stable-worldmodel
Summary
stable-worldmodel (swm) is an open-source platform for reproducible world-model research and evaluation. It provides a shared interface for collecting trajectory data, training or plugging in world models, running model-predictive control solvers, and evaluating robustness under controllable visual, geometric, and physical factors of variation.
Official Artifacts
- Paper: https://arxiv.org/abs/2605.21800
- Code: https://github.com/galilai-group/stable-worldmodel
- Documentation: https://galilai-group.github.io/stable-worldmodel/
- Official X thread: https://x.com/lucasmaes_/status/2060022309759389774
Role In The Wiki
This entity should be tracked as infrastructure, not as a model family. Its value is that it makes action-conditioned world-model claims easier to audit:
observation history + action sequence
-> learned latent rollout
-> solver-selected control inputs
-> standardized success / latency / robustness metricsThat makes it directly relevant to World Models, JEPA, Robotics Time-Series Modeling, and Time-Series Benchmark Hygiene.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in stable-worldmodel-2026 rather than duplicating verdict rows here.
At the entity level, stable-worldmodel is the object card for the codebase, documentation, paper, and X announcement. The transfer hypothesis is that its standardized trajectory store, solver interface, and factor-of-variation evaluation could inspire comparable infrastructure for action-conditioned multivariate time-series and digital-world operational systems. The current released evidence remains visual/control-simulation evidence, not numeric telemetry, graph time-series, EHR, or observability intervention evidence.