One-sentence Summary

Researchers from NVIDIA and Stanford introduce RD-VLA, a vision-language-action model that dynamically scales computation via latent-space iterative refinement rather than token-based reasoning, enabling 80× faster inference and 90%+ success on complex robotic tasks while maintaining constant memory usage.

Key Contributions

• RD-VLA introduces a recurrent, weight-tied action head that enables adaptive test-time compute via latent iterative refinement, eliminating fixed-depth constraints and supporting arbitrary inference depth with constant memory usage.
• The model is trained with truncated backpropagation through time and dynamically stops inference based on latent convergence, allowing it to allocate more compute to complex tasks while rapidly saturating on simpler ones—boosting success rates from 0% to over 90% on challenging manipulation tasks.
• Evaluated on LIBERO and CALVIN benchmarks, RD-VLA achieves state-of-the-art performance with 93.0% success on LIBERO and 45.3% task-5 success on CALVIN, while delivering up to 80× faster inference than prior reasoning-based VLA models.

Introduction

The authors leverage a recurrent, weight-tied architecture to enable adaptive test-time compute in Vision-Language-Action (VLA) models, addressing a key limitation of prior systems that expend fixed computational resources regardless of task complexity. Existing reasoning-based VLAs rely on explicit token generation—like Chain-of-Thought—which scales memory linearly and forces reasoning through lossy, discretized output spaces, making them inefficient for continuous robotic control. RD-VLA instead performs iterative refinement entirely within a fixed-dimensional latent space, using truncated backpropagation through time for training and an adaptive stopping criterion at inference to dynamically allocate compute. This design enables up to 80x faster inference than token-based reasoning models while maintaining constant memory usage, and achieves state-of-the-art performance on benchmarks like LIBERO and CALVIN, including real-world transfer to tasks like towel folding and bread toasting.

Method

The authors leverage a novel architecture called Recurrent-Depth Vision-Language Action (RD-VLA) to decouple computational depth from the fixed structural constraints of pretrained vision-language backbones. Rather than relying on fixed-depth MLP heads or output-space iterative methods such as diffusion, RD-VLA shifts the computational burden into a weight-tied recurrent transformer core that operates entirely within a continuous latent manifold. This enables dynamic test-time compute scaling by unrolling the recurrent block to an arbitrary depth $r$r, allowing the model to allocate more computation to complex tasks and less to simpler ones, as illustrated in the performance curve on LIBERO-10.

The RD-VLA action head is designed to be backbone-agnostic and is instantiated here using a Qwen2.5-0.5B-based VLM, augmented with 64 learned latent tokens that attend to the multi-modal context during the LLM’s forward pass. After VLM execution, hidden states are partitioned into task/vision representations $h_{vis} \in \mathbb{R}^{512 \times D}$hvis​∈R512×D and latent-specific representations $h_{lat} \in \mathbb{R}^{64 \times D}$hlat​∈R64×D. These are concatenated with proprioception $p$p to form a static conditioning manifold $[h_{vis+lat}^{(24)}; p]$[hvis+lat(24)​;p], which grounds the recurrent reasoning process.

The architecture follows a functional triplet: Prelude, Recurrent Core, and Coda. The Prelude $P_{\phi}$Pϕ​ consumes $K=8$K=8 learned queries, which first self-attend bidirectionally and then cross-attend to the VLM’s middle-layer features $h_{vis+lat}^{(12)}$hvis+lat(12)​ to produce a grounded latent foundation:

In parallel, a latent scratchpad $S_0$S0​ is initialized from a high-entropy truncated normal distribution:

This noisy initialization ensures the model learns a stable refinement operator rather than overfitting to a fixed starting point.

The Recurrent Core $R_{\theta}$Rθ​ performs iterative refinement by maintaining representational stability through persistent Input Injection. At each iteration $k$k, the current scratchpad state $S_{k-1}$Sk−1​ is concatenated with the static foundation $S_{pre}$Spre​, mapped back to the manifold dimension via a learned adapter, and normalized:

The scratchpad is then updated via the weight-tied transformer block:

Here, $R_{\theta}$Rθ​ performs bidirectional self-attention across the $K$K queries and gated cross-attention using keys/values derived from the concatenated conditioning manifold. This ensures the model remains grounded in the physical observation throughout unrolling.

Once the recurrence reaches depth $r$r, the converged scratchpad $S_r$Sr​ is processed by the Coda $C_{\psi}$Cψ​, which performs final decoding by attending to self and high-level VLM features. The output is projected to the robot’s action space:

Training employs randomized recurrence: the number of iterations $N$N is sampled from a heavy-tailed log-normal Poisson distribution with $\mu_{rec} = 32$μrec​=32. Truncated BPTT is used, propagating gradients only through the final $d=8$d=8 iterations, forcing the model to learn iterative refinement from any noisy initialization into a stable manifold.

At inference, adaptive computation is implemented by monitoring convergence via the L2 distance between consecutive predicted actions:

where $\delta = 1e^{-3}$δ=1e−3. This allows the model to self-regulate compute: terminating early for simple tasks and allocating more iterations for complex ones.

Adaptive execution further couples reasoning depth with action horizon. For high-uncertainty states ($k^* > \tau$k∗>τ), the execution horizon is truncated to $H_{short}$Hshort​; otherwise, it remains $H_{long}$Hlong​. Alternatively, a linear decay schedule reduces horizon inversely with iteration count:

This ensures the agent replans more frequently under high computational demand, prioritizing safety in complex scenarios.

Experiment

• Recurrent computation significantly boosts performance on manipulation tasks, with gains plateauing after 8–12 iterations, indicating diminishing returns beyond that point.
• Task complexity varies widely, requiring different numbers of reasoning steps; adaptive computation dynamically allocates compute based on task difficulty without predefined rules.
• Adaptive strategies (especially Binary Adaptation) match fixed-depth performance while cutting inference cost by up to 34%, confirming that condition-dependent compute allocation is more effective than uniform budgets.
• Latent reasoning outperforms both end-to-end and token-level reasoning methods, achieving state-of-the-art results on LIBERO and CALVIN benchmarks with a smaller model size (0.5B parameters).
• Real-world deployment on a bimanual robot shows strong robustness across household tasks, with fixed-depth variants excelling in most scenarios and adaptive variants remaining competitive while enabling efficiency.
• The approach demonstrates viability for physical systems and opens pathways for uncertainty-aware execution, though depth generalization and saturation remain key limitations for future work.

The authors evaluate their latent reasoning approach against end-to-end and token-level reasoning methods, showing that RD-VLA achieves state-of-the-art performance on the LIBERO benchmark with significantly fewer parameters. Results indicate that both fixed and adaptive variants of RD-VLA outperform prior methods across task categories, with the adaptive version maintaining strong performance while reducing average inference cost. The findings support that iterative latent-space reasoning is more parameter-efficient and effective for robotic manipulation than token-based or direct action prediction approaches.

The authors evaluate adaptive computation strategies for their recurrent reasoning model, showing that dynamic allocation of inference steps based on task complexity achieves performance comparable to fixed-depth models while reducing average compute cost. Results indicate that different task categories naturally require varying numbers of iterations, and adaptive methods like Binary Adaptation strike the best balance between efficiency and success rate. The model’s ability to self-calibrate computation based on state uncertainty demonstrates practical viability for real-world deployment.

The authors evaluate their RD-VLA model on the CALVIN ABC→D benchmark, showing it completes longer task sequences than prior methods, achieving the highest average chain length of 3.39. Despite using only 0.5B parameters—significantly fewer than most baselines—the model demonstrates strong sequential planning capability through latent iterative reasoning. Results confirm that recurrent depth enhances performance on long-horizon manipulation tasks without requiring large-scale architectures.