Gibbs Variational Last Layers for Long-Horizon Dynamics Learning
Gibbs Variational Last Layers for Long-Horizon Dynamics Learning
This research paper proposes a method using Gibbs variational inference to improve long-horizon probabilistic predictions in dynamical systems. The authors introduce multi-step training losses that incorporate long-horizon structure and robustness via continuous ranked probability scores (CRPS) for Variational Bayesian last layers (VBLLs). Evaluated across synthetic environments, chaotic dynamical systems, and real-world data, the approach shows substantial improvements in long-horizon accuracy and distributional fit compared to traditional single-step likelihood training.
The authors introduce a new family of multi-step training losses for Variational Bayesian last layers (VBLLs) that incorporate long-horizon structure and continuous ranked probability scores (CRPS), demonstrating improved accuracy on synthetic, chaotic, and real-world dynamical systems compared to single-step likelihood objectives.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
h retains the simplicity of VBLLs while directly optimizing for multi-step predictive quality. Across illustrative synthetic environments, a diverse suite of chaotic dynamical systems, and real-world data, we show that multi-step, CRPS-based objectives substantially improve long-horizon accuracy and distributional fit, consistently outperforming single-step likelihood training.
Open source S4
reliable multi-step rollouts. We revisit VBLLs through the lens of Gibbs variational inference, which decouples posterior inference from the choice of training loss. Building on this perspective, we introduce a family of multi-step training losses that progressively incorporate long-horizon structure, robustness via continuous ranked probability scores (CRPS), and training on full autoregressive rollouts. Our approac
Open source S3
arXiv:2611.02891v1 Announce Type: cross Abstract: Accurate long-horizon probabilistic predictions of dynamical systems are a prerequisite for model-based decision-making. Variational Bayesian last layers (VBLLs) are an attractive model class for this setting, offering tractable uncertainty at near-deterministic cost, yet they are trained exclusively with single-step likelihood objectives that may be insufficient for
Open source S2
Source：arXiv Robotics — research abstracts · arxiv.org