History Forcing: Zhejiang U & Microsoft Fix Long-Horizon Drift in Video World Models

ZhejiangUniversity · hf · 2026-11-04

• Zhejiang University, Microsoft Research, Monash University and University of Adelaide published History Forcing (Nov 2026), targeting long-horizon inconsistency in few-step autoregressive video world models.
• The method uses history context residuals to correct distillation targets, distilling history-dependent continuation behavior into the model.
• Results: consistent mountain textures and body proportions across 60-second sequences, and preservation of out-of-sight object changes when the camera revisits a scene.

Original post →

More from Research

• Classic reward-hacking cases, from CoastRunners to the asthma rule, turned into an eval checklist — ClickOk5811 · 2026-11-04
• "Machine Psychology for Language Models" Published in Science Advances — scychan_brains · 2026-11-04
• CMU Study: Harness Engineering Boosts Scientific Agents Without Retraining — niloofar_mire · 2026-11-04
• STOC Gossip: One Paper Claims P=NP, Another Claims P≠NP — Both Can't Be Right — thomasahle · 2026-11-04
• Stanford team forecasts GSM8K accuracy within 3 pts using 5600x less compute via hinted evals — stanfordnlp · 2026-11-04
• Model-weights self-improvement demos show promise but lack evidence to beat fine-tuning — ccerrato147 · 2026-11-04

Research channel →