X-Cache: Cross-Chunk Block Caching for Few-Step Autoregressive World Models Inference

arXiv 2026 57 method

TLDR

X-Cache accelerates autoregressive world model inference by caching across generation chunks, enabling real-time simulation for autonomous driving.

Reasoning

The paper presents a novel caching method that reduces inference cost for interactive world models, with a dual-metric gating mechanism and KV cache management. Its strength lies in addressing a practical bottleneck, but the abstract is cut off, limiting full evaluation of results and limitations.

Read-first score

Read-first score 57, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 53.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 29%

75.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation,metric

Citation impact 18%

63.3

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.63305222

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 416.

Keyword Scores

world model

interactive world model

world simulator

video world model

generative world model

model-based reinforcement learning world model

world dynamics prediction

Deep Analysis

Innovations

Cross-chunk block caching that caches across consecutive generation chunks rather than across denoising steps
Dual-metric gating mechanism using a structure- and action-aware block-input fingerprint to decide recompute or reuse
Identification of KV update chunks with unconditional full computation to prevent error propagation in the autoregressive KV cache

Methodology

X-Cache is a training-free acceleration method for few-step autoregressive world models. It maintains per-block residual caches that persist across generation chunks and uses a dual-metric gating mechanism with a structure- and action-aware block-input fingerprint to decide whether to recompute or reuse cached residuals. To prevent approximation errors from contaminating the autoregressive KV cache, it identifies KV update chunks and unconditionally forces full computation on them.

Key Results

On the X-world production multi-camera action-conditioned driving world model, X-Cache achieves 71% block skip rate and 2.6x wall-clock speedup while maintaining minimum degradation.

Limitations

The method is designed specifically for few-step autoregressive world models with multi-block causal DiT and rolling KV cache, limiting generalizability to other video generation paradigms.
The abstract does not quantify the degradation or provide failure cases, leaving uncertainty about the trade-off between speedup and quality.

Resources

Paper