A vision-language-action edit kept 52% of control performance

The fast edit is not consistently local

A vision-language-action policy can be adapted after training by subtracting the weight change associated with one learned skill. The attraction is speed: given precomputed task-specific experts, the new preprint says the subtraction needs neither data nor gradients at edit time. Its August 5 audit asks the harder question: after removing one robot behavior, do the behaviors meant to remain still work through a full closed-loop rollout?^[1]

Across all 10 tasks in the LIBERO-Goal simulation suite, the answer split three ways. Five edits suppressed their intended target while retaining enough average control performance to meet the authors' separation rule; three target skills resisted suppression; and two edits drove both target and every evaluated control to zero success. Each of the 100 cells in the full target-by-control matrix used 20 fresh rollouts, so the headline is not one selected demonstration.^[1]

The useful denominator is the controls that survive

The paper's strongest result is less reassuring than a clean target removal. For the five target-suppressing edits, held-out initial states left mean baseline-normalized control retention at 52 percent, with individual target edits spanning 78 percent down to 4 percent. In other words, the method erased the intended task in half of the 10-skill audit, but the unrelated skills did not survive uniformly. That is a more operational measure of locality than observing a changed action token or a single forward pass.^[1]

The scope is deliberately narrow. LIBERO is a procedural robot-manipulation benchmark with task suites, fixed initial states and human-teleoperated demonstrations; it is not a customer site or a physical safety test. The authors list simulation-only evaluation and one full training seed among the main limits. Their reported OpenVLA-7B and flow-matching checks broaden the architecture coverage, but remain simulation results rather than evidence that a deployed robot can safely forget one behavior.^[1,2,3]

The reproduction path is visible; the hardware test is not

The underlying evaluation stack is at least inspectable. LIBERO publishes its code, data-download path and evaluation script, while the official MergeVLA repository supplies implementation and publicly linked expert models used to build the study's main policy family. That makes the reported matrix a testable research claim, not an opaque benchmark graphic. It does not make the new audit independently reproduced, and it does not establish the same behavior under sensor noise, contact variation, resets or physical safeguards.^[2,3]

The decision change is simple: teams considering post-training removal of an unsafe, deprecated or restricted robot skill should treat target suppression and retained-control performance as separate release gates. The next measurable checkpoint is a public reproduction across additional seeds and physical robots, with every retained behavior rolled out after the edit rather than inferred from static model metrics.^[1,2,3]