Jul 2026
RL2-VLA: Adaptive RL Latent Compositional Steering with Test-Time Scaling for Vision-Language-Action Models
This work introduces an adaptive inference-time steering framework that leverages Reinforcement Learning on VLA Latents, and discovers that inference-time steering follows fundamentally different scaling laws under success and failure states, revealing that action diversity is most beneficial when the base VLA is likel...
Derek Ming Siang Tan, S. Shailesh, Srikrishna Iyer et al.
· arXiv.org · 0 citations