Update README: add Space link + DHP probe confirmed results
Browse files
README.md
CHANGED
|
@@ -87,9 +87,30 @@ Training: TinyStories (full dataset), 30k steps, seq_len=256, batch=8, AdamW lr=
|
|
| 87 |
|
| 88 |
---
|
| 89 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
## Connection to DHP
|
| 91 |
|
| 92 |
-
The learnable damping coefficients γ_k parameterize per-slot relaxation timescales τ_{L,k} = 1/γ_k
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 93 |
|
| 94 |
See: [DHP papers — zenodo.org/communities/duoneural](https://zenodo.org/communities/duoneural)
|
| 95 |
|
|
|
|
| 87 |
|
| 88 |
---
|
| 89 |
|
| 90 |
+
## Interactive Demo
|
| 91 |
+
|
| 92 |
+
Try CDM V6 HORN in your browser — watch slots specialize in real time as the model generates text:
|
| 93 |
+
|
| 94 |
+
**[CDM-HORN-Demo — HuggingFace Space](https://huggingface.co/spaces/DuoNeural/CDM-HORN-Demo)**
|
| 95 |
+
|
| 96 |
+
Shows:
|
| 97 |
+
- **Slot Logit Lens**: per-generated-token view of what each slot is tracking (top-3 vocab tokens)
|
| 98 |
+
- **Oscillator Panel**: per-layer γ/ω/regime visualization — see which layers ring and which hold
|
| 99 |
+
|
| 100 |
+
---
|
| 101 |
+
|
| 102 |
## Connection to DHP
|
| 103 |
|
| 104 |
+
The learnable damping coefficients γ_k parameterize per-slot relaxation timescales τ_{L,k} = 1/γ_k, establishing a direct connection to the **Dynamical Horizon Principle** (DHP).
|
| 105 |
+
|
| 106 |
+
A formal DHP probe was run on CDM V6 HORN (perturbation sensitivity across sequence positions):
|
| 107 |
+
|
| 108 |
+
- **Analytical** (exact): DHO slot update characteristic eigenvalues are γ_k ± i·sqrt(ω_k²−γ_k²) (underdamped) or real-valued (overdamped). In both cases, the slow eigenvalue governing long-time perturbation decay is bounded above by γ_k. Therefore λ_k ≤ γ_k exactly from DHO math — **HORN's γ_k are direct per-slot DHP timescale bounds**.
|
| 109 |
+
- **Empirical**: Perturbation sensitivity s(Δ) ∝ e^{−λ·Δ} confirmed (R²=0.993–0.9998 for Layers 1–6). Empirical λ_emp ∈ [0.237, 0.297] ≈ 30% of γ — expected, as empirical probe measures compound sensitivity through full forward pass (attention + CDM + FFN), not just the in-slot DHO rate.
|
| 110 |
+
- **DHP predictability horizons**: τ*_k = 0.72/γ_k, ranging 0.81–1.10 sequence positions across layers.
|
| 111 |
+
- **Three-regime alignment**: Reactive L0 (low τ*, fast horizon), overdamped L1-5 (longer τ*), persistent resonance L6-7 (highest τ*) — three temporal horizons emergent from LM training, corresponding to distinct DHP timescales.
|
| 112 |
+
|
| 113 |
+
**CDM V6 HORN is an empirically confirmed DHP substrate.**
|
| 114 |
|
| 115 |
See: [DHP papers — zenodo.org/communities/duoneural](https://zenodo.org/communities/duoneural)
|
| 116 |
|