AIAny
Icon for item

The Past Frames the Future: Memory for Autoregressive Video Generation

Surveys memory mechanisms for autoregressive video generation, framing memory as persistent historical information that influences future generation. Organizes work by Forms, Functions, Operations, Learning, and Evaluation, and synthesizes challenges for long-horizon consistency and memory-aware learning.

Introduction

Most autoregressive video generators rely on bounded context windows, which causes essential historical details—object identities, spatial layouts, event states—to drop out long before they are no longer relevant. This paper argues that treating retention as an explicit memory problem (what to store, how to update it, when to read it) is the key to reliable long-horizon, autoregressive video generation. The survey reframes diverse techniques under a unified operational view and pinpoints where current methods fail under closed-loop rollouts.

Key Findings
  • Memory as a design lens: effective memory requires not just capacity but accuracy, accessibility, and causal influence; representations range from raw visual windows and KV caches to compressed latents, 3D priors, and retrieval banks, each trading fidelity for efficiency.
  • Lifecycle operations matter: writing, reading, updating, managing, and integrating memory are distinct problems that demand specialized mechanisms (salience scoring, routers, retrieval, state-conditioned updates) and matching training strategies to avoid drift.
  • Learning & evaluation gaps: self-rollout training, trustworthy state-updating supervision, and standardized evaluation tasks that reveal true long-term retention remain underdeveloped; many methods optimize short-horizon metrics that mask memory failure.
  • Practical trade-offs: approaches that extend context (large KV caches or rolling windows) scale poorly; structured or compressed memories reduce footprint but need robust update rules to prevent inconsistency.
Who it's for and trade-offs

Great fit if you research long-horizon video generation, world-models, or memory-augmented generative systems and need a structured taxonomy plus pointers to recent mechanisms and benchmarks. Look elsewhere if you only need short-video synthesis or fidelity-focused diffusion recipes—this survey emphasizes temporal persistence, evaluation design, and memory-specific training over implementation details or single-model engineering.

Information

  • Websitearxiv.org
  • OrganizationsHKUST, CityUHK, FDU, ZODA, CMU, NYU, HKUST(GZ), NUS, Georgia Tech, PKU, MBZUAI, NVIDIA, UCF, UNITN, NTU, UC Merced
  • AuthorsHarold Haodong Chen, Rongjin Guo, Disen Lan, Wen-Jie Shu, Hongfei Zhang, Hanzhe Hu, Shengtao Yao, Zixin Zhang, Guibin Zhang, Zhefan Rao
  • Published date2026/09/23

Categories

More Items

Introduces a rubric-based approach for video reward modeling that generates explicit, query-adaptive evaluation criteria before scoring to reduce scalar drift; proposes RGPO, a two-stage training (seed warm-up + joint optimization) that achieves strong pointwise and pairwise evaluation with high data efficiency.

Learns a camera-queryable implicit 3D-aware memory that compresses multi-view history into target-view tokens to enable long-horizon, camera-controllable video generation. Improves revisit consistency and camera-control accuracy and supports streaming exploration from a single image or text prompt.

Evaluates whether MiniMax-H3 can infer latent physical events from incomplete multimodal evidence and produce videos consistent with those inferences. Builds a 517-instance benchmark across four reasoning scenarios (multi-view, audio disambiguation, video continuation, audiovisual integration) and quantifies modality-specific failure modes.