QuickMerge++: Fast Token Merging with Autoregressive Prior

Dong Liu, Yanxuan Yu

As generative models scale to larger inputs across language, vision, and video domains, the cost of token-level computation has become a key bottleneck. While prior work suggests t…