Research-quality prompts
Central claim: the forcing-method lineage progressively aligns training conditions, causal architecture, and inference-time rollouts.
Hidden assumption: a student's own generated history remains informative enough to serve as a useful training distribution.
Failure boundary: very few steps, horizons beyond training, prompt changes, and constrained KV memory amplify appearance and motion drift.
Connection: the topic links long-video generation, real-time control, causal memory, and event-level evaluation.
Concrete experiment: keep backbone and data fixed while separately changing teacher history, student rollout, and causal-teacher initialization.
Discussion seed: after a student leaves the teacher distribution, what signal can still provide a correct and stable causal update?
Open questions
- What is the minimal recipe shared across forcing methods?
- When does self-rollout supervision become error reinforcement?
- Is a separate causal teacher necessary?
- Is there a structural conflict between few-step quality and long-horizon stability?
- How should immediate quality, motion, and delayed event consistency be evaluated separately?