MTLoRA is a novel parameter-efficient training framework for multi-task learning (MTL) that introduces Task-Agnostic and Task-Specific Low-Rank Adaptation modules. These modules effectively disentangle the parameter space in MTL fine-tuning, enabling the model to balance task specialization and interaction. MTLoRA is applied to hierarchical-transformer-based MTL architectures, adapting them to multiple downstream dense prediction tasks. Experiments on the PASCAL dataset show that MTLoRA achieves higher accuracy on downstream tasks compared to fully fine-tuning the MTL model while reducing the number of trainable parameters by 3.6×. MTLoRA establishes a Pareto-optimal trade-off between the number of trainable parameters and the accuracy of downstream tasks, outperforming current state-of-the-art parameter-efficient training methods in both accuracy and efficiency. The code is publicly available. MTLoRA introduces two variants: Task-Agnostic Low-Rank Adaptation (TA-LoRA) and Task-Specific Low-Rank Adaptation (TS-LoRA). TA-LoRA captures shared features across tasks, while TS-LoRA learns task-specific features. MTLoRA+ further improves performance by adding low-rank decomposition modules to the patch-merging layers in vision transformers. Experiments show that MTLoRA outperforms other parameter-efficient training methods, including adapters, BitFit, VPT, Compacter, LoRA, VL-Adapter, Hyperformer, and Polyhistor. MTLoRA achieves higher accuracy with fewer parameters, demonstrating its effectiveness in multi-task learning. The framework is implemented using PyTorch and is publicly available on GitHub. The method is evaluated on the PASCAL MTL dataset, showing significant improvements in accuracy and efficiency. MTLoRA is also tested on larger backbones and pre-training datasets, demonstrating scalability and effectiveness. The results show that MTLoRA provides a Pareto-optimal trade-off between parameter efficiency and task accuracy, making it a promising approach for multi-task learning.MTLoRA is a novel parameter-efficient training framework for multi-task learning (MTL) that introduces Task-Agnostic and Task-Specific Low-Rank Adaptation modules. These modules effectively disentangle the parameter space in MTL fine-tuning, enabling the model to balance task specialization and interaction. MTLoRA is applied to hierarchical-transformer-based MTL architectures, adapting them to multiple downstream dense prediction tasks. Experiments on the PASCAL dataset show that MTLoRA achieves higher accuracy on downstream tasks compared to fully fine-tuning the MTL model while reducing the number of trainable parameters by 3.6×. MTLoRA establishes a Pareto-optimal trade-off between the number of trainable parameters and the accuracy of downstream tasks, outperforming current state-of-the-art parameter-efficient training methods in both accuracy and efficiency. The code is publicly available. MTLoRA introduces two variants: Task-Agnostic Low-Rank Adaptation (TA-LoRA) and Task-Specific Low-Rank Adaptation (TS-LoRA). TA-LoRA captures shared features across tasks, while TS-LoRA learns task-specific features. MTLoRA+ further improves performance by adding low-rank decomposition modules to the patch-merging layers in vision transformers. Experiments show that MTLoRA outperforms other parameter-efficient training methods, including adapters, BitFit, VPT, Compacter, LoRA, VL-Adapter, Hyperformer, and Polyhistor. MTLoRA achieves higher accuracy with fewer parameters, demonstrating its effectiveness in multi-task learning. The framework is implemented using PyTorch and is publicly available on GitHub. The method is evaluated on the PASCAL MTL dataset, showing significant improvements in accuracy and efficiency. MTLoRA is also tested on larger backbones and pre-training datasets, demonstrating scalability and effectiveness. The results show that MTLoRA provides a Pareto-optimal trade-off between parameter efficiency and task accuracy, making it a promising approach for multi-task learning.