Vision-Language-Action (VLA) models have shown impressive zero-shot cross-task transfer, but a new arXiv paper argues that adaptation methods are often necessary to achieve consistent performance on novel tasks. The authors frame this as a practical gap between general transfer and reliable deployment.

The paper, titled 'Post-Training, Finetuning, and Steering: A Field Guide for Curating Training Data,' organizes adaptation into three approaches. It positions these as complementary ways to improve VLA performance on tasks where zero-shot inference falls short.

The field guide focuses on how to curate training data for each approach, giving practitioners a structured path from pretrained models to task-specific behavior. The abstract is truncated, but the core argument is clear: adaptation is often required, and data curation is central to doing it well.