Event

NestedVLA Uses Hypernetworks to Generate Action Expert Parameters for VLA Skill Integration

1 reports · 1 sources

Guodong Du, Wanyu Lin, and others proposed the NestedVLA method, which uses a nested hypernetwork to generate task-adaptive action expert parameters for the target task and combines them with a frozen base VLA, replacing parameter aggregation with parameter generation to alleviate multi-task training interference. The paper was accepted by NeurIPS 2026. On the three reinforcement learning benchmarks MetaWorld, ManiSkill, and CALVIN, NestedVLA averaged 41 percentage points higher than the strongest model-merging baseline; on ManiSkill, its gap from per-task fine-tuning was within 1 percentage point, and on CALVIN it also notably narrowed the gap.