enver

/

ayncoding-qwen3-8b-slim

Like 0

Paper Release: Hierarchical Symbolic-Neural Mixture of Experts (H-MoE)

Author: Enver AynEngine
Affiliation: Sovereign Epistemic AI Research / AynEngine Project
Preprint PDF: Download H_MoE_RESEARCH_PAPER.pdf
arXiv Source Package: Download arxiv_package.tar.gz

Highlights

• 95.0% Pass@1 on OpenAI HumanEval at ~7.55s per problem on pure dual-socket Intel Xeon CPU (a 27.3x speedup over raw 8B autoregressive decoding).
• Sub-1ms Symbolic AST Logic Gate eliminating $90%+$ of heavy parameter DRAM streaming over inter-socket buses.
• Classical Arabic Epistemic Governance (Al-Ḥadd bi al-Dhātiyyāt, Dafʿ al-Dawr, Dafʿ al-Tasalsul, ʿAdam al-Tanāquḍ, Sībawayh Governance) eliminating vague identifiers, cyclic deadlocks, infinite recursion, and silent error swallowing.

We welcome feedback and contributions from the machine learning, systems, and software engineering communities!