NEO-VLA removes vision encoder, hits 90.9% frozen on LIBERO

A new robotics paper argues that the standard VLA recipe of pretraining a vision encoder and then bolting on an action head may be holding manipulation policies back. NEO-VLA, described in a preprint on Hugging Face, skips the separately pretrained visual encoder entirely: lightweight convolutional layers tokenize pixels straight into a 40-layer, 2B-parameter vision-language backbone that image and text tokens traverse jointly, with a flow-matching action expert reading cached keys and values from selected backbone depths.

Under matched training conditions, the authors report NEO-VLA reaching "90.9% average success" on LIBERO with the backbone frozen, "compared with 79.5% for Qwen3-VL and 84.5% for Gemma-3." Full fine-tuning brings the average to 95.3%. Layer-wise probing produces the result the paper leans on hardest: "action-critical signals often peak well before the final visual representation," and the authors note that full fine-tuning and added scale narrow but do not erase the encoder-free backbone's lead.

The paper also argues against a cheaper fix. Routing a Qwen3-VL action expert to intermediate vision-tower features, which adds no parameters, actually hurts: "Encoder access lowers frozen-backbone performance from 79.5% to 64.9%, while full fine-tuning brings the two interfaces to similar scores, 90.9% and 90.8%." On RoboTwin 2.0's randomized split NEO-VLA holds 69.40 against its own 75.78 clean score, and across three real-world tasks fine-tuned from 50 demonstrations each it scores 12/20, 17/20, and 13/20 over 20-trial evaluations.

The authors flag their own caveat on the cross-paper tables: "simulator versions, camera setups, training data, and evaluation protocols differ across prior work, so these results should be viewed as benchmark context rather than controlled comparisons." NEO-VLA lands as the 185th robotics story in our last 90 days of tracking.