Frozen Transformers in Language Models Are Effective Visual Encoder Layers

5. Q: Additional ablation studies.

We appreciate the reviewer’s excellent suggestions. We would like to again emphasize that our work is the first to discover that “despite being trained solely on text data, a frozen LLM transformer can be used as a visual encoder layer for purely visual tasks, in the absence of language”. So, we prioritize validating the general applicability of this discovery across a wide range of visual tasks, in the most straightforward, simplest, and unified manner – achieved by inserting a single LLM layer at the end of visual backbones. We do not claim that this simple strategy is the optimal approach for achieving the best performance. While we acknowledge that there are various variants to utilize specific LLM layers and insert them into visual backbone architectures which may further improve performance, we deliberately avoid optimizing model selection with more complicated or task-specific architectures.

Following the reviewer’s suggestions, the table below presents an ablation study on different architecture variants. While some variants improve the performance while others decrease the performance, these ablation results contribute to strengthening our understanding and are orthogonal to our main discovery. We will include the results in the revision. And we leave a more comprehensive investigation on these and additional variants as interesting future work.

Experimental setting: We conduct the ablation experiments under our ablation setting of the original submission: ViT-S with 100 epochs on ImageNet. This is shorter than a full training of 300 epochs, due to our resource constraints. However, the results in the table are all compared under the identical and fair setup.

• Q: Inserting more LLM blocks. Please refer to Row 6 in the table above "+ 2 LLaMA Blocks." We use the final 2 LLaMA blocks, instead of the final single block, to evaluate the impact of using more LLaMA blocks. We find that using more LLM blocks seems beneficial for the overall performance. We think this is an interesting direction to further broaden our discovery and will include it in the revised paper.

• Q: Other places to insert. Please refer to Rows 4 and 5 in the table above "+ LLaMA at Middle" and "+ LLaMA at Head." They perform worse than inserting the LLM transformer block at the end of ViT (Row 2), supporting our design choice. Additionally, inserting at the end of encoders is a simple and unified strategy that can be employed across a wide range of tasks and diverse backbones; in contrast, achieving insertion at the middle may require scrutinizing backbones for each task individually.

4. Q: Efficiency information.

First, we agree that sharing detailed configurations is beneficial from an application perspective, and per the reviewer’s request, the table below presents these comparisons. Using a frozen LLM transformer adds negligible trainable parameters and slightly decreases the speed in image classification, while mainly resulting in an increase in overall parameters and FLOPs.

(The FLOPs numbers are evaluated with flops_profiler using a standard image of shape $224\times 224\times 3$. FPS comes from running the inference on 50k ImageNet-val images using batch size of 1 and 256 and a single A6000 GPU).

Second, and more importantly, we would like to clarify the fundamental differences between our work and directly scaling up ViTs. As mentioned earlier, our approach is not tailored for optimizing ViTs, and studies focusing on image classification can develop larger models that surpass our paradigm with proper design and tuning. In contrast, we would like to reiterate that our goal is to reveal the novel and intriguing discovery that despite being trained solely on text data, a frozen LLM transformer can surprisingly function as a visual encoder layer for diverse purely visual tasks, in the absence of language. Therefore, our approach is not intended to optimally scale up and enhance individual tasks, but to reveal that a frozen set of parameters can be shared across various tasks and backbones. Even from the perspective of improving models, we would like to clarify:

• Scaling up ViT can lead to improvement in image classification, but the scaled-up layers may not necessarily extend to diverse tasks as we do, such as point clouds, videos, and vision-language tasks.
• To the best of our knowledge, simply scaling up ViTs does not exhibit the emergent behavior of concentrating on relevant regions as in our information filtering hypothesis. Furthermore, our approach empirically shows such patterns not limited to images. Our discovery of this behavior covers videos, point clouds, 2D vision-language, and 3D vision-language tasks. We summarize three representative results (Figures 3, 5, C from our original submission) in [this anonymous image link], for the convenience of checking.

3. Q: Missing baselines.

We agree with the reviewer that baselines in addition to ViT-S-MLP will strengthen our paper. However, we are not sure about the suggestion to “append frozen ViT layers behind the original ViT,” because the LLaMA transformers have different parameter numbers compared to ViT transformers. Inferring from the mention of “to keep the total number of parameters the same,” we presume the reviewer is suggesting baselines with equal parameter counts to ViT-S-LLaMA. Therefore, we add the following two sets of experiments and hope they align with what the reviewer suggests. If our interpretation is incorrect, we kindly request the reviewer to provide further clarification, and we are more than willing to address any concerns.

• We first add the experiment of training from a randomly initialized LLaMA block, which has the same number of parameters but more trainable parameters. The tables below cover all the tasks investigated in our paper, including image classification, point cloud recognition, action recognition, motion forecasting, 2D vision-language, and 3D vision-language. We will include these results in the revision. The results in the tables show two main findings which validate that the performance gains truly stem from our method: (1) Our method of adding a frozen LLaMA transformer block consistently outperforms using a randomly initialized transformer block. (2) Naively adding a large transformer block can be detrimental for many tasks, resulting in inferior performance even compared to the plain baseline without the addition of a transformer block. This is because the naive strategy increases the challenges for optimization and suffers from a small training dataset compared to LLM pretraining.

• The second evidence of aligning with the equal parameter counts is presented in Figure 2 (Page 6) of our original submission. We analyze how the performance of image classification changes with different LLaMA or OPT layers. As shown in the figure, the models can have varied accuracy under the same trainable parameter and total parameter counts. Therefore, it indicates the importance of having proper pretrained weights in the inserted LLM blocks, such as our final LLaMA block, to provide benefits across diverse visual tasks.

We appreciate the reviewer for the insightful feedback. We are glad to hear that the reviewer considers our paradigm “simple yet effective” and recognizes our detailed investigation on how it works. Below we address each of the reviewer’s concerns.

---

1. Clarification of our contribution on diverse tasks.

The listed summary and strengths by the reviewer focus on ViT. We would like to kindly clarify that our discovery of “a frozen LLM transformer transfers to visual tasks” is not limited to ViT-based architectures, but applicable to a diverse range of tasks and backbones different from ViT, such as point cloud recognition, motion forecasting, and 3D vision-language tasks. Along this line, our major objective is to reveal the surprising generalizability of a frozen LLM transformer, trained only on text data, to a wide range of visual tasks, which is different from aiming to scale up ViT or optimize individual tasks. We hope that this clarification better highlights our contribution and addresses the reviewer’s overall concerns.

---

2. Q: Clarification of method.

We apologize for the confusion and thank the reviewer for pointing this out. Our method is consistent with Figure 1 (Page 2), i.e., the LLaMA transformer is behind the last FFN block in ViT. We will revise the statement to “behind the last transformer block in ViT” to avoid confusion.

We appreciate the reviewer for the insightful feedback. We are glad to hear that the reviewer finds our discovery of frozen LLM transformers being visual encoders intriguing, and recognizes our efforts in covering a wide range of tasks and developing the analysis of the information filtering hypothesis. Meanwhile, we understand the reviewer’s concerns about our baselines. Below we address all the points.

Following the reviewer’s suggestions, we conducted additional experiments with two sets of baselines: (1) using additional MLPs, aligning the number of trainable parameters; (2) using randomly initialized LLM transformer blocks, aligning the total number of parameters, to compare with our approach. The tables below summarize the results for all the tasks investigated in our paper, including image classification, point cloud recognition, action recognition, motion forecasting, 2D vision-language, and 3D vision-language.

We will include these results in the revision. The results in the tables show several important findings which validate that our performance gains stem from our method rather than increased network capacity:

• Our method of adding a frozen LLaMA transformer block consistently outperforms “-MLP” models.
• More importantly, different from the observation in the image classification task (ablation in Table 6 of our submission), simply adding MLPs is not a consistently beneficial strategy for all visual tasks and can, in fact, be detrimental, resulting in inferior performance on some tasks even compared to the pain baseline. This is because naively adding large MLPs without the guidance from the pre-trained LLM transformer block may lead to challenges in optimization, especially for modalities and tasks characterized by weaker semantics, such as point clouds and motion forecasting.
• Similarly, our method of adding a frozen LLaMA transformer block consistently outperforms using a randomly initialized transformer block.
• Similarly, naively adding a large transformer block can be even detrimental for many tasks, resulting in inferior performance even compared to the baseline without the addition of a transformer block. This is because the naive strategy increases the challenges for optimization and suffers from a small training dataset compared to LLM pretraining.

2. Q. Significance of improvements.

We understand the reviewer’s concerns regarding the significance of our improvements. And we believe that the MLP experiments detailed above have already addressed most of these concerns. Additionally, we would like to emphasize that our objective is not solely focused on improving specific tasks, but rather showing the general applicability of frozen LLM transformers across a wide range of tasks.

Here, we would like to further clarify that the performance improvements we achieved are significant for each task, according to the benchmarks established in the published papers within the specific subfield.

• Image classification. In our paper (Table 1 on Page 4, Table 6 on Page 6), our ViT-LLaMA model outperforms ViT-S and ViT-S-MLP by 0.6% and 0.3%, respectively. We believe that such improvements are significant for ImageNet1K. To contextualize, we refer to a notable study that includes exhaustive comparison between neural network architectures. ConvNext (Liu et al.) compared to its previous state of the art, SwinTransformer (Liu et al.) in Table 1, showing a 0.1% improvement for small architectures and a 0.3% for base architectures, which are smaller or comparable to our achieved improvement.

Liu et al. A ConvNet for the 2020s. CVPR 2022.

Liu et al. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. ICCV 2021.

• Point cloud recognition. In our paper (Table 2 on Page 4), we mainly focus on the challenging ScanObjectNN dataset to evaluate point cloud recognition, where we improve by 0.7% in the hardest T50 setting. As a reference, in the ScanObjectNN's paper (Table 4 in ScanObjectNN, Uy et al.), the top-performing methods have the accuracy of 77.9%, 78.1%, and 78.5%, with gaps of 0.2% and 0.4%, which are smaller than our improvement.

Uy et al. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. ICCV 2019.

• Action recognition. We achieve a ~1% improvement on SSv2, which is significant. As a reference, VideoMAE shows a ~0.4% improvement over its previous best method in Table 6 of Tong et al., which is smaller than our gain. This supports that our improvement is significant.

Tong et al. VideoMAE: Masked autoencoders are data efficient learners for self-supervised video pre-training. NeurIPS 2022.

• Motion forecasting. We improve minADE of mmTransformer from 1.10 to 1.08 (the smaller the better). In mmTransformer's original paper (Table 1, Li et al.), its improvement over previous state of the art is also around 0.02. This supports that our improvement is significant.

Liu et al. Multimodal motion prediction with stacked transformers. CVPR 2021.

• 2D Vision-language. We refer to the METER (Dou et al.) paper and VQAv2 as a reference. We improve the performance of METER on VQAv2 by 0.6. According to METER's Tables 2-4, the advantage of selecting the best visual backbone or language encoder also shows a similar scale of advantage over the remaining models. Hence, we believe that our improvement also significantly benefits the visual encoding.

Dou et al. An empirical study of training end-to-end vision-and-language transformers. CVPR 2022.

• 3D Vision-language. We improve the baseline of SQA3D by around 0.9, while in SQA3D's original paper (Table 3, Ma et al.), the best method exhibits a 0.6 advantage over the second one. This verifies the significance of our improvement.

Ma et al. SQA3D: Situated question answering in 3d scenes. ICLR 2023.

5. Q: How will it scale?

We thank the reviewer for the insightful question. Based on our current empirical evaluation, we do not have conclusive evidence indicating the performance gain in scale-up scenarios. Moreover, it seems that this may also vary depending on specific tasks. For example, the improvement from LLM blocks decreases from ViT-Tiny to ViT-Small (Table 1, Page 4) in image classification, but increases from ViT-Small to ViT-Base (Table 3, Page 5) in action recognition. As mentioned by the reviewer, “this is difficult to evaluate,” requiring substantially increased computational resources, and we leave this line of exploration as interesting future work.

On the other hand, we would like to further emphasize that our approach is not intended for an optimal direction to scale up models for individual tasks. Instead, it is designed to reveal that a frozen LLM transformer with unchanged parameters can surprisingly function as an encoding layer for diverse visual tasks.

---

6. Q: Clarification of Tab. 6 and Fig. 2?

Sorry for the confusion and we will revise the paper to better clarify this. Both Table 6 and Figure 2 (Page 6) are for image classification on ImageNet. The only difference is: Table 6 is trained for the standard 300 epochs, while the ablation in Figure 2 is trained for a more efficient 100 epochs, due to our limited resources. Specifically, fully training one model on ImageNet would occupy most of our GPUs for 3-4 days.

---

7. Q: Questions on the analysis of the information filtering hypothesis?

The visualization in Figure 3 and the quantitative evaluation in Figure 4 are consistent with our information filtering hypothesis. Below we clarify the specific questions one by one.

• Q: Which model is evaluated in Fig. 3? Figure 3 uses ViT-S and ViT-S-LLaMA models for visualization.

• Q: Why $F_{LM}^A$ has a lower score on ViT-B? Note that Figure 3 shows the visualization on ViT-S and ViT-S-LLaMA. Now we present additional visualization on ViT-B and ViT-B-LLaMA, shown in [this anonymous image link]. This image explains why $F_{LM}^A$ in ViT-B-LLaMA has a lower score than ViT-B: the attention layer $F_{LM}^A$ still clearly distinguishes relevant regions from irrelevant regions, but the relevant regions are assigned lower activations, which corresponds to a smaller IoU in Figure 4. However, the illustration in the new ViT-B figure and Figure 4 are still consistent with our information filtering hypothesis, because the LLM layer finally converts the relevant regions into highly-activated tokens to enlarge their contribution in $F_{LM}^F$ and $F_L^{2}$.

• Q: Is the observation consistent across architectures? Our information filtering hypothesis has consistent observation across architectures: the features processed by our LLM blocks ( $F_{LM}^F$ and $F_{L}^2$) consistently show a clearer concentration to relevant regions. Meanwhile, we acknowledge that the specific behaviors and utilities of each layer remain an open-problem, e.g., the observed phenomenon that $F_{LM}^A$ can discern tokens in diverse ways.

3. Q: Additional ablation studies.

We appreciate the reviewer’s excellent suggestions. We would like to first emphasize that our work is the first to discover that “despite being trained solely on text data, a frozen LLM transformer can be used as a visual encoder layer for purely visual tasks, in the absence of language”. So, we prioritize validating the general applicability of this discovery across a wide range of visual tasks, in the most straightforward, simplest, and unified manner – achieved by inserting a single LLM layer at the end of visual backbones. We do not claim that this simple strategy is the optimal approach for achieving the best performance. While we acknowledge that there are various variants to utilize specific LLM layers and insert them into visual backbone architectures which may further improve performance, we deliberately avoid optimizing model selection with more complicated or task-specific architectures.

Following the reviewer’s suggestions, the table below presents an ablation study on different architecture variants. While some variants improve the performance while others decrease the performance, these ablation results contribute to strengthening our understanding and are orthogonal to our main discovery. We will include the results in the revision. And we leave a more comprehensive investigation on these and additional variants as interesting future work.

Experimental setting: We conduct the ablation experiments under our ablation setting of the original submission: ViT-S with 100 epochs on ImageNet. This is shorter than a full training of 300 epochs, due to our resource constraints. However, the results in the table are all compared under the identical and fair setup.

• Q: Inserting more LLM blocks. Please refer to Row 6 in the table above "+ 2 LLaMA Blocks." We use the final 2 LLaMA blocks, instead of the final single block, to evaluate the impact of using more LLaMA blocks. We find that using more LLM blocks seems beneficial for the overall performance. We think this is an interesting direction to further broaden our discovery and will include it in the revised paper.

• Q: Other ways to insert. Please refer to Rows 4 and 5 in the table above "+ LLaMA at Middle" and "+ LLaMA at Head." They perform worse than inserting the LLM transformer block at the end of ViT (Row 2), supporting our design choice. Additionally, inserting at the end of encoders is a simple and unified strategy that can be employed across a wide range of tasks and diverse backbones; in contrast, achieving insertion at the middle may require scrutinizing backbones for each task individually.

---

4. Q: Inserting layers from a visual model.

Following the reviewer’s suggestion, we use the last transformer layer of CLIP-B/16 and insert it into ViT-S/16 for image classification. The result in the table below shows that this variant does not improve the performance. We hypothesize that this result may indicate that the pre-encoded knowledge in CLIP does not bring in additional useful knowledge for training visual backbones within our current paradigm.

We appreciate the reviewer’s perspective on interpreting our method as a more general approach to transferring knowledge from any foundation model beyond LLMs. As clarified earlier, our paper focuses on LLM transformers, for the reason exactly pointed out by the reviewer – “the transfer from a language model to a vision model is interesting and most surprising that it works.” Investigating other types of foundation models contributes to strengthening our understanding and is orthogonal to our main discovery on LLM transformers. In addition, the result in the table above suggests that investigating other types of foundation models is not trivial, and we leave this line of exploration as interesting future work.

We appreciate the reviewer for the thoughtful feedback. We are glad to hear that the reviewer finds our discovery intriguing, covering a wide range of tasks, and the analysis of the information filtering hypothesis insightful. Indeed, we hope that our surprising yet widely applicable findings will motivate new ideas in the community. Below we address each of the reviewer’s concerns.

---

1. Q: Comparing with “-MLP” models.

We agree that this is a valid concern and conducted additional experiments with the “-MLP” baseline for all the tasks investigated in our paper, including point cloud recognition, action recognition, motion forecasting, 2D vision-language, and 3D vision-language, in addition to the already reported result for image classification, as shown in the tables below. We will include these results in the revision. The results in the tables show two main findings which validate the advantage of our method:

• Our method of adding a frozen LLaMA transformer block consistently outperforms “-MLP” models.
• More importantly, different from the observation in the image classification task, simply adding MLPs is not a consistently beneficial strategy for all visual tasks and can, in fact, be detrimental, resulting in inferior performance on some tasks even compared to the pain baseline. This is because naively adding large MLPs without the guidance from the pre-trained LLM transformer block may lead to challenges in optimization, especially for modalities and tasks characterized by weaker semantics, such as point clouds and motion forecasting.

6. Q: Attention mask and positional embedding.

We would like to clarify that our approach uses the attention mask strategies and positional embeddings identical to those used in the corresponding visual tasks. We do not use the auto-regressive attention masks and rotary positional embeddings designed for LLMs, since they are not directly applicable to visual tasks, and thus we do not include experiments for this. Specifically, most visual models do not use auto-regressive attention masks or rotary positional embedding, because their tokens are 2- or 3-dimensional and should be considered all at once, compared to a 1-dimensional temporal sequence in language. For example, ViT (Dosovitski et al.) for classification uses learnable positional embedding and no attention masks; and VectorNet (Gao et al.) for motion forecasting encodes 3D coordinates into features and only uses attention masks to indicate padded tokens. Therefore, we directly follow the common practice of each visual task.

Dosovitski et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021.

Gao et al. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. CVPR, 2020.

5. Q: Additional ablation studies.

We appreciate the reviewer’s excellent suggestions. We would like to first emphasize that our work is the first to discover that “despite being trained solely on text data, a frozen LLM transformer can be used as a visual encoder layer for purely visual tasks, in the absence of language”. So, we prioritize validating the general applicability of this discovery across a wide range of visual tasks, in the most straightforward, simplest, and unified manner – achieved by inserting a single LLM layer at the end of visual backbones. We do not claim that this simple strategy is the optimal approach for achieving the best performance. While we acknowledge that there are various variants to utilize specific LLM layers and insert them into visual backbone architectures which may further improve performance, we deliberately avoid optimizing model selection with more complicated or task-specific architectures.

Following the reviewer’s suggestions, the table below presents an ablation study on different architecture variants, together with some ablation provided in the original submission. While some variants improve the performance while others decrease the performance, these ablation results contribute to strengthening our understanding and are orthogonal to our main discovery. We will include the results in the revision. And we leave a more comprehensive investigation on these and additional variants as interesting future work.

Experimental setting: We conduct the ablation experiments under our ablation setting of the original submission: ViT-S with 100 epochs on ImageNet. This is shorter than a full training of 300 epochs, due to our resource constraints. However, the results in the table are all compared under the identical and fair setup.

• Q: Inserting more LLM blocks. Please refer to Row 6 in the table above "+ 2 LLaMA Blocks." We use the final 2 LLaMA blocks, instead of the final single block, to evaluate the impact of using more LLaMA blocks. We find that using more LLM blocks seems beneficial for the overall performance. We think this is an interesting direction to further broaden our discovery and will include it in the revised paper.

• Q: Other ways to insert. Please refer to Rows 4 and 5 in the table above "+ LLaMA at Middle" and "+ LLaMA at Head." They perform worse than inserting the LLM transformer block at the end of ViT (Row 2), supporting our design choice. Additionally, inserting at the end of encoders is a simple and unified strategy that can be employed across a wide range of tasks and diverse backbones; in contrast, achieving insertion at the middle may require scrutinizing backbones for each task individually.

• Q: Other LLM blocks. In the original submission, we analyzed different blocks from LLaMA and OPT in Fig. 2 (Page 6), under the same ablation setting. We found that different transformer blocks from LLaMA or OPT can improve image classification accuracy. Note that it is time and resource-prohibitive for us to try all possible combinations across all of our tasks. Hence, we simply chose the last block from LLaMA-7B, because it is sufficient for us to convey our discovery. As for the guidance of selecting blocks, our observation from Fig. 2 indicates that the final few blocks from LLM consistently lead to improvement.

3. Q: Point cloud recognition performance on ModelNet40.

We discussed and explained the marginal performance gain on ModelNet40 in the last paragraph of Page 4 in the original submission. ModelNet40 is an earlier dataset in point cloud classification, and the performance on it has saturated in recent years (Ma et al.). Hence, the observed slight performance drop (~0.2%) in the easier settings (1k, 4k points) is within the norm. Given this context, we concentrated on the more challenging and recent ScanObjectNN dataset, Table 2 (Page 4), for a more informative comparison. On ScanObjectNN, our approach achieves consistent improvements as shown in the table below (copied from Table 2 on Page 4), especially in the hardest T50 setting.

Ma et al. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. ICLR 2022.

---

4. Q: Theoretical justification.

In Sec. A.4 (Page 16) of the original submission, we discussed a potential theory that supports our information filtering hypothesis – “usable information theory” (Xu et al.). Specifically, Xu et al. posit that a well-trained neural network can be treated as a decipher and increases the usable information in latent features, so that subsequent layers can translate features into correct predictions/labels more easily. We think this theory aligns well with our observation: a pretrained LLM block amplifies the relevant tokens and naturally enables the decoder to predict better labels.

Additionally, we would like to point out that understanding LLMs and the training mechanism of neural networks remains open problems; so our paper primarily focuses on empirical discoveries and investigations. We leave a more in-depth theoretical analysis as interesting future work. Meanwhile, we hope that our information filtering hypothesis can offer additional perspectives to understand LLM transformers.

Xu et al. A theory of usable information under computational constraints. ICLR 2020.

We appreciate the reviewer for the thorough feedback and the recognition of our contribution, which generalizes pretrained LLM transformers to diverse visual tasks and introduces the information filtering hypothesis as explanation. Below we address each of the reviewer’s concerns.

---

1. Q: Feature visualization clarification.

We would like to first clarify that visualizing feature activations and attention maps is a widely-adopted strategy in related studies to investigate the emergent behaviors of ViT features, such as in DINO (Caron et al.) and AbsViT (Shi et al.). This is because these visualizations represent the most direct output and intermediate variables from transformers. In this work, we followed a similar strategy to demonstrate how features concentrate on relevant regions with our frozen LLM transformer, providing empirical validation for our information filtering hypothesis. Importantly, our visualizations extend beyond those conducted in DINO and AbsViT, by encompassing diverse tasks (videos, 3DVQA, etc.) and introducing the “frequency” activation to better capture the patterns of features. If the reviewer has alternative suggestions for visualization and analysis methods, we are happy to discuss and utilize them.

Caron et al. Emerging Properties in Self-Supervised Vision Transformers. ICCV 2021.

Shi et al. Top-Down Visual Attention from Analysis by Synthesis. CVPR 2023.

---

2. Q: Training from a randomly initialized LLM block.

In the tables below, we present the ablation studies on training from a randomly initialized LLM block for all the tasks investigated in our paper, including image classification, point cloud recognition, action recognition, motion forecasting, 2D vision-language, and 3D vision-language. We will include these results in the revision. The results in the tables show two main findings which validate that the performance gains truly stem from our method:

• Our method of adding a frozen LLaMA transformer block consistently outperforms using a randomly initialized transformer block.
• Naively adding a large transformer block can be detrimental for many tasks, resulting in inferior performance even compared to the baseline without the addition of a transformer block. This is because the naive strategy increases the challenges for optimization and suffers from a small training dataset compared to LLM pretraining.