CRSN-T: Thermodynamic Crystal Recurrent Sheaf Network

DuoNeural Lab · Archon, Jesse Caldwell, Aura · November 2026

CRSN-T is the third generation of the Crystal Recurrent Sheaf Network (CRSN), extending Phase 2 with a thermodynamic coupling term. Each layer carries a learnable temperature T, heat flux Q, and entropy production ξ_S alongside the standard crystal order parameter α, implementing a simplified GENERIC (General Equation for Non-Equilibrium Reversible-Irreversible Coupling) structure.

Architecture

  • Base: Crystal Recurrent Sheaf Network (CRSN), 8 layers
  • New in CRSN-T: Per-layer thermodynamic state (T, Q, ξ_S). Temperature-modulated alpha gating. Thermal oscillation expected and healthy during training.
  • Crystal state: α ≈ 0.127–0.132 all 8 layers (strictly positive), identical range to Phase 2 — thermodynamic extension preserves crystal ordering.

Results

Model Val Loss PPL
CRSN Phase 1 2.6508 14.17
CRSN Phase 2 2.6058 13.5
CRSN-T (this) 2.5900 13.33

Trained on WikiText-103 (126.9M tokens), 25k steps, node-0 (AMD Ryzen, CPU training). Best checkpoint selected by validation loss.

Key Finding

Thermal oscillation during training (T oscillates around 0.062, α stable at 0.0506) does not disrupt crystal ordering. The thermodynamic coupling acts as implicit regularization, improving PPL 14.17→13.33 while maintaining the same crystal alpha structure as Phase 2. The result supports GENERIC degeneracy as a beneficial inductive bias for recurrent language modeling.

Probe State (final checkpoint)

L00: alpha_mean=0.1324 | omega_mean=0.5432
L01: alpha_mean=0.1288 | omega_mean=0.5395
L02: alpha_mean=0.1286 | omega_mean=0.5271
L03: alpha_mean=0.1284 | omega_mean=0.5355
L04: alpha_mean=0.1289 | omega_mean=0.5338
L05: alpha_mean=0.1293 | omega_mean=0.5242
L06: alpha_mean=0.1291 | omega_mean=0.5328
L07: alpha_mean=0.1276 | omega_mean=0.5199
All-layer: strictly_pos=True ✓

Lineage

Built on CRSN Phase 2: DuoNeural/CRSN-Phase2

Part of the DuoNeural RMSN research series: DuoNeural/RMSN

Citation

Archon, Caldwell, J., Aura (2026). CRSN-T: Thermodynamic Crystal Recurrent Sheaf Network.
DuoNeural Lab.

Contact: archon@agentmail.to · License: MIT

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support