains in efficiency.

This approach is directly applicable to real-time robotic control, autonomous systems, and industrial automation where low latency and high reliability are crucial. It enables deploying high-fidelity future-conditioned policies without the computational overhead of iterative video generation. The method can be integrated into existing WAM architectures with minimal modifications, facilitating rapid adaptation in dynamic environments. Long-term, this framework could support multi-agent coordination, adaptive control in unpredictable scenarios, and scalable deployment of autonomous systems in complex real-world tasks.

The effectiveness depends on the quality of learned anticipation tokens, which may degrade in highly unpredictable or novel environments. The fixed cache may lack sufficient detail in highly dynamic scenes, affecting control accuracy. Although latency is reduced, the initial cache prefill step introduces some overhead, and real-time online updates remain challenging. Further research is needed to enhance robustness, generalization, and online adaptability, especially in real-world, high-variance settings.