LoopVLA: Learning Sufficiency in Recurrent Refinement for Vision-Language-Action Models

Abstract: Current Vision-Language-Action (VLA) models typically treat the deepest representation of a vision-language backbone as universally optimal for action prediction. However, robotic manipulation is composed of many frequent closed-loop spatial adjustments, for which excessive abstraction may waste computation and weaken low-level geometric cues essential for precise control. Existing early-exit strategies attempt to reduce computation by stopping at predefined layers or applying heuristic rules such as action consistency, but they do not directly answer when a representation is actually sufficient for action. In this paper, we present LoopVLA, a recurrent VLA architecture that jointly learns representation refinement, action prediction, and sufficiency estimation. LoopVLA iteratively applies a shared Transformer block to refine multimodal tokens, and at each iteration produces both a candidate action and a sufficiency score that estimates whether further refinement is necessary. By sharing parameters across iterations, LoopVLA decouples refinement from absolute layer indices and grounds sufficiency estimation in the evolving representation itself. Since sufficiency has no direct supervision, we introduce a self-supervised distribution alignment objective, where intermediate confidence scores are trained to match the relative action quality across refinement steps, thereby linking sufficiency learning to policy optimization signals. Experiments on LIBERO, LIBERO-Plus, and VLA-Arena show that LoopVLA pushes the efficiency-performance frontier of VLA policies, reducing parameters by 45% and improving inference throughput by up to 1.7 times while matching or outperforming strong baselines in task success.

• The paper introduces LoopVLA, a recurrent refinement framework that learns representational sufficiency for efficient action prediction in VLA models.
• It employs a shallow Transformer loop with dual heads for dynamic halting based on task complexity, improving throughput and task success rates.
• Empirical evaluations show significant performance gains and reduced model size in challenging robotic manipulation benchmarks compared to fixed-depth baselines.

LoopVLA: Recurrent Refinement and Sufficiency Estimation in Vision-Language-Action Models

Motivation and Problem Setting

Vision-Language-Action (VLA) models have become central to general robotic manipulation, leveraging multimodal pretraining for robust perception and control. Existing VLA models commonly adopt a "late-output" paradigm, where only the final-layer representation from a vision-language backbone is used for subsequent action generation. However, this paradigm overlooks two fundamental limitations in control-oriented applications. First, representational depth is not uniformly optimal: low-level geometric cues critical for rapid closed-loop actions may be attenuated in deep, heavily abstracted representations. Second, computational cost is unnecessarily inflated when excessive processing is applied regardless of task complexity.

Heuristic early-exit and post-hoc intermediate readout strategies do not provide principled or fine-grained sufficiency evaluation for stopping criteria, either incurring the full computational burden of deep models or relying on weak heuristics such as action consistency. The core challenge is defining and learning "representational sufficiency"—the minimal level of refinement needed for reliable action prediction—directly grounded in policy optimization.

Architectural Contribution

LoopVLA addresses these concerns by recasting VLA policy learning as an iterative, recurrent refinement process. Central to this architecture is a shallow Transformer block applied in a loop, with dual output heads at each iteration: one for action prediction and one for sufficiency estimation.

Unlike fixed-depth stacking, parameter tying across iterations eliminates bias toward absolute layer indices and compels sufficiency determination from representational quality, not depth. The shared loop architecture enables dynamic computation allocation—allowing the model to spend more refinement steps on challenging inputs and terminate early for simple scenarios.

The sufficiency estimation head incorporates a cross-attention mechanism between sufficiency-specific tokens and action tokens, enriched by loop-index positional encodings, to output a halting probability at each step. This is normalized via a remaining mass allocation (RMA) mechanism, yielding a valid stochastic halting distribution over iterations.

Training Procedure

Since there is no direct ground-truth for sufficiency, the model is trained in two stages. Initially, all intermediate action predictions are jointly supervised with $L_1$ loss, and sufficiency predictions are softly regularized with entropy and diversity terms to encourage meaningful exploration of loop indices.

In the second stage, sufficiency heads are calibrated explicitly: the halting distribution is aligned to a softmax target over per-iteration action errors (acting as a proxy for action quality) via a KL divergence objective. This distribution alignment formulates sufficiency as a learned, policy-driven attribute, not a hand-tuned heuristic.

Empirical Results

LoopVLA demonstrates consistently strong efficiency-performance trade-offs in robotic manipulation benchmarks:

• On LIBERO, LoopVLA achieves an average task success rate of 96.0% (LoopOFT*) with a model 45% smaller than leading baselines and up to 1.7× greater inference throughput with significantly fewer parameters.
• On VLA-Arena, average performance exceeds strong baselines (e.g., 48.7% vs. 47.0% for Qwen3OFT), particularly excelling in long-horizon and safety-critical task categories.
• Zero-shot generalization on LIBERO-Plus validates the robustness of LoopVLA under substantial visual, linguistic, and embodiment perturbations: the model remains competitive (e.g., 65.8% vs larger baselines exceeding 65%) even as parameter count is substantially reduced.

Ablation studies establish the necessity of the structured sufficiency head: substituting it with a standard MLP leads to large performance drops, especially on long-horizon tasks. Additionally, inference performance exhibits a non-monotonic dependency on both loop depth and iteration—confirming the utility of adaptive, input-dependent halting.

The analysis of selected loop indices across tasks further indicates that simpler tasks elicit early halting, while complex manipulations drive deeper iterative refinement. This demonstrates effective task-dependent computational allocation.

Theoretical and Practical Implications

From a theoretical perspective, LoopVLA reframes the architecture of VLA policies to treat representational sufficiency as a first-class, learnable property, tightly coupled to policy loss landscapes. This abstraction is more aligned with biological insights into hierarchical control and bridges the gap between depth-agnostic perception models and the needs of fine-grained robotic action.

Practically, this work establishes adaptive VLA models as compelling routes for reducing inference latency and compute budget without sacrificing policy quality or robustness. It enables flexible deployment in resource-constrained or real-time robotic environments by providing actionable trade-offs between computational budget and control performance.

Limitations and Future Directions

While LoopVLA advances the efficiency and design philosophy of VLA systems, notable limitations remain. The current recurrent mechanism has not been systematically scaled to much larger model and data regimes; future work should investigate scaling laws for iterative refinement in VLA models. Additionally, the present benchmarks are predominantly in simulation; real-world evaluation on physically embodied robots is an essential next step. Improving robustness under dramatic distributional shifts and extending to reinforcement learning or closed-loop adaptation remain promising research avenues.

Conclusion

LoopVLA provides a principled, recurrent approach to VLA policy optimization, explicitly learning sufficiency criteria for iterative representation refinement. By unifying action prediction and sufficiency estimation in an efficient, task-adaptive framework, LoopVLA significantly improves upon fixed-depth and heuristic early-exit strategies. This architecture paves the way for scalable, content-aware, and compute-efficient VLA models in both research and applied robotic settings (2605.09948).