Speculative Forcing

The Speculative Forcing framework treats the intermediate clear estimates in the denoising process as drafts and predicts which tokens can be safely accepted early.

In Speculative Forcing, once a token is accepted, it is frozen as final output and no longer participates in subsequent denoising computation, but its cached key-value is still used as context by the remaining tokens.

In the LongLive benchmark test, Speculative Forcing matches the quality of a four-step generator, requires on average only about 1.7 denoising steps per token, runs more than 1.5 times faster, and reaches generation frame rates of 41.7 and 37.8 fps.

In tests on Lingbot-World 2.0, Speculative Forcing requires on average about 2.5 denoising steps per token, runs 1.35 times faster, and reaches 24.5 fps.

The training-free verifier in Speculative Forcing judges whether a token can be accepted early based on its similarity to the clear content and its stability across steps.

The learned verifier in Speculative Forcing reads the intermediate features of the frozen generator and estimates the remaining gap of each token from its full-step output.

The core design principle of Speculative Forcing is that not all tokens need the same degree of computational investment, that is, letting each token compute as needed.

Speculative Forcing adds adaptive pruning on top of few-step generators, changing the number of steps that must be completed from a fixed value to one determined dynamically per token.

Speculative Forcing's early acceptance decisions do not disrupt the generation conditions of subsequent tokens, because the key-values of accepted tokens remain in the context.

In world generation scenarios, the average number of steps required per token (2.5 steps) is higher than in short video scenarios (1.7 steps), reflecting the higher refinement demands of dynamic content.