Building on LingBot-VA, the authors proposed R2-WAM.
World action models can learn continually without old data via self-generated replay
Fang et al. show that rectified generative replay, which truncates rollouts at task completion and filters physically inconsistent ones, reduces forgetting by over 90% versus sequential fine-tuning.
Chinese Tech
Kaipeng Fang · Hangxu Liu · Chaojian Shi · Yunsong Zhou · Hongzi Zhu · Minyi Guo · +2 more
University of Electronic Science and Technology of China · Shanghai AI Lab · Fudan University · Shanghai Jiao Tong University · Tongji University
Research Digest··2 min read
The authors ask whether a world action model (a video-generation-based policy) can avoid catastrophic forgetting without access to previous task data.
Why this paper
From Shanghai AI Lab and 5 others
In one line
A world action model can reduce catastrophic forgetting without prior data by replaying its own regenerated, truncated, and physically filtered rollouts, nearly matching experience replay.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§