Research map / Video / Video generation
Co-Evolving Latent Action World Models
Yucen Wang, Fengming Zhang, De-Chuan Zhan, Li Zhao, Kaixin Wang, Jiang Bian
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 warm-up phase (frozen world model supplies gradients to IDM) before joint training (ablation) | CoLA-World joint IDM + OpenSora training · latent action codebook metrics | codebook utilization/entropy | Improved |
| joint training without warm-up (random or pretrained IDM init) (ablation) | IDM + pretrained OpenSora world model · codebook metrics | codebook utilization | Hurt |
| longer warm-up before joint training (ablation) | CoLA-World · codebook metrics | joint training stability | Improved |
| joint training instead of two-stage with same 60K budget | OpenSora world model · OXE video prediction | FVD: 291.30 -> 278.90 | Improved |
| joint training instead of two-stage on real-action adaptation | OpenSora world model · LIBERO real action | FVD: 115.45 -> 93.68 | Improved |
| joint training instead of two-stage on visual planning | OpenSora world model · VP2 RoboDesk (5 tasks) | average success rate: 7.73% -> 21.20% | Improved |
| keep LAM frozen after warm-up instead of co-evolving with world model (ablation) | CoLA-World · world model video prediction | video generation quality | Hurt |
| continue E2E training with unfrozen world model instead of pure warm-up (frozen world model) (ablation) | CoLA-World LAM · LIBERO linear probing | probing loss | Improved |
| joint training instead of two-stage for LAM quality | linear probing of latent actions · six robotics datasets | L1 probing loss: 0.1614 -> 0.1623 on LIBERO | No clear effect |
| scale world model size (6 to 28 DiT blocks) (ablation) | CoLA-World · LIBERO linear probing | probing loss | Improved |
| increase batch size from 64 to 128 (ablation) | CoLA-World · video prediction and LIBERO probing | FVD: 261.78 -> 233.06 | Improved |
| swap OpenSora backbone for Wan2.1 (ablation) | CoLA-World · OXE, EgoCentric, AgiBot, LIBERO | FVD: 169.82 -> 155.58 on OXE; 161.91 -> 168.96 on LIBERO | Mixed |
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 1 of 1 shared result
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
- AdaWorld: Learning Adaptable World Models with Latent Actions (ICML 2025 · similarity 0.91)
- Generalizing Action-Conditioned Latent World Models with Video Model Rewards (NeurIPS 2026 · similarity 0.90)
- Pre-Trained Video Generative Models as World Simulators (AAAI 2026 · similarity 0.89)
- Learning Visual Feature-Based World Models via Residual Latent Action (NeurIPS 2026 · similarity 0.89)
- World Models as Group Actions (NeurIPS 2026 · similarity 0.88)
- Zero-shot World Models via Search in Memory (NeurIPS 2025 · 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.
