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

Most vision-language-action (VLA) models see the world through a separately pretrained vision encoder. The action policy only receives that encoder's final output, after local geometry, edges and depth cues have already been abstracted away. NEO-VLA drops the encoder. Pixels go through a lightweight convolutional patch embedder straight into a native vision-language backbone, and a flow-matching action expert reads from whichever backbone layers we choose.

Because the readout depth becomes a design choice, we can ask where in the backbone the information needed for control actually lives. Three findings come out of this study:

1. Action-relevant information sits in the shallow layers of the VLM. Reading early layers beats reading late layers at matched expert size, and intervening on layers 0–10 changes the predicted action far more than intervening on later layers.
2. Attaching Qwen's vision encoder to the action expert does not help. The gain comes from native vision-language pretraining. Routing the expert to Qwen3-VL's intermediate vision-tower features lowers frozen-backbone LIBERO success from 79.5% to 64.9%. The frozen native backbone reaches 90.9%.
3. Understanding data matters in post-training. On RoboTwin 2.0 Clean2Rand, co-training with NEO's understanding corpus gives 69.4%, a caption-only mix gives 47.2%, and action data alone gives 10.7%.

VideoNEO-VLA in two minutes

Robot clips are sped up 2–4×; the speed is marked in each shot.

Finding 1Action-relevant information sits in the shallow layers of the VLM

Each layer of the action expert reads the cached keys and values of one backbone layer. We keep the expert fixed at 28 layers and change only which backbone layers it reads: skip-28 spreads its reads across all 40 layers, end-28 reads layers 12–39, and first-28 reads layers 0–27.

The cleanest comparison is skip-28 vs. end-28: both keep all 40 backbone layers and the same 28-layer expert, and differ only in whether early layers are read. first-28 and first-14 also shorten the backbone, so they change depth and compute as well as readout position.

With a frozen backbone, dropping access to the first 12 layers costs 4.2 points (90.9 → 86.7). Full fine-tuning narrows the gap but does not close it (95.3 vs. 92.5).

Where the expert is sensitive

To see which attachment points matter, we intervene on one at a time. Replacing the keys and values at one layer with those from a different observation and task raises translation error most at layers 0–10. Noise of the same magnitude stays within 0.004 cm of the unchanged error, so the effect comes from the content being replaced, not the size of the disturbance. Masking the wrist image hurts most around layers 6–10; masking the instruction hurts most at layer 0, with a smaller rise near layer 17.

Control draws on early, spatially detailed computation inside the backbone, and a policy that only sees the last layers leaves that on the table.

Finding 2Attaching Qwen's vision encoder to the expert does not help; the gain comes from native pretraining

We put three backbones under the same action expert and robot-training corpus: Qwen3-VL-2B, Gemma-3-4B and NEO-2B. With the backbone frozen, only the expert learns, so the policy must work with the pretrained representations as they are. NEO-2B reaches 90.9% here, 11.4 points above Qwen3-VL and on par with both encoder-based models after full fine-tuning.

^a Updates the vision encoder and action expert with other layers frozen; uses a replan interval of 4 and horizon of 8 instead of 12 and 20.

Giving the expert direct access to Qwen's vision tower

A natural objection is that NEO-VLA wins only because its expert can reach early visual features, which an encoder-based VLA hides behind the encoder output. We tested this directly: in Qwen3-VL we route 13 expert layers to intermediate vision-tower features instead. This adds no parameters.

Encoder access makes the frozen Qwen3-VL policy worse (79.5 → 64.9), and after full fine-tuning the two interfaces are equal (90.8 vs. 90.9). Early visual features are only useful when they are learned jointly with language and presented in the space the rest of the model uses. NEO's encoder-free pretraining produces such features; an encoder bolted on in front of a language model does not.

Caveat: the backbones differ in pretraining data and capacity, so this comparison does not isolate the effect of removing the encoder alone. It does show that the native backbone is a strong starting point for action learning.

Exposing more of a conventional encoder is not enough. The benefit of NEO-VLA comes from native, encoder-free vision-language pretraining.

Finding 3Understanding data matters in post-training

VLA post-training is usually done either on action data alone or with some vision-language data mixed in. We compared three recipes on RoboTwin 2.0, trained on clean scenes and evaluated on randomized ones (Clean2Rand): co-training with NEO-2B's own multimodal understanding corpus, co-training with ShareGPT4V-COCO captions, and action data only.

Training on action data alone collapses under randomization. Simple captions recover part of the gap, and the backbone's own understanding corpus recovers most of it. In our recipe, understanding examples are packed into every other optimizer step with an auxiliary next-token loss weighted by 0.1.

Keeping the backbone's understanding ability alive during action training is what lets the policy hold up when scenes change, and the closer the co-training data is to the backbone's pretraining corpus, the better.

Real robotThree tasks, 50 demonstrations each

We built a small real-world benchmark to test precise placement, multi-step execution and object reorientation. Each model is fine-tuned on 50 episodes per task and evaluated over 20 trials: 10 in-domain and 10 out-of-domain. All methods use the same real data and their own recommended post-training settings.

Twenty trials per task is a small sample; differences of one or two successes are within noise.

BenchmarksSimulation results

Simulator versions, camera inputs, training data and evaluation protocols differ across papers, so the tables below give context rather than controlled comparisons. Numbers for other methods are as reported in prior work. Open marks methods with both code and weights public.

LIBERO

NEO-VLA: 2,000 episodes per suite at step 27,500. LIBERO is close to saturated, so we report it mainly to show in-distribution performance is retained.

LIBERO-Plus

LIBERO-Plus evaluates seven kinds of distribution shift separately. NEO-VLA is strongest when the way the scene is observed changes (camera, noise), and weaker when the robot or the layout changes.

The total is task-weighted rather than the mean of the seven axes. Other results follow Qwen-RobotManip.

RoboTwin 2.0

Under the official setting (2.5K clean training episodes), the main difference shows up when the policy is evaluated on randomized scenes.

LIBERO-Pro

LIBERO-Pro changes object appearance, object positions, instruction wording and task goals. NEO-VLA handles reworded instructions well but, like the other policies here, mostly fails when positions or goals change. Robust perception alone does not yet give flexible task execution.

TrainingData and recipe

Backbone. NEO-2B is a 40-layer decoder-only native vision-language model, trained in three stages (image-text pretraining, mid-training on captioning, conversation, OCR and detection data, then supervised instruction tuning). It uses native rotary position embeddings over time, height and width, causal attention for text and bidirectional attention within each image.

Robot pretraining corpus. Nine public dataset families (AgiBot-World, Bimanual YAM, RoboCOIN, Open X-Embodiment, Galaxea Open-World, RoboMIND, RH20T, DROID and the simulated InternData-A1), 2.23M episodes and about 9,900 hours. Unreliable joint readings, constant channels, action outliers and corrupt episodes are masked or dropped, and camera calibration is checked against projected arm motion.

Recipe. Stage 1 trains the action expert on a frozen backbone (113,400 steps, batch 8,192). Stage 2 unfreezes everything for 75,000 steps and co-trains with NEO-2B's understanding data. Stage 3 continues full fine-tuning on in-house action data for another 75,000 steps.

CitationBibTeX

@article{fan2026neovla,
  title   = {From Pixels to Actions: Towards Native Vision-Language-Action Policies},
  author  = {Fan, Weichen and Diao, Haiwen and Wu, Penghao and Hu, Siyuan and
             Huang, Ziqi and Liu, Tianqi and Sun, Mingze and Wang, Ziyi and
             Cai, Zhongang and Deng, Hanming and Wang, Quan and Yang, Lei and
             Lu, Lewei and Hong, Fangzhou and Chen, Zhaoxi and Lin, Dahua and
             Liu, Ziwei},
  year    = {2026}
}

NEO-VLA builds on the NEO native vision-language backbone: Diao et al., From Pixels to Words: Towards Native Vision-Language Primitives at Scale, ICLR 2026.