Enregistré dans:
| Auteur principal: | |
|---|---|
| Format: | Recurso digital |
| Langue: | |
| Publié: |
Zenodo
2025
|
| Accès en ligne: | https://doi.org/10.5281/zenodo.18003638 |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
Table des matières:
- <h2><strong>Abstract</strong></h2> <p>We propose a post-training framework for inducing <strong>self-organized criticality (SOC)</strong> in large language models by tuning <strong>meta-cognitive control loops</strong> governing chain-of-thought generation, reflection, and context feedback. Rather than enforcing criticality at the level of neural activations or weights, we show that SOC can emerge behaviorally when slow adaptive parameters regulate fast inference dynamics. We define order parameters for reasoning depth, recursion, and self-revision, and describe a feedback protocol that drives their distributions toward scale-free statistics without task-specific tuning. This approach offers a principled mechanism for dynamically allocating cognitive effort, balancing robustness and adaptability, and may explain why near-critical regimes appear favorable for general intelligence.</p> <h2><strong>Core Thesis</strong></h2> <blockquote><strong>Criticality should be engineered at the level of meta-cognitive dynamics, not raw inference.</strong></blockquote> <p>Key claim:</p> <ul> <li> <p>LLMs already possess rich internal representational structure.</p> </li> <li> <p>What they lack is a <strong>self-organizing control regime</strong> that allocates reasoning depth adaptively.</p> </li> <li> <p>Self-organized criticality naturally emerges when:</p> <ul> <li> <p>Fast generative processes are regulated by</p> </li> <li> <p>Slow adaptive feedback variables</p> </li> <li> <p>Without external fine-tuning.</p> </li> </ul> </li> </ul> <p>This mirrors SOC in physical, biological, and cognitive systems.</p>