They designed FlexiWorld, a latent world model based on JEPA (joint embedding predictive architecture) that does not reconstruct images but predicts in latent space.
Flexible action chunking improves long-horizon planning in JEPA world models
A JEPA-based world model with mixed-span goal supervision and variable-length action chunks achieves 89.29% mean success across four benchmarks, outperforming fixed-chunk baselines.
Big Tech
Shidu Ren · Qilin Gu · Zhenghao Ni · Junhan Sun · Jiaqi Wang · Damien Scieur · +1 more
University of Toronto · Zhejiang University · Tencent Jarvis Lab · Mila & Université de Montréal · Samsung SAIL
Research Digest··2 min read
The authors introduce FlexiWorld, a JEPA-based world model that trains with arbitrary goal spans and variable-length action chunks, along with an actor-residual planning method (ARCEM).
Why this paper
From Samsung SAIL and 5 others
In one line
FlexiWorld achieves 89.29% mean success across four benchmarks using variable-length action chunks and mixed-span goal supervision.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§