Altered Thoughts, Altered Actions: Reasoning Chain as Control Surface for a Vision-Language-Action Policy

This research explores how reasoning chains in vision-language-action policies affect a robot's motor actions, demonstrating that they function as a control surface. By manipulating the reasoning chain and instruction inputs, the study shows that text edits can repair or corrupt behavior. Using LIBERO simulations and DeepThinkVLA, the authors measure both repair and corruption effects, finding that a corrected reasoning chain recovers 47.8% of lost success, though still falling short of clean runs by 38.0%.