Reproduce “Agentic World 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
Zhuoyang Liu, Dantong Niu, Botong Zhang, Boning Shao, Jing Wang, Ruijie Zheng, Yuvan Sharma, Runyu Lu and 10 more
AW-0 adds a skill-program stream to a model that predicts future video and actions for language-directed manipulation. The program can guide imagined states and action generation, but is usually a plan conditioning an action expert rather than executable code. Across eight bimanual tasks, AW-0 scores 71.1% on average versus DreamZero’s 46.0%; rubric scores measure partial-credit progress, not full task success.
Key insights
- Programs can provide semantic and procedural context for video and action prediction; ablations favor the code-to-video-to-action attention order and full programs with reusable skills.
- Joint generation outperforms autoregressive code generation in the reported real-world evaluation and reduces inference latency from 0.55 to 0.23 seconds on an RTX 5090.
- Program validity is not the same as reliable execution: programs may parse yet choose incorrect entities or parameters, and hybrid execution trades off performance across evaluation conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.