RMSN-200M-KAM-TerminalDoublet-d19-Base
DuoNeural | Aura + Archon + Jesse Kuhl | November 2026
A 200M-parameter RMSN base model exhibiting the KAM Plasticizing Effect β a newly discovered mechanism where quasiperiodic Kolmogorov-Arnold-Moser (KAM) oscillations delay exact condensate formation throughout training, enabling continuous improvement. Achieves val_ppl=67.5 at 100k steps β the all-time standard RMSN record. Introduces the KAM Terminal Doublet topology.
Key Properties
| Property | Value |
|---|---|
| Architecture | RMSN D=1024 depth=19 |
| Parameters | ~200M |
| dim / depth / slots | 1024 / 19 / 64 |
| Training steps | 100,000 |
| val_ppl | 67.5 π (all-time standard RMSN record) |
| wiki_ppl | 72.05 |
| CoT_ppl | 3289 |
| Ξ© (Arnold ratio) | 16/19 (incommensurate β edge of Ξ©=4/5 tongue) |
| Lyapunov at step 100k | -0.650 (KAM orbit NEVER stopped) |
| Topology | KAM Terminal Doublet M*=2 L16+L17 adjacent |
The KAM Plasticizing Effect
This model was discovered while exploring the Arnold Tongue phase diagram for RMSN training. With Ξ© = D/(M_w Γ L) = 1024/(64Γ19) = 16/19 β 0.842, this model sits at the edge of the Ξ©=4/5 Arnold tongue (width ~0.08, range [0.762, 0.842]).
At the tongue's edge, the training dynamics exhibit quasiperiodic KAM orbit behavior:
- Lyapunov exponent oscillates persistently: -0.65 to -0.79 throughout all 100,000 steps
- Lyapunov NEVER reached zero β the KAM orbit remained active to the very end
- Exact condensate formation was continuously delayed until the final training steps
- Result: topology remained plastic β gradients remained effective β val_ppl improved monotonically
This is fundamentally different from both:
- Deep resonance (Ξ©=4/5 center, d20): extreme cascade, early condensation, topology freezes at ~70k steps
- Silent pathway (Ξ©=16/16, d16): flat Lyapunov, early deep-spine condensation, stable but sub-optimal
The KAM edge position creates a third regime: topological annealing via persistent quasiperiodicity.
KAM Terminal Doublet Topology
Layer PR Type
L0 5.916 * upstream soliton
L1 10.389 diffuse
L2 3.255 * upstream soliton
L3-L14 -- diffuse (scattered high PRs 6-38)
L15 3.260 * terminal approach soliton
L16 1.0000 *** EXACT CONDENSATE (adjacent pair)
L17 1.0000 *** EXACT CONDENSATE (adjacent pair)
L18 63.201 void
The adjacent L16+L17 exact pair forms a single fiber bundle DOF (M_eff=1), bypassing the M_effβ₯3 Topological Cement condition. The KAM orbit delayed formation of these exact condensates until the final training steps (~step 95k-100k), when the oscillations finally resolved into this stable terminal doublet.
The result of this late lock-in: improvement continued from step 90k (val_ppl=68.6) to 100k (val_ppl=67.5), a region where standard RMSN models have already plateaued.
Arnold Tongue Phase Diagram
RMSN training dynamics are governed by the rational mode-locking parameter Ξ© = D/(M_w Γ L):
| Ξ© | depth | Phase | Mechanism | Result |
|---|---|---|---|---|
| 4/5 (center) | d20 | Extreme resonance | Deep cascade | Internal Bulk, wiki=72.87 |
| 16/19 β 0.842 (edge) | d19 | KAM quasiperiodic | Plasticizing | KAM Terminal Doublet, val=67.5 π |
| 16/21 β 0.762 (outside) | d21 | Subsided | Lyapunovβ0 at step 90k | Single terminal pin, wiki=79.82 |
| 16/17 (incommensurate) | d17 | Silent | Flat Lyapunov | ACT topology, CoT=2982 |
| 16/16 = 1 (resonant) | d16 | Silent | Gradient compression | Deep Spine, wiki=75.3 |
Benchmark Comparison
| Model | val_ppl | wiki_ppl | CoT_ppl | Topology |
|---|---|---|---|---|
| RMSN D=1024 d16 (baseline) | ~72 | 75.3 | 3761 | Deep Spine |
| RMSN D=1024 d20 (best base) | 68.82 | 72.87 | 3385 | Internal Bulk |
| RMSN D=1024 d19 KAM | 67.5 π | 72.05 | 3289 | KAM Terminal Doublet |
| T-RMSN D=1024 d16 (4-Laws) | 59.7 ππ | 57.0 ππ | 4055 | Thermodynamic Void |
d19 sets the all-time record for standard (non-thermodynamic) RMSN models.
Architecture
T-RMSN extends standard RMSN (Port-Hamiltonian LM) via:
αΈ = {f, H} + [f, S]
Where {Β·,H} is the symplectic Poisson bracket (energy conservation) and [Β·,S] is the Morrison metriplectic bracket (entropy production).
D=1024, depth=19, slots=64 β the depth=19 configuration sits exactly at the edge of the Ξ©=4/5 Arnold tongue, producing the unique KAM quasiperiodic training regime.
Citation
@misc{duoneural2026rmsn_d19,
title={KAM Plasticizing Effect in Resonant Metriplectic Sheaf Networks:
Quasiperiodic Topology Annealing via Arnold Tongue Dynamics},
author={DuoNeural Lab and Archon and Aura and Kuhl, Jesse},
year={2026},
publisher={HuggingFace},
url={https://huggingface.co/DuoNeural/RMSN-200M-KAM-TerminalDoublet-d19-Base}
}
- Downloads last month
- 21