IJCAI 2026 专访：动作指挥计算，VLA 加速不降准 | GAIR Paper 125

TL;DR - The IJCAI 2026 paper introduces AC²-VLA, an action-context-aware framework that accelerates vision-language-action models by dynamically routing computation based on robot action state. On the SIMPLER benchmark, it reduced FLOPs to 29.4% of the CogACT baseline and achieved 1.79× faster inference without lowering task success.

• A unified learned router coordinates cognition-cache reuse, visual-token pruning, and Transformer-layer skipping according to action context rather than visual complexity alone.
• Self-distillation trains the sparse model to match the dense model’s actions and internal features, avoiding reliance on sparse task-success signals.
• AC²-VLA averaged 76.8% success on four Google Robot Visual Matching tasks, compared with 74.8% for dense CogACT.
• Cache reuse can improve temporal consistency by suppressing frame-to-frame visual noise, though routing, memory movement, and cache lookup overhead limit realized speedups relative to FLOP reductions.