arXiv:2606.23641v2 Announce Type: replace Abstract: Vision-language-action (VLA) models have the potential for open-world generalization by leveraging pretrained vision-language representations, yet downstream finetuning on limited robot data often degrades these representations. This leads to brittle action policies that ignore language instructions in favor of visual shortcuts, despite the base VLM having this knowledge, which is a failure mode we term instruction blindness. We hypothesize that standard finetuning with limited data applies gradients to a sparse set of points, which manifests as a sharp loss landscape with high-curvature minima. We analyze this effect, then show that regularization through explicit flatness-preserving optimization is a simple yet effective technique while finetuning on the exact same data. Specifically, we demonstrate that applying sharpness-aware minimization (SAM) during VLA finetuning significantly improves instruction following by over 55% across multiple simulation and real-world benchmarks without additional data, architectural modification, or pretraining. Explicitly preserving flatness better preserves pretrained knowledge compared to other regularization techniques that simply anchor model weights or constrain learning. We further analyze the effect of selective sharpness, quantify its effects, and show that such approaches are complementary to existing guidance techniques.