DeCoPrune: Efficient KV-Cache Pruning for...

DeCoPrune. Efficient KV-Cache Pruning for Autoregressive Video Diffusion via Denoising Consistency. By Zeqi Xiao and five others, from Nanyang Technological University, Tsinghua University, and one other institution. Published on the fifth of October. Its code repository is worked out: about one hundred and ten source files, with a readme.

Autoregressive video diffusion generates video chunk by chunk, which is what makes streaming generation and interactive control possible. The cost is a KV cache that keeps growing with the generated history. Existing ways to compress that cache fall into two camps. Some discard history using fixed windows. Others select tokens through local attention and similarity signals. Neither directly measures the thing that matters: whether the current chunk contributes information beyond the context already retained. A token that merely repeats what the cache already knows is safe to drop. A token carrying something new is not, and the existing signals do not tell the two apart.

DeCoPrune is a training-free method that treats cache compression as a problem of denoising consistency. Its starting point is an empirical observation. How hard a token is to denoise turns out to be a useful proxy for how valuable it is to keep in the long run. Tokens whose prediction moves a lot between an intermediate step and the final result tend to carry visual evidence that is less predictable from the retained context. In other words, the tokens the model finds hard to settle are the ones holding something the cache does not already have.

For each token in the current chunk, DeCoPrune measures that difficulty as the discrepancy between the token's intermediate clean prediction and its final denoised value. High-discrepancy tokens are retained in the long-term cache. Low-discrepancy tokens are pruned. The signal comes from the denoising process the model already runs, so nothing extra has to be trained.

Testing whether information is really retained needed a new benchmark, so the authors built CMBench. It comprises fifty-eight context episodes, each about a minute long and either generated or real-world, and one hundred and sixteen continuation tasks of two kinds, Reappear and Revisit. Each task requires recalling a specific object or scene observed earlier.

Experiments use LingBot World v2. DeCoPrune preserves long-range recall close to that of the full, unpruned KV cache while pruning more than eighty-five percent of historical KV tokens. Continuation generation becomes more than four times faster. At comparable budgets, it substantially outperforms the compression baselines it was evaluated against.

The broader claim is that denoising consistency can act as a model-intrinsic signal: a way for the generator itself to say which parts of its past deserve to be kept, while the cost of autoregressive inference comes down. The project homepage, the code, and the CMBench dataset are all public.