Fielix - A Novel Neural Network Architecture

English | 中文

---

English

Fielix is a novel neural network architecture, different from Transformer, RNN, SSM and other existing architectures.

Core Innovations

1. Field Effect Propagation

Information propagates like a physical field in feature space, replacing traditional attention mechanisms.

2. Dynamic Topology

Connection patterns are dynamically generated based on content, rather than predefined.

3. Spiral Memory

Multi-timescale memory system with spiral-style compression storage.

4. Emergent Position Encoding

Position information emerges from token interactions, not predefined.

Experimental Results

Metric	Fielix (27M)	Transformer (21M)
Initial Loss	3.0	7.9
Final Loss	1.66	2.59
Learning Efficiency	Higher	Baseline
Training Speed	1.94x slower	Baseline

Fielix achieves better results with fewer training steps.

Installation

pip install torch
cd nexus

Quick Start

from models.nexus_model import FielixConfig, FielixForCausalLM

config = FielixConfig(
    vocab_size=32000,
    dim=512,
    num_layers=6,
    attention_type='field',  # Field effect attention
    use_memory=True,         # Enable spiral memory
)

model = FielixForCausalLM(config)

Architecture Components

fielix/
├── core/
│   ├── field_propagation.py   # Field effect propagation
│   ├── dynamic_topology.py    # Dynamic topology
│   ├── spiral_memory.py       # Spiral memory
│   ├── emergent_position.py   # Emergent position encoding
│   ├── feedforward.py         # Adaptive feedforward network
│   └── nexus_block.py         # Fielix basic block
├── models/
│   └── nexus_model.py         # Complete model
└── train/
    └── train_sharegpt.py      # Training script

TODO

• Training speed optimization (CUDA kernel fusion)
• FlashAttention-style field propagation
• Long context testing (8K, 32K, 128K)
• Continuous learning with spiral memory

License

Apache License 2.0 - Free to use and modify, but must credit the original author.

Citation

@misc{fielix2026,
  title={Fielix: A Field Effect Neural Network Architecture},
  author={Cherry},
  year={2026},
  url={https://github.com/aspect-love/fielix}
}

Contributing

PRs and Issues are welcome! Especially:

• Training speed optimization
• Long context experiments
• New application scenarios

---

中文

Fielix 是一个全新的神经网络架构，不同于 Transformer、RNN、SSM 等现有架构。

核心创新

1. 场效应传播 (Field Effect Propagation)

信息像物理场一样在特征空间中传播，替代传统的注意力机制。

2. 动态拓扑 (Dynamic Topology)

连接模式根据内容动态生成，而非预定义。

3. 螺旋记忆 (Spiral Memory)

多时间尺度的记忆系统，信息螺旋式压缩存储。

4. 涌现位置编码 (Emergent Position Encoding)

位置信息从 token 交互中涌现，而非预定义。

实验结果

指标	Fielix (27M)	Transformer (21M)
初始 Loss	3.0	7.9
最终 Loss	1.66	2.59
学习效率	更高	基准
训练速度	慢 1.94x	基准

Fielix 用更少的训练步数达到更好的效果。

安装

pip install torch
cd nexus

快速开始

from models.nexus_model import FielixConfig, FielixForCausalLM

config = FielixConfig(
    vocab_size=32000,
    dim=512,
    num_layers=6,
    attention_type='field',  # 场效应注意力
    use_memory=True,         # 启用螺旋记忆
)

model = FielixForCausalLM(config)

待优化方向

• 训练速度优化 (CUDA kernel fusion)
• FlashAttention 风格的场传播
• 长上下文测试 (8K, 32K, 128K)
• 螺旋记忆的持续学习能力

许可证

Apache License 2.0 - 可自由使用和修改，但需标注原作者

引用

@misc{fielix2026,
  title={Fielix: A Field Effect Neural Network Architecture},
  author={Cherry},
  year={2026},
  url={https://github.com/aspect-love/fielix}
}

贡献

欢迎提交 PR 和 Issue！特别欢迎：

• 训练速度优化
• 长上下文实验
• 新的应用场景