VisuaLLM watches a language model think, one token at a time. For every word it generates, you see the candidates it weighed, how confident it was, and which one it actually picked.