HERO: Hierarchical Extrapolation and Refresh for Efficient World Models
TLDR
Generation-driven world models create immersive virtual environments but suffer slow inference due to the iterative nature of diffusion models.
Reasoning
Fallback reasoning generated from available title and abstract metadata: Generation-driven world models create immersive virtual environments but suffer slow inference due to the iterative nature of diffusion models. While recent advances have improved diffusion model efficiency, directly applying these techniques to world models introduces limitations such as...
Read-first score
Read-first score 35.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals.
Field roles
Rank sensitivity
Stability: volatile; rank range: 199.
Deep Analysis
Innovations
- Identification of feature coupling phenomenon in world models: shallow layers exhibit high temporal variability while deeper layers yield stable feature representations.
- Patch-wise refresh mechanism with frequency-aware tracking for efficient token recomputation in shallow layers, avoiding extra metric computation and compatible with FlashAttention.
- Linear extrapolation scheme for deeper layers that directly estimates intermediate features, bypassing computations in attention modules and feed-forward networks.
Methodology
HERO is a training-free hierarchical acceleration framework for world models. In shallow layers, it uses a patch-wise refresh mechanism with frequency-aware tracking to selectively recompute tokens, avoiding extra metric computation and remaining compatible with FlashAttention. In deeper layers, it applies a linear extrapolation scheme to directly estimate intermediate features, completely bypassing attention and feed-forward network computations.
Key Results
HERO achieves a 1.73× speedup with minimal quality degradation, significantly outperforming existing diffusion acceleration methods.