DeCoPrune: training-free pruning cuts video diffusion memory tokens 85%, 4.14x faster
XingangP · x · 2026-10-12
DeCoPrune is a training-free token pruning method for autoregressive video diffusion. Key insight: denoising difficulty reveals what is worth remembering—keep tokens hard to predict from context and prune redundant ones, with no extra forward passes. On LingBot World v2 it cuts historical KV tokens by 85.43%, speeds continuation generation 4.14x, and keeps long-range visual recall near FullKV. The authors also release CMBench, a benchmark for long-range visual memory, along with paper, code and dataset.
Related event: DeCoPrune: Training-Free Pruning Speeds Video Generation 4x(2 posts)→
More from Research
- TestSuiteBench: coding agents kill only ~55% of bugs with generated test suites — lbeurerkellner · 2026-10-13
- Berkeley study: 10 minutes of ChatGPT use reduces persistence once removed — RachelVT42 · 2026-10-13
- Radiolab's 'Math Vs Machine': a mathematician grapples with AI's rise — stevenstrogatz · 2026-10-13
- Brain microprotein atlas with 1,000+ microproteins lands on Hugging Face — huggingface · 2026-10-13
- Looped Transformers Stuck at 8% on OOD Multi-hop; DiscoLoop Fix Hits ~100% — jasondeanlee · 2026-10-13
- Microsoft's StudentSim simulates individual students to train AI tutors before real trials — Microsoft Research · 2026-10-13