Title: Closed-Loop Neural Activation Control in Vision-Language-Action Models

Original: arXiv:2606.00269v1 Announce Type: new

Abstract: While Vision-Language-Action (VLA) models allow for test-time steering through interventions on semantically significant internal directions, current approaches rely on a static steering coefficient, effectively functioning in an open-loop manner. This rigidity is ill-suited for embodied control scenarios, where task states and conceptual errors fluctuate over time, frequently leading to overcorrection, oscillatory behavior, and diminished task success—particularly affecting temporal attributes like speed and smoothness. To address this, we introduce CTRL-STEER, a closed-loop framework that substitutes static intervention strength with adaptive, time-varying control signals. Our core concept involves decoupling representation from regulation: instead of presuming that temporal concepts are governed by specific neurons, we direct interventions along motion-aligned residual directions while employing a feedback controller to dynamically adjust the intervention magnitude. We implement this framework using both PID and reinforcement learning-based controllers. Evaluations using a fine-tuned OpenVLA policy across four LIBERO task suites demonstrate that CTRL-STEER offers superior concept regulation stability and an improved trade-off between steering and task success compared to baselines with fixed coefficients, all without requiring modifications or retraining of the underlying model.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC