AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/11/2026, 04:18:24 UTC.

Briefings

A new vision-language-action framework called NEO-VLA removes fixed visual encoders, allowing policies to access evolving visual computations directly. Controlled experiments show native architectures outperform encoder-based models, and full fine-tuning only partially closes the gap. Layer-wise probing indicates action-critical signals peak earlier in the hierarchy than final visual representations. The study suggests embodied control benefits from accessing visual representations across the computational hierarchy rather than relying on a single fixed encoder output.

Key points

• NEO-VLA is a structurally encoder-free vision-language-action framework where images are tokenized by lightweight patch convolutional layers and evolve jointly with language throughout the backbone.
• Layer-wise probing reveals that action-critical visual signals peak well before the final visual representation, suggesting access to evolving visual computation is more effective than fixed encoders.

Context

Vision-language-action models typically compress perception first, but manipulation often relies on visual signals vulnerable to this abstraction. This research introduces NEO-VLA to address the question of whether action policies should inherit a fixed visual endpoint or directly access evolving visual computation.

Limitations

Results are based on controlled comparisons and full fine-tuning experiments. Additional scale alone cannot recover the performance gap between native and encoder-based models. The study covers diverse manipulation settings but does not address distribution shifts beyond the tested settings.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 09/11/2026, 04:11:11 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiflnenxbfcxofgmnovdfu5bqnivnifr6mv6em3fotfmadob75gecq