Does a robotic arm make mistakes easily at step 10? A 2B model resolves this with dynamic attention adjustment | IJCAI 2026
The team from East China Normal University and Shanghai Jiao Tong University proposed S²-VLA, which introduces belief states and an adaptive dynamic gating mechanism. With only 2B parameters and 7GB of video memory, it achieves a long-term control success rate of over 96.4%, outperforming most 7B and even 8.5B models. This approach addresses the problem of cumulative error amplification in traditional VLA models due to static attention allocation during long-term tasks. By dynamically adjusting the weight ratios of visual, intention, and action self-attention channels, the model can flexibly allocate computing resources according to task phases (such as positioning, grasping, or switching). Experiments show that this method can spontaneously perceive progress and deviations without manual labeling of phase labels, effectively suppressing error propagation and improving execution stability. Additionally, S²-VLA has an inference throughput of 80.8Hz, making it suitable for deployment on edge devices.