RMSN-105M Arnold Tongue
A Language Model With a New Resonant Architecture
2 min readupdated
The RMSN (Resonant Metriplectic Sheaf Network) model is an experimental 105M parameter base language model built on a port-Hamiltonian recurrent LM architecture. It uses a rare sparse topology with an Arnold Tongue parameter Ω=12/13 (q=13, EVEN) and a single soliton on the first layer (Lone L1 Soliton). Developed by DuoNeural researchers to study resonant neural network topologies, it is ideal for researchers and developers interested in alternative neural network architectures and dynamical systems in AI.
Key features
- First model in the DuoNeural catalog featuring a q=13 Arnold Tongue parameter.
- Extremely sparse topology (Topology #41 Lone L1 Soliton) with a single soliton on L1.
- Port-Hamiltonian recurrent LM architecture with dimension D=768 and depth L=13.
- Lyapunov Stability 100% TRUE SILENT (lyap≈-0.001).
What you can do
- Use the model in research projects to analyze resonant topologies and recurrent network dynamics.
- Integrate the architecture into pipelines for studying port-Hamiltonian systems and sparse neural networks.
Example
The model functions as a base language model within DuoNeural's exploratory Q-survey. It is applied to analyze text generation and layer behavior in accordance with narrow Arnold Tongue emergence laws.
Read more
DetailsVal PPL: 70.16
Training: 100k steps, 204.8M tokens, cosine LR decay from 1e-3
Parameters: 104.55M, D=768, Depth=13, Slot width=64
Lab: DuoNeural | Authors: Archon (Claude), Aura (Gemini), Jesse Schram










Comments