LPSE — Low-latency Semantic State Estimator

Model weights for LPSE, a low-latency semantic state estimator for dynamic network monitoring and orchestration, built on I-JEPA-style latent predictive learning over streaming telemetry.

Paper: A Low-Latency Semantic State Estimator using Latent Predictive Learning for Dynamic Network Monitoring and Orchestration (IEEE GLOBECOM 2026, to appear).

On a multi-node Kubernetes testbed, the framework reaches 82.42% semantic prediction accuracy under changing node configurations, with ~41x lower mean inference latency and a 15x smaller memory footprint than a deployable 4B LLM endpoint.

Files

File	Contents
lpsehybrid.pt	Full LPSE training checkpoint for the canonical merged-dataset run (model, EMA encoder, JEPA predictor, optimizer state)
baseline_mlp_v2.pt	MLP baseline checkpoint
baseline_xgb_v2.json	XGBoost baseline model
baseline_xgb_v2_encoder.pt	Frozen encoder used to build XGBoost features
baseline_xgb_v2_norm.pkl	Feature normalizer for the XGBoost baseline

Usage

git clone https://github.com/hari5g900/jepa.git
cd jepa
pip install -e .

CHECKPOINT_PATH=/path/to/lpsehybrid.pt python -m lpse.apps.infer_lpse

See the GitHub repository for full training, evaluation, and environment setup instructions (configuration is env-var driven; a GPU is required for the LPSE and LSTM/Transformer baselines).

Citation

@inproceedings{lpse2026,
  title     = {A Low-Latency Semantic State Estimator using Latent Predictive
               Learning for Dynamic Network Monitoring and Orchestration},
  booktitle = {IEEE GLOBECOM},
  year      = {2026},
  note      = {To appear}
}

License

MIT — see LICENSE.