Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models

TL;DR

The paper tackles the overhead and representational mismatch of explicit chain-of-thought in Vision-Language-Action models by introducing LaRA-VLA, which internalizes reasoning in continuous latent space and eliminates CoT generation at inference. A three-stage curriculum progressively replaces explicit multi-modal CoT with latent representations, anchored by a latent-augmented backbone and a diffusion-based action generator. Through automated data pipelines and structured CoT datasets, LaRA-VLA achieves state-of-the-art performance on simulated benchmarks and long-horizon real-robot tasks, while delivering up to a 90% reduction in inference latency. These findings suggest latent reasoning as a scalable, efficient paradigm for real-time embodied control that better aligns with continuous perception and action than discrete CoT approaches.

Abstract

Vision-Language-Action (VLA) models benefit from chain-of-thought (CoT) reasoning, but existing approaches incur high inference overhead and rely on discrete reasoning representations that mismatch continuous perception and control. We propose Latent Reasoning VLA (\textbf{LaRA-VLA}), a unified VLA framework that internalizes multi-modal CoT reasoning into continuous latent representations for embodied action. LaRA-VLA performs unified reasoning and prediction in latent space, eliminating explicit CoT generation at inference time and enabling efficient, action-oriented control. To realize latent embodied reasoning, we introduce a curriculum-based training paradigm that progressively transitions from explicit textual and visual CoT supervision to latent reasoning, and finally adapts latent reasoning dynamics to condition action generation. We construct two structured CoT datasets and evaluate LaRA-VLA on both simulation benchmarks and long-horizon real-robot manipulation tasks. Experimental results show that LaRA-VLA consistently outperforms state-of-the-art VLA methods while reducing inference latency by up to 90\% compared to explicit CoT-based approaches, demonstrating latent reasoning as an effective and efficient paradigm for real-time embodied control. Project Page: \href{https://loveju1y.github.io/Latent-Reasoning-VLA/}{LaRA-VLA Website}.

Figures (11)

• Figure 1: Comparison of CoT formulations in VLA models. (a) Textual CoT-based VLA explicitly generates discrete textual reasoning tokens, which are subsequently decoded into actions via downstream modules such as autoregressive (AR) policies or action experts (AE). (b) Most visual CoT-based VLA represents reasoning through discrete visual goal tokens, obtained via detokenization or alignment, before action generation. (c) Our model internalizes both textual and visual reasoning into continuous latent representations. Visual goal latents are aligned with perceptual features and serve a dual role: they encode future-oriented task information and provide implicit supervision for textual CoT latents.
• Figure 2: Overview of LaRA-VLA. Training proceeds in three stages: (i) explicit CoT fine-tuning with aligned visual prediction latents and inverse-dynamics supervision for actions; (ii) a curriculum-based transition from explicit CoT to compact text latents, gradually reducing the number of text tokens while increasing reliance on latent reasoning, where the latent representations are also implicitly supervised by visual and action signals; and (iii) adaptation of latent-conditioned VLM features to an action expert for efficient action generation without explicit CoT at inference time.
• Figure 3: Attention mechanism used in LaRA-VLA.
• Figure 4: Real-world setup of four long-horizon tasks.
• Figure 5: Real-world results.