PathComp — Persistent VLA Anchors
Preserving acquisition-time interfaces during continual VLA fine-tuning.
Path Compatibility (PathComp) preserves acquisition-time input–response relations during continual fine-tuning of vision–language–action policies. Under review at ICLR 2027.
- Record modal tokens together with the fused latents and action distributions they produced when a task was acquired.
- Train current fusion and action modules to preserve those persistent responses from cached inputs.
- Use a separate modal anchor to constrain encoder drift on stored observations.
The study evaluates held-out action error under its continual-learning protocol. It does not equate offline action error with rollout success, and does not claim matched storage or total training compute.