Addressing depth generalization limits in latent recurrent VLA models

Develop architectural innovations or training protocols for Recurrent-Depth VLA (RD-VLA) and related latent iterative reasoning-based visuomotor policies that prevent state saturation and performance degradation when recurrence depth exceeds the empirically optimal range, thereby enabling reliable scaling of test-time compute in robotics.

Background

The paper introduces Recurrent-Depth VLA (RD-VLA), a vision–language–action architecture that performs iterative refinement in a latent space using a weight-tied recurrent transformer core. Experiments show that performance improves with increasing recurrence depth up to an optimal range (typically 8–12 iterations), after which further unrolling can cause state saturation or performance degradation rather than continued improvement.

In the Discussion, the authors explicitly identify a limitation concerning depth generalization: although recurrence enables adaptive test-time compute, pushing beyond the optimal depth can harm performance. They state that resolving this limitation to achieve robust scaling of latent reasoning is an open challenge, suggesting potential solutions through architectural changes or specific training protocols.

References

A key limitation observed in our experiments is the boundary of depth generalization. While performance scales predictably with the number of recurrent steps up to some optimal number of iterations, extending recurrence beyond this number of iterations may lead to state saturation or performance degradation rather than continued refinement. Addressing this problem—perhaps through architectural innovations or specific training protocols remains an open challenge for scaling latent reasoning in robotics.

— Recurrent-Depth VLA: Implicit Test-Time Compute Scaling of Vision-Language-Action Models via Latent Iterative Reasoning  (2602.07845 - Tur et al., 8 Feb 2026) in Discussion and Limitations