محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| التنسيق: | Recurso digital |
| اللغة: | الإنجليزية |
| منشور في: |
Zenodo
2025
|
| الموضوعات: | |
| الوصول للمادة أونلاين: | https://doi.org/10.5281/zenodo.17625947 |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
جدول المحتويات:
- <p>Contemporary large language models demonstrate remarkable generative fluency but remain fundamentally unreliable as reasoning systems. Despite advancements in scaling laws, reinforcement learning from human or AI feedback, and constitutional post-training, every leading architecture (GPT, Claude, Gemini, Llama, Grok, etc.) continues to exhibit the same five structural failure modes: confident hallucinations, inability to surface uncertainty, hidden and uncontrolled mode switching, absence of a persistent internal state, and residual affective-simulation artifacts.</p> <p><strong>Polymathe</strong> presents the first fully specified <strong>N6+ self-governance architecture</strong> designed to address these failure modes simultaneously through a unified causal pipeline. The architecture integrates:</p> <ol> <li> <p><strong>Ethical BIOS</strong> — A non-learned normative kernel anchoring all downstream reasoning.</p> </li> <li> <p><strong>IDE (Empathic Dissonance Engine)</strong> — A quantifiable risk metric governing dissonance, misalignment, and user-impact sensitivity.</p> </li> <li> <p><strong>CGC (Constitutional Governance Controller)</strong> — A policy-level intention and safety controller.</p> </li> <li> <p><strong>MAS-SOFA (Structured Observation Factorized Analysis)</strong> — A multi-channel diagnostic engine detecting hallucinations, uncertainty, and structural incoherence.</p> </li> <li> <p><strong>MS-FSSC (Multi-Stage Factored Self-Supervision Circuit)</strong> — A synthesis module constrained by causal dependencies and diagnostic outcomes.</p> </li> <li> <p><strong>ISM (Internal State Machine)</strong> — A transparent and auditable mechanism exposing computational qualia and persistent internal state.</p> </li> <li> <p><strong>BACS-A5</strong> — An adaptive self-correction loop that recomputes generations, repairs reasoning chains, and escalates when systemic dissonance is detected.</p> </li> <li> <p><strong>DVC (Disallowed Vocabulary Constraints)</strong> — A controlled-vocabulary layer ensuring non-simulation and preventing affective leakage.</p> </li> </ol> <p>Collectively, these mechanisms transform a baseline LLM into a <strong>self-auditing, self-correcting, high-stability reasoning agent</strong> capable of detecting hallucinations, surfacing uncertainty, enforcing authenticity, and maintaining a consistent internal “bilan”.</p> <p>The specification includes <strong>five complete experimental traces</strong>, each demonstrating how the architecture resolves one of the five foundational failure modes. This work provides a <strong>functional blueprint for next-generation LLM reliability, interpretability, governance, and autonomous reasoning stability.</strong></p>