LaminarNet: Linear-Time Multi-Scale State Propagation for Efficient Language Modeling

Transformer architectures dominate modern language modeling but incur O(N²) computational and memory costs with respect to sequence length, limiting scalability. We introduce LaminarNet, a linear-time architecture that replaces quadratic self-attention with two key mechanisms: Geometric Drift Field (GDF) for selective state propagation and Cross-Stratum Routing (CSR) for multi-scale hierarchical token interaction. LaminarNet achieves significant improvements in efficiency and performance. In a parameter-matched benchmark (49M parameters), it reduces perplexity by 38.3%, increases throughput by 63.3%, and reduces VRAM usage by 38.8% compared to Transformers. Additionally, a 437M-parameter model trained on 9.56B Turkish tokens demonstrates strong scaling behavior, achieving a perplexity of 11.38. Code: https://github.com/Uunan/LaminarNet Model: https://huggingface.