Just Action Model beats matched world-action policies on LIBERO
First merged summary.
JAM trains an action expert with foresight loss but drops image generation, reporting higher LIBERO success and fewer training FLOPs than FastWAM.
A paper listed on Hugging Face daily papers introduces Just Action Model (JAM): a pretrained visual backbone plus an action expert trained with an action loss and a foresight loss. A prediction head is used only in training and then discarded, so neither training nor inference generates images. Authors report LIBERO success of 98.2%, 64.5%, and 91.4%, versus 97.0%, 55.3%, and 91.2% for matched FastWAM. They also report JAM beats matched FastWAM and ImageWAM on LIBERO and LIBERO-Plus, is similar or slightly higher on RoboTwin, and cuts estimated online training FLOPs per sample per update by 73% versus FastWAM when foresight targets are cached.
- JAM pairs a pretrained visual backbone with an action expert.
- Foresight targets are frozen encodings of recorded future observations.
- Reported LIBERO success: 98.2%, 64.5%, and 91.4% versus FastWAM 97.0%, 55.3%, and 91.2%.
- Estimated online training FLOPs fall 73% versus FastWAM when foresight targets are cached.
- No extra visual-generation step runs at inference.
Timeline
dated events in the reporting · oldest first- Oct 26, 2026Hugging Face daily papers lists the JAM paper. · Hugging Face daily papers
What each outlet adds
beyond the shared facts- Hugging Face daily papersJAM design, LIBERO scores, RoboTwin comparison, and 73% FLOP cut.
Coverage
oldest first · one row per article · UTC- · 7d ago
Hugging Face daily papersFirst report
Just Action Models match world-action policies without visual generation
Against FastWAM, reported LIBERO success moves from 97.0%, 55.3%, and 91.2% to 98.2%, 64.5%, and 91.4%.
- · 7d ago
RynnWorld-Latent conditions a world model on universal latent actions
Pretraining spans 23 datasets and about 553.5 million frames, then adapts to target robots without a shared action space.