Most reinforcement learning agents only get a simple pass/fail reward, which hides how good or bad their attempts really were.
STEP3-VL-10B is a small (10 billion parameters) open multimodal model that sees images and reads text, yet scores like much larger models.