Industrial world-model research line — JEPA-style latent state prediction on the O-series liquid core.
Same next-generation route as LeCun's JEPA family: abandon raw-signal reconstruction and autoregressive generation; predict and evolve abstract latent states instead. JEPA validates the route on vision/robotics; this repo pushes it down the industrial lane — continuous time, O(1) memory, edge-native — across audio, sensor, event-stream and physics modalities.
An experiment layer on top of AwareLiquid/M1
(vendored read-only as m1/). M1 owns the core (mt_lnn); this repo owns the
world-model objectives and training loop only. No M1 files are modified.
AwareLiquid-World/
├── m1/ # git submodule -> AwareLiquid/M1 (read-only reference)
├── world/ # world-model experiment layer
│ ├── objectives.py # JEPA-style objectives (masked latent prediction,
│ │ # SIGReg, latent-variable head, action conditioning)
│ ├── data.py # modality frontends reuse: log-mel, sensor streams, events
│ └── train.py # training loop skeleton
├── PLAN.md # the roadmap: what exists, what to build, GPU needs
└── tests/ # smoke tests (no training)
| Module | Role in the world-model line |
|---|---|
mt_lnn/world_model.py — PredictiveStateHead |
V-JEPA-style latent predictor (BYOL/V-JEPA EMA target, collapse-free across 3 seeds). Already the evaluation head. |
mt_lnn/imagination.py — LatentImagination |
multi-step latent rollout, 0 params |
mt_lnn/sensory_frontend.py |
raw sensor stream → backbone tokens (the modality path) |
mt_lnn/physics_ops.py |
Hamiltonian + symplectic rollout (the physics modality) |
mt_lnn/salience_events.py / pipeline.py |
salience ignition + dual-speed loop (the "awareness" wiring) |
O1-Sound (AwareLiquid/O1-Sound) |
audio frontend reference (log-mel, streaming C kernel) |
- Masked latent-state prediction — mask 70–90% of a temporal segment and
predict the latent state of the masked region (never reconstruct the raw
signal). Objective in
world/objectives.py. - SIGReg (isotropic Gaussian regularisation) — constrain the latent state distribution toward a unit Gaussian; replaces teacher-student/EMA plumbing for collapse prevention, zero inference cost.
- Explicit latent variables — a latent-factor dimension in the state that produces multi-path future rollouts (per-scenario evolution, not one number).
- Action-conditioned latent prediction — drive the state transition with an action vector for end-to-end trajectory/world-model planning.
Research scaffolding, nothing measured yet. The modules exist in M1 and are tested in their own contexts; the world-model objectives here are untrained. Milestones and success criteria in PLAN.md. No GPU is currently attached — the first training run needs an A100-class machine.
git clone --recurse-submodules https://github.com/AwareLiquid/AwareLiquid-World
cd AwareLiquid-World
pip install -r m1/requirements.txt -r requirements.txt # after they exist- AwareLiquid/M1 — the MT-LNN core
- AwareLiquid/O1-Sound — audio modality reference
- awareliquid.ai — benchmarks and retractions