Reproduce “What Matters for World Modeling in Vision-Language-Action Models?” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qisen Ma, Yixiang Chen, Yuan Xu, Peiyan Li, Jiabing Yang, Kai Wang, Jianhua Yang, Jianlou Si and 5 more
The authors compare four ways of adding future-scene prediction to robot policies and test them across manipulation tasks. Under a shared Qwen3-VL-2B policy, WM-MoT reaches 76.2% episode-weighted binary success on LIBERO-Plus, versus 71.8% for the action-only baseline. With current DINOv3 features held constant, future supervision raises WM-MoT from 73.6% to 76.2%; results vary by representation and architecture.
Key insights
- Representation quality depends on the prediction architecture: no encoder or integration paradigm leads in every setting, though DINOv3 and SigLIP provide comparatively consistent gains.
- World modeling is especially useful for long-horizon and dynamic tasks; short tasks can see little benefit or slight regressions, and some designs are more sensitive to reduced training data.
- Separate predictive capacity can help with moving objects and pretrained VLAs, but auxiliary future-prediction objectives can also interfere with an existing policy, as seen with WM-Align on π₀.₅.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.