The paper introduces History Forcing, a unified on-policy distillation framework designed to improve long-horizon consistency in few-step autoregressive video diffusion models by exploiting richer historical supervision. The method computes a residual between teacher predictions conditioned on informative historical context and those without, then adds this weighted residual to the score or velocity target used for student distillation. Historical context is sourced from earlier generated chunks exhibiting less accumulated drift or a fixed-view teacher rollout that maintains scene visibility while the student camera looks away. This corrected target aligns the student generator with long-range historical constraints, mitigating inconsistency when subjects or scenes reappear after going out of sight. Evaluated on a 1.3B parameter model, the approach achieves state-of-the-art results, recording a VBench quality score of 85.07 for 15-second generation and a VBench-Long aggregate score of 79.82 for 60-second rollouts. On out-of-sight world modeling benchmarks, Evolution Quality on PlayWorld improves from 2.35 to 2.70, while Object Consistency on MemoBench rises from 58.43 to 62.18, demonstrating more plausible scene evolution and consistent state recovery upon camera return.