Quatrix: A New Neural Network Architecture Replacing Attention Mechanism with Reinforcement Learning Value Navigation

Quatrix proposes the Q-Compass architecture, which replaces traditional attention mechanisms with a value navigation mechanism based on reinforcement learning Q-functions. It significantly reduces KV cache requirements while maintaining performance. This architecture has advantages such as high parameter efficiency and structured sparsity, and supports multi-modal (language, image, audio) and cross-domain applications (e.g., cancer research).

Since the advent of Transformer, the attention mechanism has become the core of deep learning, but the expansion of model scale has led to a surge in computational complexity and memory requirements. Quatrix aims to replace similarity-based attention with value navigation to solve the resource consumption problem of traditional attention.

The Q-Compass block only requires three projection matrices (W_s, W_a, W_o). It determines the information flow path by calculating Q-values between the state (current position) and actions (other positions). The absence of W_V design avoids information loss. The SAVO variant introduces projection of Q-value products to balance value and content expression capabilities, leading to a significant improvement in perplexity metrics with the same number of parameters.

KV cache compression effect is significant: when rank r=H/8, it is 12.5% of traditional MHA; when r=H/16, it is only 6.25% with minimal performance loss. SAVO outperforms standard MHA across multiple parameter scales, performs well in cross-modal applications (language, image, audio) and cancer mutation feature models, and has strong transferability.

Quatrix provides a Python implementation and can be installed via pip. Example codes demonstrate pure text and multi-modal configurations, supporting quick start (e.g., training on the TinyShakespeare dataset).

Quatrix has published two academic papers proving the effectiveness of the value navigation mechanism. Future research will focus on the NanoG1 cancer foundation model. This architecture rethinks the essence of sequence mixing and provides a new path for large model efficiency and multi-modal unification.