MoRA: High-Rank Updating for Parameter-Efficient Fine-Tuning

1. Unclear Compression and Decompression Operations: The compression and decompression operations described in equations (8) and following are not clearly explained. The rationale behind choosing these specific operations is unclear and requires further elaboration.

2. Rotation Operation: While the rotation operation is likely intended to preserve positional information after compressing the input, the explanation provided in the paper is not sufficiently clear. The authors need to clarify why rotation is added and how it enhances the matrix's ability to distinguish input information. The current description, especially the statement "rotation information can help the square matrix to distinguish the input information," is vague and needs more technical justification.

3. Performance Discrepancy in Table 1: In Table 1, MoRA with rank 8 achieves 100% accuracy, whereas LoRA with the same rank only achieves 52%. This discrepancy suggests that the LoRA model may have been undertrained or improperly tuned. The paper should address this discrepancy to ensure fair comparisons between methods.

4. Limited Improvement Across Tasks: MoRA shows clear improvements only in continual pre-training tasks, with little or no improvement in other downstream tasks. This suggests that MoRA is a task-specific method, which contradicts the goal of parameter-efficient fine-tuning methods that are expected to generalize across various tasks. This task-specific nature makes it challenging to justify its broad application without further experimentation, which defeats the purpose of an efficient fine-tuning method.

5. Uncertain Advantage of Rotation: The paper mentions that rotation helps the square matrix distinguish input information, but this benefit is not clearly demonstrated, particularly in higher-rank scenarios. At rank 256, the performance differences between the Sharing, Decouple, and Rotation methods are minimal. This raises the question of whether rotation is truly necessary at higher ranks, and the paper should clarify under what conditions (if any) rotation provides a real advantage.

6. For table 5, could you also provide the number of trainable parameters for both the methods under both ranks?

1. The primary advantage of MoRA lies in tasks requiring significant enhancement of LLM knowledge and capabilities, such as memorizing UUID pairs and continual pretraining (CP). However, it may be more practical to directly use full finetuning in such cases, particularly for CP, which generally demands high computational resources to achieve optimal performance. This raises questions about the suitability of applying PEFT in such high-cost scenarios. As shown in Table 2, MoRA cannot outperform full FT in CP tasks. Additionally, on downstream tasks outside of CP, MoRA does not exhibit a clear advantage over baseline methods and generally results in decreased performance compared with standard LoRA, which limits its practicality.

2. In the final update matrix $\Delta W$ obtained with MoRA, many entries are zero. This could be a drawback, as some of these parameters might be essential for specific downstream tasks, yet MoRA is unable to update them. This limitation potentially restricts MoRA's effectiveness for LLM updates across a diverse range of downstream tasks.

3. The method section could benefit from improved organization, especially as there are no flowcharts or diagrams to visually illustrate the MoRA method. Currently, certain lines, such as the definitions of $f_{comp}$ and $f_{decomp}$ are somewhat confusing and could be clarified for better comprehension.

1. In sec 5.1 and sec 5.3, MoRA is only compared with standard LoRA, more baselines(e.g., [1][2]) are needed to substantiate your hypothesis that high-rank updating benefits memory-intensive tasks.
2. It appears that the experiments in Section 5.2 are aimed at demonstrating MoRA's superiority in memory-intensive tasks and its comparable performance on other tasks. However, results indicate that MoRA underperforms other baselines in mathematical reasoning, which I consider memory-intensive, as increasing the rank from 8 to 256 significantly improves the performance of both LoRA and MoRA.
3. The writing needs significant refinement:

• Line 160: "we employ LLaMA-2 7B as base model, utilizing 1,000 pairs per batch and conducting 100 epochs".
• Line 165: "we observe low-rank updating are hard to memorizing new knowledge compared to FFT".
• Line 191: "enhances the output dimension".

[1] AdaLoRA: Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning

[2] DyLoRA: Parameter-Efficient Tuning of Pre-trained Models using Dynamic Search-Free Low-Rank Adaptation

1. The authors note that “increasing rank alleviates this problem” of memorizing UUID pairs. Although MoRA converges faster than LoRA at rank 256, they both end up converging. What are the benefits of using MoRA rank 8 vs. LoRA rank 256? According to Table 2, they get roughly the same performance. If MoRA is a strict pareto improvement over LoRA for the same number of parameters, the paper should state so unambiguously (the paper states this in a few separate places for specific tasks). If there are any caveats these should be stated clearly as well.

2. There is one recent study that is quite similar to MoRA: LoRA-XS (first appeared May 2024) “LoRA-XS: Low-Rank Adaptation with Extremely Small Number of Parameters” (Bałazy et al.) https://arxiv.org/abs/2405.17604. This is very similar to the approach taken by MoRA, and it should be cited as contemporaneous work. Another related (but less similar) study is “VeRA: Vector-Based Random Matrix Adaptation” (Kopiczko et al. 2023) https://arxiv.org/pdf/2310.11454

3. For different datasets, it would be very helpful to compare the total number of samples and tokens. The authors make a distinction between IFT with few samples (e.g. LIMA) and more samples, but don’t give enough detail to compare continual pretraining. How large are the biomedicine and finance corpora? The LoRA Learns Less and Forgets Less paper (https://arxiv.org/abs/2405.09673) makes a distinction between IFT and continual pretraining based on the total number of tokens (they find LoRA with rank 256 does worse than full finetuning when trained on >5B tokens). Are these data regimes similar?

Minor comments:

What is the hardware use for the speed/memory usage? This would be informative for Table 5. Otherwise it is somewhat meaningless The authors don’t define continual pretraining; presumably this means that the corpus is used to train with next token prediction. A handful of paragraphs suffer from poor/confusing grammar

Typos:

130: Compare → Compared

200: f_comp → f_decomp

256:use → uses

321: adapt to the → adapt to

342: 5.2.2 Baselines and Implements → Implementations

350: we search

355: The details parameters

362: grammar

369: grammar

399: leave modules layernorm → leave layernorm modules