DepthMuun/gfn-gssm-mniah-k2

📍 G-SSM mNIAH Solver (Official Checkpoint)

This repository contains a Geodesic State Space Model (G-SSM) optimized for the Multi-Needle-in-a-Haystack (mNIAH) retrieval task.

🚀 Highlights

• Architecture: Geodesic State Space Model (G-SSM).
• Parameters: 8,109 (PyTorch verified).
• Scalability: Verified 100% accuracy up to 32,000 tokens context.
• Inference Efficiency: Constant $O(1)$ VRAM.

💡 Technical Note on $K$: This model is trained for exactly $K=2$ needles. Its deterministic behavior (no hallucination of state changes) is a direct result of geodesic integration. Variable $K$ support is a matter of training curriculum, not architectural limitation.

🔬 Scientific Significance

This model solves the Multi-Needle retrieval task using only 8,109 parameters, demonstrating that long-term dependencies can be integrated into a physical state through geodetic flow rather than quadratic self-attention. This represents a ~100x parameter reduction compared to Transformer baselines solving similar context lengths.

💻 Technical Usage (Inference)

To run inference locally, you need the GFN Framework installed.

1. Install GFN Framework

pip install gfn

2. Clone this repository

git lfs install
git clone https://huggingface.co/DepthMuun/gfn-gssm-mniah-k2
cd gfn-gssm-mniah-k2

3. Run Inference Script

Use the included inference.py script to test sequence retrieval:

python inference.py

Python API Example

As shown in inference.py, use the gssm high-level API:

import torch
from gfn import gssm

# Load model (automatically detects config.json)
model = gssm.load("mniah_model_final.pt", device="cpu")
model.eval()

# Sequence test
# Model returns (logits, state, info)
logits, _, _ = model(sequence)
# Geometric decoding logic is detailed in inference.py

📜 Citation

If you use this work, please cite:

@article{sturtz2026geometry,
  title={Geometric Flow Networks: A Physics-Informed Paradigm for Sequential Intelligence},
  author={Stürtz, Joaquín},
  journal={Zenodo Preprints},
  year={2026},
  doi={10.5281/zenodo.19141133},
  url={https://doi.org/10.5281/zenodo.19141133}
}

🔗 Resources

• Interactive Demo: Hugging Face Space
• Framework Source: GitHub: DepthMuun/gfn
• Official Paper: Zenodo