EMR-Merging: Tuning-Free High-Performance Model Merging

The success of pretrain-finetune paradigm brings about the release of numerous model weights. In this case, merging models finetuned on different tasks to enable a single model with multi-task capabilities is gaining increasing attention for its practicability. Existing model merging methods usually suffer from (1) significant performance degradation or (2) requiring tuning by additional data or training. In this paper, we rethink and analyze the existing model merging paradigm. We discover that using a single model's weights can hardly simulate all the models' performance. To tackle this issue, we propose Elect, Mask & Rescale-Merging (EMR-Merging). We first (a) elect a unified model from all the model weights and then (b) generate extremely lightweight task-specific modulators, including masks and rescalers, to align the direction and magnitude between the unified model and each specific model, respectively. EMR-Merging is tuning-free, thus requiring no data availability or any additional training while showing impressive performance. We find that EMR-Merging shows outstanding performance compared to existing merging methods under different classical and newly-established settings, including merging different numbers of vision models (up to 30), NLP models, PEFT models, and multi-modal models.

Paper

Similar papers

Peer review

Reviewer b5vt5/10 · confidence 4/52024-07-07

Summary

The paper presents a method called EMR-MERGING (ELECT, MASK & RESCALE-MERGING) for merging models finetuned on different tasks into a single model with multi-task capabilities without the need for additional tuning or training. This method addresses the limitations of existing model merging techniques, which often suffer from performance degradation or require additional data and tuning.

Strengths

- **Theoretical and Empirical Analysis:** The paper provides a solid theoretical foundation for the effectiveness of the proposed method, including detailed proofs and empirical analyses. This adds robustness to the claims made by the authors. - **Tuning-Free:** One of the major advantages of this method is that it does not require any additional data, tuning, or training, making it highly practical for real-world applications where data availability is limited or privacy concerns restrict access.

Weaknesses

- **Additional Computational Overhead:** The proposed method introduces extra computational overhead during inference due to the use of masks and rescalers. This aspect should be discussed in the paper, particularly regarding its impact on the overall efficiency and performance in real-world applications. - **Storage Considerations for PEFT Models:** Pre-trained models finetuned on downstream tasks using Parameter-Efficient Finetuning (PEFT) methods typically require only the storage of the pre-trained model weights and additional adapter weights. These adapter weights consume significantly less storage space compared to the masks utilized in EMR-MERGING. The authors should address this comparison and discuss the storage implications in detail. - **Impact of Model Quantization:** With the trend towards quantization of models, reducing their precision from 32-bit to 8-bit, 2-bit, or even 1-bit, the storage overhead of masks becomes increasingly significant. Quantization is an essential step in the development of large models, and as such, the relative storage cost of the masks in EMR-MERGING will be more pronounced in the future. The authors should consider discussing the impact of model quantization on their method and the potential challenges it poses. - **Limited Dataset Size in Experiments:** The datasets used in the experiments are relatively small, which undermines the persuasive power of the results. For example, in the case of visual tasks, large models like CLIP's ViT-L exhibit strong generalization capabilities and can achieve high zero-shot classification performance on these small datasets. As a result, the performance differences among various model merging methods might not be significant. The authors should consider using larger, more diverse datasets to better demonstrate the effectiveness and robustness of their approach.

Questions

NA

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

NA

Reviewer b5vt2024-08-09

The author seems to have misunderstood my points regarding Weaknesses 2 and 3. I would like to clarify them as follows: - Weakness 2: A pre-trained model with N adapters can also perform multiple tasks. In this case, I don't quite understand why model fusion is necessary. What advantages does the proposed EMR-MERGING method have over the PEFT approach? PEFT allows for fine-tuning downstream tasks while reducing storage space requirements and achieving satisfactory results. Furthermore, PEFT consumes significantly less storage space compared to storing a mask. - Weakness 3: In cases of highly quantized models (e.g., 8-bit), EMR-MERGING would occupy 1/8 of the original model’s size, which significantly limits the effectiveness of the method. In contrast, PEFT does not encounter this issue.

Authorsrebuttal2024-08-09

Rebuttal by Authors

Thank you so much for your review and feedback. We further address your comments regarding Weakness 2 and Weakness 3 as follows: Answer to Weakness 2: 1) **Model finetuning and model merging are fundamentally different tasks with different targets**. For model fine-tuning, it consists of full-parameter fine-tuning that pursues fine-tuning performance and parameter-efficient fine-tuning (PEFT) that pursues fine-tuning efficiency, both of which have the same target of customizing models for specific domains, **necessitating labeled datasets and sufficient computational resources for supervised training**. For model merging, under the background that there exists an exponentialy increasing number of pre-trained or finetuned model weights, it targets to take advantage of these existing weights and obtain a single model with multi-task abilities, **without the need for labeled data or supervised training**. 2) **Model merging can be combined with kinds of model fine-tuning methods**. More specifically, the proposed **EMR-Merging can be applied to not only fully finetuned models but also PEFT models**, as shown in Tab. 6 of our paper. We only need the fine-tuned adapters on each task, and EMR-Merging can merge these adapters into a single adapter with a few masks and rescalers, whose overhead is less than storing all the adapters while achieving performance close to that of multiple individual adapters. 3) **Model finetuning and model merging are both significant techniques**. In this paper, our EMR-Merging focuses on realizing tuning-free and high-performance model merging instead of reducing storage space requirements as much as possible like PEFT. Besides, **model merging and PEFT can be easily combined to achieve further parameter reduction**, which is a potential area for future work. For example, there have been some studies focusing on merging PEFT models [1,2]. Answer to Weakness 3: 1) When **compared to multiple individual 8-bit models**, EMR-Merging that consists of one 8-bit model and several masks and scalars can still significantly reduce the parameter numbers. 2) When **compared to other merging methods**, EMR-Merging shows significantly better performance requring no tuning, demonstrating its greater applicability. 3) **A potential solution to your concern is using PEFT techniques to train the quantilized models and merge the PEFT modules using EMR-Merging**. This can minimize the number of parameters on multiple tasks, which will be included in our future work. ---- **References** [1] Parameter efficient multi-task model fusion with partial linearization, ICLR 2024. [2] Composing Parameter-Efficient Modules with Arithmetic Operations, NeurIPS 2023.

Reviewer b5vt2024-08-14

Thank you for your responses, I appreciate the effort and detail you have provided in addressing the concerns raised. I have no additional questions at this time and will determine my final score after conferring with the Area Chair and the other reviewers.

Authorsrebuttal2024-08-14

Thanks for your comments

We appreciate your comments, which can help improve our paper's quality greatly and inspire our future work. The discussion about the impact of model quantization and PEFT will be included in the revision. Thank you for your efforts. You are welcome to discuss with us if you have any other questions.

Reviewer Qs8J5/10 · confidence 5/52024-07-07

Summary

The authors identify two issues in current model merging methods: significant performance degradation from the multi-task counterpart and requirement of additional data and training. To tackle those issues, they propose EMR merging, which first creates a unified task vector, then selects masks based on each task-specific model’s correspondence with the unified task vector, finally rescales it. EMR merging shows strong empirical performance among several benchmarks.

Strengths

1. The paper is well written and the algorithm is coupled with adequate amount of analysis. 2. The proposed algorithm is computationally lightweight, yet obtains significant performance improvement.

Weaknesses

1. The novelty is limited. The three steps are essentially a combination of TIES and DARE, where only the masking step is slightly different from TIES. Due to this high similarity, it is unclear why the proposed method performs so much better than TIES and DARE. It would be better to provide a more disentangled analysis for each of the three steps through ablation studies, so that the readers can have a better understanding where exactly the performance improvement comes from. 2. Several related works are missing [1,2] (both on ArXiv in Feb 2024). It is important to do a thorough comparison, especially when the paper basically claims SoTA results. 3. Not all methods are compared on all benchmarks. For instance, AdaMerging is not compared with on NLP benchmarks. Although the original paper of AdaMerging does not provide results in NLP, I do not see any algorithmic obstacle that prevents it from being implemented in the NLP setting. [1] Representation Surgery for Multi-Task Model Merging. Yang et al. ICML 2024. [2] Merging Multi-Task Models via Weight-Ensembling Mixture of Experts. Tang et al. ICML 2024.

Questions

1. What is the sparsity of the masks? I know that it is kind of task/model dependent, but it would be nice to show those results. 2. Do EMR still need the hyperparameter tuning on the scaling factors? Or the one from rescalder is final? Because in DARE, they tune the scaling factors even after they compute the rescaling factors.

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors adequately addressed the limitations.

Reviewer S3RQ5/10 · confidence 4/52024-07-09

Summary

Model merging directly fuses multiple independently trained models at the weight level to obtain a single comprehensive model, which is a current research hotspot. This paper proposes a new model merging method EMR-MERGING to reduce the performance gap between the merged model and the independent models.

Strengths

- The proposed EMR-MERGING method does not require additional training, data, and tuning, which enhances the applicable scenarios of the method. - This paper conducts extensive experiments, including visual models, NLP models, and multi-modal models. In particular, the visual model is extended to 30 tasks. - This paper has a clear structure and the proposed methods are easy to implement.

Weaknesses

- **Loss of parallel ability**: Existing model merging methods obtain a unified model for all tasks, so samples from all tasks can be included in one inference (i.e., a batch of data). However, EMR-MERGING needs to configure an independent model for each task during inference, thus losing the ability of parallel inference. In real scenarios, a batch of data may come from different tasks rather than a single task. - **Important baseline missing**: It is unclear what the benefits and advantages of EMR-MERGING are compared to "sparse independent models". Ties-Merging shows that removing 90% of the neurons in each independent model has almost no change in performance. So why not just use the sparse version of the task vector instead of using EMR-MERGING? - **Symbol or spelling error**: - The symbols are not uniform, and the uniform vector is sometimes $T_F$ and sometimes $T_{uni}$. - Table 11: “huper-parameter settings” -> “hyper-parameter settings” - Sec 4.2.2: “using 1eleven datasets” -> “using eleven datasets” - Finally, there is a lack of datasets and checkpoints for CV (30 tasks), NLP, and multimodal in the anonymous link. If the authors can release them, it will help the community research.

Questions

- Is it possible to select only the maximum value in the Electing operation? Is it also possible to select the average value? - In Figure 4(c), regarding Cos similarity, TV > Ties-Merging > AdaMerging, but in terms of accuracy, TV < Ties-Merging < AdaMerging. Figure 4(b) is similar. These phenomena seem to be contradictory. How does the author explain this?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors discuss some limitations of the method: for example, it consumes additional memory compared to other model merging methods, and all rely on the same pre-trained model. However, as in Weaknesses, this method may lose the ability to parallelize inference on multiple tasks, and it is unclear how much benefit it has compared to sparse independent models.

Reviewer Lqs26/10 · confidence 4/52024-07-11

Summary

This paper aims to improve the performance of model merging in the multi-task learning domain. By electing a unified model, identifying a mask and a rescaling factor for each task, the proposed method EMR-Merging is able to significantly improve the merging performance over the previous state-of-the-art, such as Adamerging, Ties-Merging, Task Arithmetic, etc. Besides the CV model merging, EMR-Merging also effectively merges the large language models, approaching the individual fine-tuned performance.

Strengths

1. I like the idea that we need to elect a unified model first, and then identify different masks for different tasks. Directly merging models can lead to function conflicts. 2. The experiments are extensive, including VIT model, language models, and multi-modal models, with considerable good performance. 3. The ablation study is also provided.

Weaknesses

1. I am curious why the proposed way works for the unified model. One specific case is that, if the the average of the task vector is close to zero, but not exactly equal to zero, e.g., 0.0001, which will give us a positive sign. If the magnitude of this weight is large, a positive sign multiplied by this large magnitude weight can lead to a large error. Can the authors elaborate more the intuition behind the unified model, mask, and rescaler? 2. Can you explain the process of obtaining the mask? I could not find the definition of $\tau_F$ after reading the paper in detail. 3. What is the configuration of Figure 4? More details are needed here. 4. How to interpret the t-SNE visualization? 5. The distance between figures is over-reduced. It can be confusing and misleading. For instance, figure 5 and figure 4. 6. How expensive to extend EMR-Merging to LLMs, such as LLaMa3? 7. How does the proposed method perform compared to SOTA https://arxiv.org/pdf/2402.00433?

Questions

Please see the above weaknesses.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Provided.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC