generate text one token at a time, which is slow. uses a small drafter to guess several tokens at once, then the main model checks them in parallel. But the drafter often hesitates on hard tokens.
The researchers introduce Parallel Reconsideration: the drafter identifies uncertain guesses and generates extra candidates for those tokens, all in parallel. This adds almost no extra time. The system is trained to correct its own mistakes.
Across models and tasks, the method improved the average number of accepted tokens per round by 5.3% to 10.2%, and sped up overall generation. This matters because it makes language models faster without more powerful hardware.