Gabriel Herbert Sarch

ViGoRL is a vision-language model trained with reinforcement learning to ground each reasoning step in image coordinates, improving performance on spatial and web-based reasoning tasks through better attention and visual verification.