Multi-step consistency targets for latent world models
An auxiliary loss that pulls a world model's multi-step rollout toward targets anchored in real observations. The theory says when it is harmless and when it must help; the experiments so far confirm the theory and have not yet shown an advantage over the standard target.
- Status
- In progress
- Topics
- world models, self-supervised learning
Acoustic early warning, and where it breaks
A Random Forest classifier on MFCC features with local inference, reaching strong held-out metrics that then failed on real recordings while misflagging ambient noise. The generalization gap, not the accuracy number, is the finding.
- Status
- 2026
- Topics
- audio ML, sim-to-real, embedded