Defierithos: The Lonely Warrior Rises from Resonance

Libo Wang

Keywords: Wave, Sub-band, Self-Resonant Field (SRF) Mechanism, Wormhole Memory, Dynamic Chain-of-Thought

TL;DR: This work has been published on TechrXiv: https://www.techrxiv.org/doi/full/10.36227/techrxiv.174181746.67591790

Abstract: In view of the long-term computational complexity of the Transformer architecture in long sequence processing, as well as the redundancy and long-distance dependency attenuation caused by global self-attention, this work proposes a new architecture design named "Defierithos" that is based on self-resonance field (SRF). It replaces the traditional self-attention mechanism with partial resonant interference, uses spectrum modulation and phase superposition for information transmission, and effectively reduces the computational burden. To verify the effectiveness, the researcher used simulation-based training to build a training environment in a Python simulator based on custom GPTs to complete the execution of text generation (synthetic data), code programming (HumanEval), and mathematical reasoning (GSM8k). The experimental results show that Defierithos is superior to the GPTs version of Transformer in six indicators: ROUGE-L, METEOR, Pass@k, MMLU, accuracy, and ARC-AGI. It proves that as an architecture system that replaces tokens with waves instead of traditional self-attention mechanisms, it provides experience that can be used as a reference for the architecture of the next generation of natural language processing.

Submission Number: 151