In plain words
Language models usually write one token at a time. ProVe is a decoding method that lets a specially trained model suggest a block of future tokens in one pass, then uses another pass to check how much of that block to keep.
In tests with a 32B Qwen model, its self-checking version kept the model’s ordinary greedy output exactly and averaged higher throughput than ordinary decoding across six instruction tests. The result is for the tested setup, not every model or task.
The research
Language models that generate one token per pass require repeated serial work. Standard speculative decoding uses a smaller model to draft serially, so the unresolved question is whether a large model trained to propose a block at once can make generation faster, and how much checking those proposals costs.
ProVe lets a block-trained language model propose future tokens, then checks the block to preserve greedy output or trade some accuracy for speed.
To speed up language-model generation, can a large model draft several future tokens at once instead of relying on a smaller model that drafts serially? ProVe trains a model to propose a block, then uses a second pass to check the actual proposed tokens and decide how much to keep.
Across six instruction tests, self-checking reproduced the proposer’s ordinary greedy output and averaged 2.54 times its relative throughput.
With a smaller verifier, the authors measured faster averages alongside small mean accuracy drops at tested settings, supporting a speed-and-accuracy tradeoff.
The experiments use greedy decoding with Qwen models at a 32B proposer scale; instruction and reasoning models are tested on different sets of tasks. The results do not establish performance for other model families or scales, sampling, or more optimized serving systems.