How does a language model decide its next word? This interactive app lets you explore how language models generate text under different decoding strategies. You can:
Move through text generation one token at a time, comparing the choices allowed by greedy, top-k, and top-p decoding.
Watch how the next-token probability distribution shifts after each step, with colors highlighting the tokens considered by each strategy.
Explore the Beam Search tab to see a tree of candidate sequences develop step-by-step.