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

• URL: http://arxiv.org/abs/2605.15735v2
• Date: Mon, 18 May 2026 07:08:58 GMT
• Title: UAM: A Dual-Stream Perspective on Forgetting in VLA Training
• Abstract summary: Vision--language--action (VLA) models are typically built by fine-tuning a pretrained vision--language model (VLM) on action data.<n>We show that this standard recipe systematically erodes the VLM's multimodal competence.<n>Inspired by the two-stream organization of biological vision, we trace this degradation to a structural bottleneck.
• Score: 20.50709311546184
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Vision--language--action (VLA) models are typically built by fine-tuning a pretrained vision--language model (VLM) on action data. However, we show that this standard recipe systematically erodes the VLM's multimodal competence, a side effect we call the embodiment tax. But do VLAs have to forget? Inspired by the two-stream organization of biological vision, we trace this degradation to a structural bottleneck: current VLAs ask a single encoder to support both language-grounded semantics and control-relevant visual features, whereas biological vision separates recognition and visuomotor control into distinct pathways. Building on this view, we propose the Unified Action Model (UAM), which adds a parallel Dorsal Expert, an analog of the brain's dorsal pathway. To make the Dorsal Expert an effective second pathway and reduce the control-learning burden on the VLM, we initialize it from a pretrained generative model and train it with a mid-level reasoning objective that predicts visual dynamics. This design allows us to train the whole VLA end-to-end on action data alone: with no parameter freezing, no gradient stopping, and no auxiliary VL co-training, UAM retains over $95\%$ of the underlying VLM's multimodal capability and at the same time achieves the highest average success rate among baselines on a variety of manipulation tasks that probe out-of-distribution generalization, including unseen objects, novel object--target compositions, and instruction variation. Together, these results suggest that semantic preservation in VLAs can emerge from architectural separation itself, rather than being enforced by frozen weights or auxiliary data replay, and that this preserved semantic capability can naturally transfer from VLMs to semantic generalization in actions.

Related papers

• Two Bridges, One Pathway: From VLMs to Generalizable VLAs with Embodied Trajectory-Coupled Data [93.7685703383343]
  Vision-language models (VLMs) are powerful general-purpose reasoners, yet converting them into robot control policies is surprisingly difficult.<n>We argue that this gap can be bridged gradually with the right intermediate data.<n>We show that mixing task-relevant out-of-distribution ETC data with a small amount of action data enables the model to generalize to novel visual-language conditions.
  arXiv  Detail & Related papers  (2026-06-07T08:57:51Z)
• QuoVLA: Quotient Space for Vision-Language-Action Models [51.02329790939691]
  Vision-Language-Action (VLA) models commonly adapt pretrained VLM latents to robot control by mapping visual observations and language instructions to continuous actions.<n>Our textitQuotient Theory for VLA shows that pretrained VLM latents are not action-insufficient but action-sufficient.<n>We propose QuoVLA, a quotient-space framework for VLA that compresses pretrained VLM latents into action-sufficient representations.
  arXiv  Detail & Related papers  (2026-05-24T06:28:53Z)
• DIAL: Decoupling Intent and Action via Latent World Modeling for End-to-End VLA [72.9197085473598]
  DIAL is a framework bridging high-level decision making and low-level motor execution through a differentiable latent intent bottleneck.<n>A VLM-based System-2 performs latent world modeling by latent visual foresight within the VLM's native feature space.<n>A lightweight System-1 policy then decodes this predicted intent together with the current observation into precise robot actions.
  arXiv  Detail & Related papers  (2026-03-31T15:02:27Z)
• Look Before Acting: Enhancing Vision Foundation Representations for Vision-Language-Action Models [66.96421290733126]
  Vision-Language-Action (VLA) models have emerged as a promising paradigm for robotic manipulation.<n>We propose textbfDeepVision-VLA, built on a textbfVision-Language Mixture-of-Transformers (VL-MoT) framework.<n>DeepVision-VLA outperforms prior state-of-the-art methods by 9.0% and 7.5% on simulated and real-world tasks, respectively.
  arXiv  Detail & Related papers  (2026-03-16T17:59:54Z)
• ActionCodec: What Makes for Good Action Tokenizers [106.78093973045526]
  Vision-Language-Action (VLA) models have demonstrated superior instruction-following and training efficiency.<n>Central to this paradigm is action tokenization, yet its design has primarily focused on reconstruction fidelity.<n>We introduce textbfActionCodec, a high-performance action tokenizer that significantly enhances both training efficiency and VLA performance.
  arXiv  Detail & Related papers  (2026-02-17T07:07:15Z)
• VISTA: Enhancing Visual Conditioning via Track-Following Preference Optimization in Vision-Language-Action Models [26.542479606920423]
  Vision-Language-Action (VLA) models have demonstrated strong performance across a wide range of robotic manipulation tasks.<n>Despite the success, extending large pretrained VLA models to the action space can induce vision-action misalignment.<n>We propose a training framework that explicitly strengthens visual conditioning in VLA models.
  arXiv  Detail & Related papers  (2026-02-04T20:59:29Z)
• Vision-Language-Action Models for Autonomous Driving: Past, Present, and Future [125.92052530850425]
  Vision-Language-Action (VLA) frameworks integrate perception with language-grounded decision making.<n>VLA frameworks offer a pathway toward more interpretable, generalizable, and human-aligned driving policies.<n>This work aims to establish a coherent foundation for advancing human-compatible autonomous driving systems.
  arXiv  Detail & Related papers  (2025-12-18T16:57:44Z)
• DualVLA: Building a Generalizable Embodied Agent via Partial Decoupling of Reasoning and Action [62.70893433854428]
  We propose DualVLA, which enhances action performance through carefully designed post-training while still preserving reasoning capability.<n>Experiments show that DualVLA achieves an average success rate of 61.0 in SimplerEnv and an average score of 65.4 across eight competitive multimodal benchmarks.
  arXiv  Detail & Related papers  (2025-11-27T06:03:53Z)
• iFlyBot-VLA Technical Report [25.330744626382977]
  We introduce iFlyBot-VLA, a large-scale Vision-Language-Action (VLA) model trained under a novel framework.<n>The main contributions are listed as follows: (1) a latent action model thoroughly trained on large-scale human and robotic manipulation videos; (2) a dual-level action representation framework that jointly supervises both the Vision-Language Model (VLM) and the action expert during training; and (3) a mixed training strategy that combines robot trajectory data with general QA and spatial QA datasets.
  arXiv  Detail & Related papers  (2025-11-01T06:24:56Z)
• Don't Blind Your VLA: Aligning Visual Representations for OOD Generalization [42.41263928527529]
  Vision-Language-Action (VLA) models can endow agents with transferable world knowledge and vision-language grounding.<n>Yet when these VLMs are adapted to the action modality, it remains unclear to what extent their original visual representations and knowledge are preserved.<n>We conduct a systematic study of representation retention during VLA fine-tuning, showing that naive action fine-tuning leads to degradation of visual representations.
  arXiv  Detail & Related papers  (2025-10-29T15:20:10Z)
• Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better [58.559985503802054]
  Vision-language-action (VLA) models combine end-to-end learning with transfer of semantic knowledge from web-scale vision-language model (VLM) training.<n>The most powerful VLMs have tens or hundreds of billions of parameters, presenting an obstacle to real-time inference.<n>Recent VLA models have used specialized modules for efficient continuous control, such as action experts or continuous output heads.<n>We show that naively including such experts significantly harms both training speed and knowledge transfer.
  arXiv  Detail & Related papers  (2025-05-29T17:40:09Z)
• CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models [89.44024245194315]
  We introduce a method that incorporates explicit visual chain-of-thought (CoT) reasoning into vision-language-action models (VLAs)<n>We introduce CoT-VLA, a state-of-the-art 7B VLA that can understand and generate visual and action tokens.<n>Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by 17% in real-world manipulation tasks and 6% in simulation benchmarks.
  arXiv  Detail & Related papers  (2025-03-27T22:23:04Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.