# From Pixels to Actions - Towards Native Vision-Language-Action Policies

**Authors:** Weichen Fan, Haiwen Diao, Penghao Wu, Siyuan Hu, Ziqi Huang, Tianqi Liu, Mingze Sun, Ziyi Wang, Zhongang Cai, Hanming Deng, Quan Wang, Lei Yang, Lewei Lu, Fangzhou Hong, Zhaoxi Chen, Dahua Lin, Ziwei Liu  
**Lab:** MMLab@NTU · **Published:** 4 Nov 2026 · **Upvotes:** 56 · **Level:** Advanced  
**Links:** [arXiv](https://arxiv.org/abs/2611.06918) · [PDF](https://arxiv.org/pdf/2611.06918) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.06918.html)

## Curator's notes

**TL;DR:** NEO-VLA lets action policies use visual features at all network layers, outperforming fixed-encoder models.

- **What's new:** Introduces an encoder-free architecture where vision and language interact throughout the network, unlike prior fixed-visual-encoder pipelines.
- **Key result:** Native architectures consistently outperform encoder-based counterparts across manipulation tasks and distribution shifts.
- **Why it matters:** Builders get more reliable robot control by accessing visual cues early in the model, not just at the final encoder output.

## Abstract

Vision-language-action models are typically built on a simple premise: perception is first compressed by visual encoders, and action reasoning begins only afterwards. Yet manipulation often hinges on precisely the visual signals most vulnerable to this abstraction, e.g., local geometry, boundaries, depth, and subtle state changes. This raises a fundamental question: should action policies inherit a fixed visual endpoint, or directly access evolving visual computation? We introduce NEO-VLA, a structurally encoder-free vision-language-action framework. Images are tokenized by lightweight patch convolutional layers and evolve jointly with language throughout the backbone, allowing a flow-matching action expert to draw from instruction-conditioned visual representations at different depths. This simple change reveals a consistent pattern. Native architectures outperform encoder-based counterparts under controlled comparisons; full fine-tuning reduces but does not erase the gap, and additional scale alone cannot recover it. More strikingly, layer-wise probing shows that the action-critical signals often peak well before the final visual representation. Across diverse manipulation settings and distribution shifts, NEO-VLA consequently achieves stronger and more robust control. Together, these results suggest that embodied control benefits from access to visual representations across the computational hierarchy, rather than a fixed encoder output.

## What you could build

1. Create a simulated robot arm that follows language commands using NEO-VLA for real-time grasp adjustments.
2. Develop a vision-language game agent that manipulates objects based on textual hints, leveraging layer-wise visual cues.
3. Implement a home‑assistant robot that adapts to new objects by fine‑tuning NEO-VLA without a separate visual encoder.

## Quiz yourself

What architectural change does NEO-VLA make compared to traditional vision-language-action models?

**Answer:** It removes the fixed visual encoder and lets visual tokens be processed jointly with language throughout the network, allowing access to intermediate visual representations.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.