TopoTuner: Topological Finetuning of Large Language Models
Artificial Intelligence
Summary
The authors developed TopoTuner, a method to fine-tune large language models more efficiently by only updating select parts of the model instead of all weights. They measure changes in the model's components using topological tools to decide which parts to keep frozen during training. TopoTuner learns patterns from one dataset and reuses them to speed up fine-tuning on new tasks, training far fewer parameters while keeping performance close to full fine-tuning. Their approach also reduces training time compared to popular methods like LoRA. This work suggests that fine-tuning strategies can be transferred across different tasks by understanding how model components change topologically.
Authors
Abdulkadir Erol, Yash Mahajan, Vepaul Hariprashad, Baha Rababah, Santu Karmaker, Cuneyt G. Akcora, Mubarak Shah
Abstract
Full fine-tuning remains a strong way to adapt pretrained LLMs, but it updates all weights and can be expensive. LoRA reduces the number of trainable parameters, but it does not directly answer which pretrained components should be trained and which can be frozen during adaptation. We introduce TopoTuner, a topology-guided fine-tuning framework for selective freezing of attention projection matrices. \method treats each projection matrix as a row cloud and uses Wasserstein distances between persistence diagrams to measure how its topology changes during fine-tuning. TopoTuner learns a reusable freezing profile from a source dataset and transfers it to efficiently fine-tune models on out-of-domain datasets, evaluating whether task-specific topological drift generalizes across question answering and sentiment analysis tasks. Across LLaMA-3.1-8B, Mistral-7B-v0.3, and Qwen3-8B-Base, TopoTuner is competitive with full fine-tuning while training only 1-2\% of the model parameters, and outperforms LoRA in 7 out of 9 model-dataset settings, which can change up to 39.57\% of the projection parameters. Along with minimized updates, TopoTuner reduces training time by 20.4\% relative to full fine-tuning and 5.5\% relative to LoRA on average. TopoTuner opens a new direction for reusable freezing profiles, where fine-tuning behavior learned on one dataset can be shared across multiple tasks.