Deep pruning of vision-language-action (VLA) models involves two coupled decisions: which layers to remove and how to regain lost task performance. A failed trial can mean either the wrong layers were dropped or the recovery was insufficient. AutoPruner, a new framework from researchers on arXiv, treats this as a closed-loop experimentation problem rather than a one-shot optimization.
The system first performs offline layer interventions to measure how much each layer's removal changes the model's action outputs. An LLM agent uses that functional evidence, along with evaluation history and retained diagnoses, to propose deletion sets and recovery recipes within a search budget. Each proposal is executed and evaluated on the target tasks, and the results feed the next proposal.
Across OpenVLA-OFT, StarVLA, and FastWAM on four LIBERO suites, AutoPruner removed 16.7–87.5% of backbone layers while maintaining task success above 90%, with measured inference speedups up to 2.50×. The authors also find that functional probes expose action-sensitive layers that a representation-similarity view misses, and that adding recovery can turn previously failing pruning structures into viable ones.