Research map / Video / Action recognition
Olaf-World: Orienting Latent Actions for Video World Modeling
Yuxin Jiang, Yuchao Gu, Ivor Tsang, Mike Zheng Shou
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 |
|---|---|---|---|
| add Seq-Delta-REPA effect alignment to latent action learning | AdaWorld-style beta-VAE LAM · MIND 1st-P to 1st-P linear probe | Macro-F1: 0.6004 -> 0.8138 | Improved |
| add Seq-Delta-REPA effect alignment | AdaWorld-style beta-VAE LAM · MIND 3rd-P to 3rd-P linear probe | Macro-F1: 0.4827 -> 0.8256 | Improved |
| add Seq-Delta-REPA effect alignment | AdaWorld-style beta-VAE LAM · MIND cross-domain probe 1st-P to 3rd-P | Macro-F1: 0.4820 -> 0.6250 | Improved |
| add Seq-Delta-REPA effect alignment, adapt with 50 videos | Olaf-World vs AdaWorld · MIND 3RD-P | RPE rotation: 0.3353 -> 0.2082 | Improved |
| add Seq-Delta-REPA, 1 adaptation video | Olaf-World vs AdaWorld · MIND 1ST-P | RPE rotation: 0.6420 -> 0.4680 | Improved |
| add Seq-Delta-REPA effect alignment, video quality | Olaf-World vs AdaWorld · MIND 1ST-P, 0 videos | Image quality: 0.5600 -> 0.5400 | Hurt |
| latent-action pretraining vs direct ground-truth action conditioning | Olaf-World vs DirectAct · MIND 1ST-P, 1 video | RPE rotation: 1.2822 -> 0.4680 | Improved |
| add latent-action pretraining, unseen visual contexts | Olaf-World vs AdaWorld · OOD 50 initial frames | RPE rotation: 1.7063 -> 1.2221 | Improved |
| align temporal feature differences instead of static features (ablation) | Seq-Delta-REPA · MIND 1st-P to 3rd-P probe | Macro-F1: 0.5287 -> 0.6250 | Improved |
| use L2-normalized cosine alignment instead of MSE (ablation) | Seq-Delta-REPA · MIND 3rd-P to 3rd-P probe | Macro-F1: 0.7096 -> 0.8256 | Improved |
| use video reference encoder (V-JEPA2) instead of image encoder (DINOv3) (ablation) | Seq-Delta-REPA · MIND 1st-P to 1st-P probe | Macro-F1: 0.2334 -> 0.7044 | Improved |
| use DINOv3 image encoder as reference vs no alignment (ablation) | Seq-Delta-REPA · MIND 1st-P to 1st-P probe | Macro-F1: 0.1849 -> 0.2334 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AdaWorld — AdaWorld: Learning Adaptable World Models with Latent Actions (ICML 2025) · better on 12 of 12 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
- Generalizing Action-Conditioned Latent World Models with Video Model Rewards (NeurIPS 2026 · similarity 0.91)
- Latent Action Learning Requires Supervision in the Presence of Distractors (ICML 2025 · similarity 0.88)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.88)
- AdaWorld: Learning Adaptable World Models with Latent Actions (ICML 2025 · similarity 0.88)
- Learning Visual Feature-Based World Models via Residual Latent Action (NeurIPS 2026 · similarity 0.88)
- World Models as Group Actions (NeurIPS 2026 · similarity 0.88)
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.
