Role-Conditioned Sub-Token Routing for Efficient Vision-Language-Action Policies
RoleSub introduces a novel sub-token compression technique for Vision-Language-Action (VLA) models that partitions value representations into groups and routes them based on token, role, and language context Unlike token pruning which discards entire tokens, RoleSub retains more tokens while reducing their value width, addressing fragility issues in aggressive compression Evaluated on OpenVLA-OFT-7B across LIBERO suites, outperforming token-only controls in 33 of 36 settings with largest gains u
Analysis
TL;DR
- RoleSub introduces a novel sub-token compression technique for Vision-Language-Action (VLA) models that partitions value representations into groups and routes them based on token, role, and language context
- Unlike token pruning which discards entire tokens, RoleSub retains more tokens while reducing their value width, addressing fragility issues in aggressive compression
- Evaluated on OpenVLA-OFT-7B across LIBERO suites, outperforming token-only controls in 33 of 36 settings with largest gains under aggressive compression
- Combined visual and language compression reduces total KV cache to 9.2-11.3% of original while maintaining strong robotic control performance
Why It Matters
This work addresses a critical bottleneck in VLA deployment—expensive inference from long multimodal token sequences—by introducing a complementary compression strategy that preserves more information than aggressive token pruning. For AI practitioners building robotic systems, this enables more efficient deployment of vision-language-action models without catastrophic performance loss.
Technical Details
- RoleSub partitions each retained value representation into groups in an orthogonal space after visual token reduction, using a lightweight router conditioned on token representation, learned latent role representation, and language context
- The mechanism applies to both visual and language values, enabling dual compression without additional token removal
- Evaluated on OpenVLA-OFT-7B architecture across four LIBERO benchmark suites for robotic manipulation tasks
- Achieves 9.2-11.3% KV retention while maintaining strong control performance, with consistent outperformance (33/36 settings) versus token-only compression baselines
Industry Insight
- Sub-token compression represents a promising direction for VLA efficiency that complements existing token pruning methods, particularly for resource-constrained robotic deployment scenarios
- The role-conditioned routing approach could inspire similar adaptive compression strategies for other multimodal architectures beyond robotics
- Aggressive KV reduction to ~10% of original size while maintaining performance suggests practical pathways for real-time VLA inference on edge hardware
Disclaimer: The above content is generated by AI and is for reference only.