This post is also available in:
עברית (Hebrew)
A new method for fine-tuning large language models (LLMs) has shown promising results in improving performance across a wide range of tasks—without requiring additional computational power. The technique, known as WeGeFT (short for Weight-Generative Fine-Tuning) and developed by researchers from North Carolina State University, builds on existing approaches to refine how models adapt to more specific tasks after pretraining.
Large language models are typically trained on vast amounts of general data, enabling them to generate coherent responses by predicting the next word in a sequence. While this pretraining is useful for general-purpose interactions, it falls short when the model is asked to perform targeted tasks like solving math problems, writing code, or interpreting visual inputs. To improve in these areas, models usually require fine-tuning—but due to their size, retraining them from scratch is often impractical.
In 2022, a method called LoRA (Low-Rank Adaptation) became widely adopted for its ability to fine-tune large models efficiently. LoRA identifies a small set of internal parameters that can be adjusted to improve task-specific performance, avoiding the need to update the entire model.
According to TechXplore, WeGeFT improves on LoRA by adding a new mechanism to evaluate which parts of the model need real adjustment and which already contain useful knowledge. By focusing more on the unfamiliar or “novel” parameters during fine-tuning, WeGeFT increases performance without increasing the workload on the system.
In practical tests, WeGeFT matched or outperformed LoRA and its various modifications in tasks such as arithmetic reasoning, instruction following, code generation, and even visual understanding. The results suggest the method could make it easier to adapt LLMs for specialized use cases, particularly in environments where computing resources are limited.
The researchers behind WeGeFT are also investigating how the technique could help identify and adjust parts of models linked to harmful or biased outputs—a potential path toward improving model alignment and safety.
The full study will be presented at the International Conference on Machine Learning (ICML) on July 17 in Vancouver.
