← Back to Computer Vision cs.CV
Smart agents that know when to think hard versus react fast
Wencan Jiang, Jiangning Zhang, Jianbiao Mei, Jinzhuo Liu, Yu Yang, Xiaobin Hu, Zhucun Xue, Yong Liu, Dacheng Tao
May 18, 2026
Game-playing agents need to balance expensive reasoning with quick reactions across thousands of steps. SPIKE splits control into two: a Strategic Controller that plans and analyzes failures, and a Reactive Controller that executes fast under strict token limits. An Event Trigger watches for failures, progress stalls, and repeated actions to decide when reasoning is worth the cost. On StarDojo, this cuts token consumption by 55% while improving success rates by 38.5% relative to baselines.
Read the original paper →