Recurrent Autoregressive Diffusion: Global Memory Meets Local Attention

arXiv 2025 38 method, application

TLDR

Proposes RAD, a recurrent autoregressive diffusion framework with LSTM for long video generation, improving memory and consistency.

Reasoning

The paper introduces a novel integration of RNN (LSTM) into diffusion transformers for long video generation, addressing forgetting and spatiotemporal inconsistencies. Strengths include a consistent training-inference procedure and improved memory retention. Weaknesses are the limited evaluation on game datasets (Memory Maze, Minecraft) and lack of comparison to broader world model benchmarks.

Read-first score

Read-first score 38, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 37.

Recency 6%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 18%

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

Topical relevance 29%

52.9

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

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 199.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduces an LSTM-based recurrent neural network into the diffusion transformer framework for long-term memory retention in video generation.
Proposes the Recurrent Autoregressive Diffusion (RAD) framework with frame-wise autoregressive memory update and retrieval, eliminating training-inference gap and lack of overlap across windows.
Demonstrates that LSTM with attention matches state-of-the-art RNN blocks like TTT and Mamba2 in diffusion models.

Methodology

The method integrates an LSTM into a diffusion transformer to form the RAD framework, which performs frame-wise autoregressive memory updates and retrievals consistently during training and inference. Experiments are conducted on long video generation using the Memory Maze and Minecraft datasets.

Key Results

RAD achieves superior long video generation performance on Memory Maze and Minecraft, outperforming existing diffusion-RNN approaches and confirming the efficiency of LSTM for sequence modeling.

Resources

Paper