WAM-Cache: Staleness-Bounded KV Reuse for Efficient World Action Models
Overview
WAM-Cache is a training-free framework that reduces video DiT prefill FLOPs by 32-42% while maintaining high action accuracy. It uses pretrained video Diffusion Transformers (DiT) for generalist robot manipulation by retaining key-value representations across chunks and recomputing sparse tokens. The method combines cross-attention with visual latent surprise and an age bound to suppress compounding error. The approach is based on arXiv:2610.11401v1. Latest: WAM-Cache reduces video DiT prefill FLOPs by 32-42% across benchmarks while maintaining high action accuracy. It uses pretrained DiT for generalist robot manipulation by retaining key-value representations and recomputing sparse tokens. The method combines cross-attention with visual latent surprise and an age bound to suppress compounding error. The approach is based on arXiv:2610.11401v1.
Generated from attributed reports · Updated 49 minutes ago
Event evidence and corrections
0 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Environment: simulation
“ing within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot.”
Exact source · revision 1Source owner not reported
Reported quantity · percentage points: 0.7 other · Basis not reported · Differing source assertions
“ing within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot.”
Exact source · revision 1Source owner not reported
Reported quantity · percentage points: 2.5 other · Basis not reported · Differing source assertions
“ing within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot.”
Exact source · revision 1Source owner not reported
Report timeline
Follow attributed reports and material updates.
- arXiv Robotics — research abstractsWAM-Cache: Staleness-Bounded KV Reuse for Efficient World Action Models
WAM-Cache is a training-free framework that retains layerwise key-value representations across chunks and recomputes only a sparse refresh set of tokens. It combines cross-attention from the action expert with visual latent surprise and a strict age bound to suppress compounding error. The method reduces video DiT prefill FLOPs by 32-42% across multiple benchmarks while maintaining high action accuracy.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).