The Past Frames the Future: Memory for Autoregressive Video Generation
Autoregressive video models lose useful history once it falls outside their bounded context, and this paper frames that as a memory problem.
The authors review memory mechanisms meant to preserve information such as entity identity, dynamic state, and causal changes across long video rollouts. They define memory as historical information that can still affect generation after the original evidence is no longer locally available. The survey organizes prior work across forms, functions, operations, learning, and evaluation. It closes with open problems around resource-aware architectures, reliable state updates, self-rollout learning, and standardized tests. HF Daily Papers' note
The authors review memory mechanisms meant to preserve information such as entity identity, dynamic state, and causal changes across long video rollouts. They define memory as historical information that can still affect generation after the original evidence is no longer locally available. The survey organizes prior work across forms, functions, operations, learning, and evaluation. It closes with open problems around resource-aware architectures, reliable state updates, self-rollout learning, and standardized tests. HF Daily Papers' note
score 4