Research map / Video / Video generation
An Empirical Study of Autoregressive Pre-training from Videos
Jathushan Rajasegaran, Ilija Radosavovic, Rahul Ravishankar, Yossi Gandelsman, Christoph Feichtenhofer, Jitendra Malik
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use attention pooling instead of average pooling for probing (ablation) | Toto-large · ImageNet linear probe | top-1 accuracy (%): 53.2 -> 61.1 | Improved |
| pretrain at 128x128 then finetune at higher resolution with RoPE (ablation) | Toto-large · ImageNet linear probe | top-1 accuracy (%): 61.2 -> 63.2 | Improved |
| raise RoPE base to 50,000 during finetuning (ablation) | Toto-large · ImageNet linear probe | top-1 accuracy (%): 63.2 -> 64.4 | Improved |
| use LLaMA architecture instead of GPT2 (ablation) | Toto-large · ImageNet linear probe | top-1 accuracy (%): 48.5 -> 53.2 | Improved |
| use transformer LLaMA instead of Mamba (ablation) | Toto-large · ImageNet linear probe | top-1 accuracy (%): 40.7 -> 53.2 | Improved |
| swap tokenizer between dVAE, VQGAN and patch tokens (ablation) | Toto-large · ImageNet linear probe | top-1 accuracy (%): 61.3, 61.1, 61.2, 60.6 | No clear effect |
| use dVAE tokens as target with patch embedding input (ablation) | Toto-large · ImageNet linear probe | top-1 accuracy (%): 60.6 -> 58.5 | Hurt |
| use full attention instead of causal attention at fine-tuning (ablation) | Toto-base · ImageNet fine-tuning | top-1 accuracy (%): 82.2 -> 82.6 | Improved |
| add MLP layers to the cross-attention probe head (ablation) | Toto-base · Kinetics-400 | top-1 accuracy (%): 59.3 -> 61.2 | Improved |
| scale model from base to 1b | Toto · Kinetics-400 | top-1 accuracy (%): 59.3 -> 74.4 | Improved |
| scale model from base to 1b | Toto · ImageNet-1k | top-1 accuracy (%): 64.7 -> 75.3 | Improved |
| Toto-base versus MVP encoder for robot cube picking | Toto-base · Franka cube picking real world | success rate (%): 75% -> 63% | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- VideoMAE — VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (NeurIPS 2022) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Video In-context Learning: Autoregressive Transformers are Zero-Shot Video Imitators (ICLR 2025 · similarity 0.85)
- Autoregressive Sequential Pretraining for Visual Tracking (CVPR 2025 · similarity 0.85)
- Video-Mirai: Autoregressive Video Diffusion Models Need Foresight (NeurIPS 2026 · similarity 0.84)
- Taming Teacher Forcing for Masked Autoregressive Video Generation (CVPR 2025 · similarity 0.84)
- Autoregressive Video Generation without Vector Quantization (ICLR 2025 · similarity 0.83)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
