Tencent ARC's Future Chunk Prediction boosts physics scores in autoregressive video generation
yshan2u · x · 2026-10-22
A Nanjing University / Tencent ARC Lab / Zhejiang University paper tackles a hidden weakness of autoregressive video models: neighboring chunks are so similar that copying clean context lowers denoising loss without learning scene evolution. Their Future Chunk Prediction (FCP) forces models to predict multiple future chunks without seeing the in-between. On controlled Wan2.2-5B experiments, Physics-IQ Verified scores rose from 30.4→33.8 (image-to-video) and 40.6→44.4 (video-to-video). Code is open-sourced as TencentARC/Looking-Ahead.
Related event: FCP: Training Trick Makes Autoregressive Video Generation More Physical(4 posts)→
More from Research
- BatchBench leaderboard released: AI cell microscopy models degrade sharply on new data sources — gklambauer · 2026-10-22
- SERA: self-evolving VLM harnesses tackle long-horizon robotic manipulation — Tingwen Du · 2026-10-22
- SpatialScout benchmark tests VLM agents on active 3D mapping across 80 scenes — Zhaoxi Zhang · 2026-10-22
- New paper: adversarial examples aren't bugs, they're superposition, with four lines of evidence — livgorton · 2026-10-22
- Judea Pearl shares four buckets of new causal inference papers, 50 pages of titles — yudapearl · 2026-10-22
- Interpretability researchers clash: is it a V neuron, and what does SAE failure to disentangle mean? — livgorton · 2026-10-22