UAM: A Dual-Stream Perspective on Forgetting in VLA Training

A recent study published on arXiv (2605.15735) indicates that the typical fine-tuning process for vision-language-action (VLA) models derived from pretrained vision-language models (VLMs) leads to a gradual decline in multimodal abilities, referred to as the 'embodiment tax.' The researchers attribute this decline to a structural limitation: existing VLAs utilize a single encoder for both language-based semantics and visual features relevant to control, unlike natural vision, which differentiates recognition from visuomotor control. To remedy this issue, they introduce the Unified Action Model (UAM), which incorporates a parallel Dorsal Expert, mirroring the brain's dorsal pathway. This Dorsal Expert is initialized from a pretrained generative model and trained with a mid-level objective to alleviate the control-learning demands on the VLM. The paper does not disclose authors or affiliations.