New Method Boosts Efficiency for Vision-Language-Action AI Models.

Vision-Language-Action (VLA) models, which integrate visual, linguistic, and action data, often face high computational and memory costs due to processing extensive multimodal token sequences. While existing methods primarily focus on reducing the number of visual tokens, this can lead to fragility by discarding entire representations. Researchers propose Role-Conditioned Sub-Token Routing (RoleSub) as a complementary approach. RoleSub works by partitioning each token's value representation into groups in an orthogonal space and then using a lightweight router to decide which groups to retain. This routing decision is dynamically conditioned on the token itself, a learned latent role, and the language context. The technique can be applied to both visual and language representations, allowing for significant compression without removing additional tokens. Evaluations on the OpenVLA-OFT-7B model across various robotic tasks demonstrated that RoleSub consistently outperforms token-only compression methods. It achieved a reduction in total Key-Value (KV) memory to as little as 9.2-11.3% of the original, while still maintaining robust control performance, proving its effectiveness for aggressive VLA compression.