ProAR: Learning Prospective Reasoning with Autoregressive Video Models

ProAR solves reasoning tasks by autoregressively generating a sequence of visual states. It equips autoregressive video models with two complementary components:

• Outcome Guidance predicts a future goal frame to guide current-chunk generation.
• Transition Guidance aligns the current representation with the clean next-chunk representation during training.

ProAR is built upon Wan2.2-TI2V-5B.

• Code: github.com/luka-group/ProAR
• Evaluation data (for VBVR-10tasks): LinghuiShen/ProAR-test
• Paper: Coming soon
• Project page: Coming soon

Model Zoo

Dataset	Model	Checkpoint	Training step
VBVR-10Tasks	AR	vbvr-10tasks/ar/model.pt	10,000
VBVR-10Tasks	ProAR	vbvr-10tasks/proar/model.pt	10,000
VideoRLVR-3Tasks	AR	videorlvr-3tasks/ar/model.pt	5,000
VideoRLVR-3Tasks	ProAR	videorlvr-3tasks/proar/model.pt	5,000
WorldArena	AR	worldarena/ar/model.pt	10,000
WorldArena	ProAR	worldarena/proar/model.pt	10,000

Download

Download all released checkpoints:

hf download LinghuiShen/ProAR \
  --local-dir checkpoints/ProAR