NEO-VLA Encoder-Free Robot Policy Beats Standard VLAs

A new robotics paper challenges the dominant recipe for building vision-language-action models, the architectures that let robots understand instructions and manipulate objects. NEO-VLA, officially released on November 5, 2026, strips out the separately pretrained vision encoder that most VLAs rely on and instead tokenizes pixels through lightweight convolutional layers directly into a shared 40-layer, 2B-parameter backbone where image and text tokens co-evolve. Under controlled comparisons on the LIBERO benchmark, the encoder-free approach consistently beats standard encoder-based baselines, and layer-wise probing suggests action-critical visual signals peak well before the final representation, raising questions about whether bolting action heads onto frozen vision encoders is holding manipulation policies back.

Architecture: No Separate Vision Encoder

Standard vision-language-action models typically freeze a pretrained vision encoder like CLIP or SigLIP and train an action prediction head on top. NEO-VLA takes a different path: lightweight patch convolutions tokenize raw pixels into the same 40-layer transformer backbone that processes language tokens, and a flow-matching action expert pulls instruction-conditioned features from multiple depths rather than only the final layer. The approach eliminates the boundary between vision pretraining and action learning, letting visual representations adapt to manipulation tasks from the start.

Benchmark Performance on LIBERO

Under matched training conditions on LIBERO, a standardized suite of simulated manipulation tasks, NEO-VLA reached 90.9% average success with the backbone frozen, compared with 79.5% for Qwen3-VL-2B and 84.5% for Gemma-3-4B. Full fine-tuning pushed NEO-VLA to 95.3%. The authors report that neither full fine-tuning nor added scale fully closes the gap for encoder-based baselines, suggesting the architectural difference is not easily overcome by parameter count or training budget alone.

Layer-Wise Probing: Action Signals Peak Early

The paper's central empirical result comes from layer-wise probing: training lightweight action heads on intermediate backbone representations reveals that action-critical visual features often peak well before the final layer. This finding contradicts the standard practice of reading only the last encoder layer. The authors tested routing a Qwen3-VL action expert to intermediate vision-tower features as a parameter-free fix, but this actually hurt frozen-backbone performance, dropping from 79.5% to 64.9%. Full fine-tuning brought both interfaces to similar scores, 90.9% and 90.8%, but the encoder-free architecture held a small edge even after full parameter updates.

Real-World and Generalization Results

On RoboTwin 2.0's randomized split, NEO-VLA scored 69.40 against its own 75.78 clean score, and across three real-world tasks fine-tuned from 50 demonstrations each it achieved 12/20, 17/20, and 13/20 over 20-trial evaluations. The authors caution that cross-paper comparisons are complicated by differences in simulator versions, camera setups, training data, and evaluation protocols, so benchmark context tables should not be read as controlled head-to-head trials. The preprint is available on Hugging Face, and the model lands as the 185th robotics story tracked in the last 90 days.

Release Date

NEO-VLA was officially released on November 5, 2026, with the preprint and model weights published on Hugging Face.

What This Means

NEO-VLA's encoder-free design suggests that the standard VLA recipe of freezing a pretrained vision encoder and training an action head may leave performance on the table. If action-critical visual signals do peak before the final encoder layer, then architectures that let vision and language co-evolve through a shared backbone could become the new baseline for embodied AI. The paper's results are strong enough to warrant replication and scaling experiments, and if the findings hold, we may see a wave of encoder-free manipulation policies in the next generation of robot learning systems.