LoRALib: A Standardized Benchmark for Evaluating LoRA-MoE Methods

Shaoheng Wang, Yao Lu, Yuqi Li, Yaxin Gao, Jiaqi Nie, Shanqing Yu, Yingli Tian, Qi XuanarXiv 2025 · DOI · arXiv

Abstract

Low-Rank Adaptation (LoRA) is a popular technique for parameter-efficient fine-tuning, but its application in Mixture of Experts (MoE) settings lacks standardization. We introduce LoRALib, an open-source benchmark comprising 680 LoRA modules across 17 architectures and 40 tasks. Our evaluation reveals that the LoRAMoE framework currently outperforms other methods like Rocket and X-LoRA. Furthermore, we demonstrate that manual expert selection based on task relevance consistently surpasses automated selection mechanisms, highlighting a significant "expert selection gap" in current automated systems.

Introduction

While LoRA has revolutionized Parameter-Efficient Fine-Tuning (PEFT), it often suffers from a lack of generalization across diverse tasks. LoRA-MoE architectures attempt to solve this by routing inputs to specialized "experts." However, the field has become a "Wild West" where inconsistent models, datasets, and hyperparameters make objective comparison difficult. LoRALib addresses this by providing a unified, "apples-to-apples" benchmarking environment to identify which architectural and selection innovations truly drive performance.

Methods and Datasets

LoRALib consists of a standardized library of 680 LoRA modules trained on 40 downstream datasets. These modules span 17 model backbones (e.g., Llama 3.2, Qwen 2.5, Mistral) with sizes from 0.5B to 9B parameters. To ensure fairness, all modules use identical hyperparameters (rank=16, alpha=16).

The benchmarking pipeline evaluates three representative LoRA-MoE methods:

  • LoRAMoE: Employs a router in the Feed-Forward Network (FFN) with balancing constraints to preserve "world knowledge."
  • X-LoRA: A dynamic gating approach mixing adapters at token and layer levels.
  • Rocket: Uses K-shot retrieval to select expert candidates before fine-tuning.

The benchmark covers 40 datasets across domains including Reasoning & Math (GSM8K, MATH), General Knowledge (MMLU, ARC), Language Understanding (GLUE), Medical (MedQA), and Ethics.

Key Insights: The Expert Selection Gap

A critical finding is that the method used to select experts from the library is as vital as the routing mechanism itself. We compared three selection mechanisms: Perplexity-based (PPL), Random, and Manual.

Comparison of PPL, Manual, and Random selection mechanisms across various models and tasks.

Manual selection—where experts are chosen based on human-perceived task relevance—consistently outperformed automated PPL selection. PPL often has a "blind spot," favoring general logic experts (like LogiQA) over domain-specific ones (like MATH) for mathematical tasks, leading to sub-optimal results.

Key Results: Method Performance

When evaluated on a level playing field, LoRAMoE emerged as the clear leader. It consistently achieved the highest average precision across all tested model backbones.

Performance comparison of Rocket, LoRAMoE, and X-LoRA methods across multiple models and evaluation benchmarks.

On the Llama-3.1-8B-Instruct backbone, LoRAMoE achieved an average score of 61.87, significantly higher than Rocket (56.96) and X-LoRA (51.37). The benefits of both superior routing (LoRAMoE) and proper expert selection were most pronounced in larger models and complex reasoning tasks.

Discussion and Future Work

While manual selection is currently superior, it is not scalable for libraries containing thousands of experts. Future research should focus on automated selection strategies that better mimic human intuition. Additionally, the performance of training-free methods like X-LoRA lags behind, suggesting a need for improved "plug-and-play" MoE architectures. LoRALib provides the necessary foundation for these future explorations.

The datasets and LoRA library are available on Hugging Face.

Summary updated

Abstract

As a parameter efficient fine-tuning (PEFT) method, low-rank adaptation (LoRA) can save significant costs in storage and computing, but its strong adaptability to a single task is often accompanied by insufficient cross-task generalization capabilities. To improve this, existing work combines LoRA with mixture-of-experts (MoE) to enhance the model's adaptability through expert modules and routing mechanisms. However, existing LoRA-MoE methods lack unified standards in models, datasets, hyperparameters, and evaluation methods, making it difficult to conduct fair comparisons between different methods. To this end, we proposed a unified benchmark named LoRALib. Specifically, we standardized datasets from $40$ downstream tasks into a unified format, fine-tuned them using the same hyperparameters and obtained $680$ LoRA modules across $17$ model architectures. Based on this LoRA library, we conduct large-scale experiments on $3$ representative LoRA-MoE methods and different LoRA selection mechanisms using the open-sourced testing tool OpenCompass. Extensive experiments show that LoRAMoE performs best, and that prioritizing LoRAs relevant to the target task can further improve the performance of MoE. We hope these findings will inspire future work. Our datasets and LoRA library are available at https://huggingface.co/datasets/YaoLuzjut/LoRAOcean_dataset and https://huggingface.co/YaoLuzjut/models.
BibTeX
@misc{wang2025loralib,
  author = {Shaoheng Wang and Yao Lu and Yuqi Li and Yaxin Gao and Jiaqi Nie and Shanqing Yu and Yingli Tian and Qi Xuan},
  title = {{LoRALib: A Standardized Benchmark for Evaluating LoRA-MoE Methods}},
  year = {2025},
  eprint = {2509.18137},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.18137},
}
Figure