NeFT: A New Neuron-Level Fine-Tuning Method for LLMs

NeFT proposes a neuron-level supervised fine-tuning method. By precisely identifying and fine-tuning the subset of neurons most relevant to the target task instead of updating all parameters, it achieves more efficient LLM fine-tuning, striking a new balance between parameter efficiency and model performance, and is applicable to various practical scenarios.

Traditional full-parameter fine-tuning of large language models has high computational cost and large memory requirements; parameter-efficient fine-tuning methods (such as LoRA, Adapter) often sacrifice performance. In recent years, the existence of "knowledge neurons" inside LLMs has been discovered, inspiring NeFT to propose a neuron-level fine-tuning paradigm that focuses on updating specific neurons.

The core idea is neuron-level fine-tuning granularity, with the process divided into three stages: 1. Neuron importance evaluation (heuristic/gradient methods, calculating the relationship between activation values and loss); 2. Selection strategy (threshold/Top-K, updating 5%-20% of neurons); 3. Parameter update (full update of selected neurons, freezing pre-trained weights for unselected ones).

1. Higher parameter efficiency: Avoids low-rank approximation loss, outperforming methods like LoRA under the same parameter budget; 2. Better interpretability: Can analyze the activation of task-related neurons to understand decisions; 3. Flexible granularity control: 1%-50% fine-tuning granularity options; 4. Cross-task transfer: Some neurons are shared across related tasks.

On the GLUE benchmark, using 10% of parameters achieves 95% of the performance of full fine-tuning, outperforming LoRA; instruction fine-tuning shows better performance; it resists forgetting in continuous learning; training memory usage is reduced by 60%-80%, and inference speed is consistent with the original model.

Applicable to scenarios such as edge device deployment, multi-tenant services, continuous learning systems, and rapid prototype development.

Limitations: Neuron evaluation has additional overhead, and currently only targets FFN layers. Future directions: Efficient evaluation algorithms, research on attention layers, exploration of robustness relationships, and multimodal expansion.

NeFT represents the trend of LLM fine-tuning towards fine-grained precision, balancing parameter efficiency and performance, and provides a new option for reducing fine-tuning costs while maintaining quality.