Our paper about how language models secretly reason through meaningless filler tokens like “.....” was accepted at NeurIPS 2026!