| --- |
| title: "LIMEN Dynamics Series: Internal Stability of LLMs" |
| tags: |
| - llm |
| - interpretability |
| - dynamics |
| - qwen |
| - hidden-states |
| - limen |
| - idchain |
| license: cc-by-nc-nd-4.0 |
| --- |
| |
| # LIMEN Dynamics Series: Internal Stability of LLMs |
|
|
| **A Comprehensive Analysis of Trajectory Instability, Regime Taxonomy, and Methodological Robustness in Large Language Models.** |
|
|
| **Author:** Jean-Denis Bosange Batuli |
| **Date:** May 2026 |
|
|
| --- |
|
|
| ## The Series |
|
|
| This dataset hosts the full technical reports and supporting data for the LIMEN Dynamics research series. We recommend reading them in order: |
|
|
| ### 1. Four Dynamical Regimes in Large Language Models: An Empirical Phase Map |
| **[Download on Zenodo](https://doi.org/10.5281/zenodo.20348878)** |
| *The Foundation.* Introduces the `ct_t` metric and identifies four consistent regimes: Underactive, Adaptive, Transition, Chaotic. Demonstrates why Qwen models exhibit unique structural stability across 158 runs and 10 models. |
|
|
| ### 2. Conditional Dynamic Signatures in Large Language Models |
| **[Download on Zenodo](https://doi.org/10.5281/zenodo.20361289)** |
| *The Rigor.* A large-scale audit across 17 open-source models (70M to 3B parameters, 1,224 runs) examining normalisation sensitivity, variance decomposition, documented falsifications, and non-stationarity profiles. Proves that dynamic stability is a conditional observable: architecture explains 7% of variance, prompt category 17%, residual variance 76%. |
|
|
| ### 3. Dynamic-Layer Controllability without Universal Semantic Recovery |
| **[Download on Zenodo](https://doi.org/10.5281/zenodo.20400171)** |
| *The Depth.* Explores how instability propagates across transformer layers and identifies specific "Dynamic Layer Signatures" that predict output regimes before generation completes. Presents the fragmented topology at panel scale and the COLLAPSE-RIVALRY cyclical structure with 84% reopening rate on 638 cycles. |
|
|
| --- |
|
|
| ## Key Findings |
|
|
| | Finding | Source | Details | |
| | :--- | :--- | :--- | |
| | **Four Dynamical Regimes** | Paper 1 | UNDERACTIVE (1.55-1.70), ADAPTIVE (2.27-2.92), TRANSITION (~2.97), CHAOTIC (4.42-35.55) | |
| | **Qwen is Structurally Stable** | Paper 1 | Qwen family is the only family observed in the ADAPTIVE zone across 10 models | |
| | **Stability is Conditional** | Paper 2 | 76% of variance is residual/prompt-dependent. Architecture alone explains only 7% | |
| | **80% Trajectories Non-Stationary** | Paper 2 | Only 13% of token-level trajectories remain in a single regime throughout generation | |
| | **Documented Falsifications** | Paper 2 | Six small-panel hypotheses did not survive scale-up to n=17 | |
| | **Layer Coordination Matters** | Paper 3 | Inter-layer phase synchronization predicts trajectory collapse before it happens | |
| | **COLLAPSE-RIVALRY Cycle** | Paper 3 | 84% reopening rate after collapse on 638 observed cycles | |
| | **Fragmented Topology** | Paper 3 | Fragmentation index 0.65, only two robust mutual pairs (opt-pythia, phi-qwen) | |
|
|
| --- |
|
|
| ## Repository Structure |
|
|
| - `Four_Dynamical_Regimes.pdf`: Paper 1. |
| - `Conditional_Dynamic_Signatures.pdf`: Paper 2. |
| - `Dynamic_Layer_Controllability.pdf`: Paper 3. |
|
|
|
|
| --- |
|
|
| ## Collaboration & Contact |
|
|
| This research is conducted by **IDChain SRL** (operating under the **Unbind** brand), focusing on real-time trajectory coherence detection for BCI and AI systems. |
|
|
| - **Website:** [unbind.world](https://www.unbind.world) |
| - **LinkedIn:** [Jean-Denis Bosange Batuli](https://www.linkedin.com/in/jean-denis-bosange-b2b693326) |
| - **Contact:** jean.bosange@gmail.com |
|
|
| --- |
|
|
| ## Citation |
|
|
| If you use this work or dataset, please cite the relevant paper: |
|
|
| ```bibtex |
| @misc{bosange_batuli_2026_limem_series, |
| author = {Bosange Batuli, Jean-Denis}, |
| title = {LIMEN Dynamics Series: Internal Stability of LLMs}, |
| year = {2026}, |
| publisher = {Zenodo / Hugging Face}, |
| |
| |
| |
| |
| |
