Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning

TL;DR

Fast-ThinkAct tackles the latency bottleneck of reasoning in Vision-Language-Action tasks by replacing long textual chain-of-thought traces with compact latent reasoning. It uses a teacher-student framework where a textual teacher generates CoTs and a verbalizable latent student distills high-quality reasoning via reward preferences, coupled with action-aligned visual plan distillation to transfer spatial planning. The latent reasoning then guides a diffusion-based action model to produce executable robot actions, enabling up to $89.3\%$ inference-latency reduction while preserving long-horizon planning, failure recovery, and few-shot adaptation across diverse benchmarks. This approach yields substantial practical impact for real-time embodied AI, combining efficient internal reasoning with robust multimodal control. All mathematical notation is preserved in $...$ format to maintain precision and compatibility with downstream indexing and search systems.

Abstract

Vision-Language-Action (VLA) tasks require reasoning over complex visual scenes and executing adaptive actions in dynamic environments. While recent studies on reasoning VLAs show that explicit chain-of-thought (CoT) can improve generalization, they suffer from high inference latency due to lengthy reasoning traces. We propose Fast-ThinkAct, an efficient reasoning framework that achieves compact yet performant planning through verbalizable latent reasoning. Fast-ThinkAct learns to reason efficiently with latent CoTs by distilling from a teacher, driven by a preference-guided objective to align manipulation trajectories that transfers both linguistic and visual planning capabilities for embodied control. This enables reasoning-enhanced policy learning that effectively connects compact reasoning to action execution. Extensive experiments across diverse embodied manipulation and reasoning benchmarks demonstrate that Fast-ThinkAct achieves strong performance with up to 89.3\% reduced inference latency over state-of-the-art reasoning VLAs, while maintaining effective long-horizon planning, few-shot adaptation, and failure recovery.

Figures (10)

• Figure 1: Overview of Fast-ThinkAct. Previous reasoning VLAs generate lengthy reasoning traces ($\sim$250 tokens). Our approach learns compact continuous tokens (e.g., 6) (blue) and parallel spatial tokens (green) as internal reasoning. The bottom-right plot shows that we achieve $9.3\times$ faster inference than ThinkAct-7B huang2025thinkact, while delivering improved performance on the SimplerEnv-Google benchmark.
• Figure 2: Overview of Fast-ThinkAct. (a) Given observation $o_t$ and instruction $l$, the Textual Teacher VLM $\mathcal{F}_\theta^T$ generates explicit reasoning chains. The Latent Student VLM $\mathcal{F}_\theta$ distills these into compact latent tokens $\mathbf{z}$ guided by reward preferences. Verbalizer LLM $\mathcal{V}_\psi$ decodes latents to text for preference-based learning via $\mathcal{L}_{\text{verb}}$, while $\mathcal{L}_{\text{distill}}$ transfers visual planning capability from teacher, and spatial tokens enable parallel visual trajectory prediction via $\mathcal{L}_{\text{ans}}$, ensuring latents are verbalizable and grounded in visual planning. (b) Reasoning-Enhanced Policy Learning. The Action Model $\pi_\phi$ is trained with $\mathcal{L}_{\text{IL}}$ while freezing the latent student $\mathcal{F}_\theta$ and state encoder.
• Figure 3: Evaluation of robot manipulation and reasoning efficiency. (a)-(e) Success rates on LIBERO liu2023libero and SimplerEnv li24simpler benchmarks compared with state-of-the-art 7B reasoning VLAs. (f) Latency comparison across 3B and 7B reasoning VLAs. Our approach achieves up to 89.3% inference latency reduction while maintaining superior task success rates.
• Figure 4: Visualization of predicted visual trajectories and action execution results on long-horizon tasks. Examples from (a) SimplerEnv-Google, (b) LIBERO-Long, and (c) RoboTwin2.0-Hard with long (278) steps. Yellow traces indicate single-arm/left gripper trajectories; red traces indicate right gripper trajectories for bimanual tasks.
• Figure 5: Failure recovery capability on RoboFAC lu2025robofac. Left: Qualitative examples (from both simulation and real robot) of corrective guidance for manipulation errors. Right: Quantitative evaluation on simulation (RoboFAC-Sim) and real-robot (RoboFAC-Real) settings.