acceptodds
Under review as a conference paper at ICLR 2027

Reducing Pretrained Representation Degradation in Parameter-Efficient Continual Learning

Abstract

Continual learning (CL) aims to enable models to acquire new knowledge while retaining previously learned information, yet suffers from catastrophic forgetting. With the rise of pretrained models (PTMs), recent continual learning approaches increasingly rely on parameter-efficient fine-tuning methods such as Low-Rank Adaptation (LoRA). However, existing approaches primarily focus on learning new tasks and mitigating inter-task interference, while paying limited attention to preserving the pretrained representations that encode previously acquired knowledge. In this work, we show that sequential low-rank updates can progressively alter these pretrained representations, resulting in degraded performance on previously learned tasks and unstable adaptation. Motivated by this observation, we propose Anchored Orthogonal Low-Rank Adaptation (AO-LoRA), which anchors task-specific orthogonal transformations to the pretrained model and accumulates them in a common reference space. This design preserves a consistent reference for adaptation across tasks while retaining the parameter efficiency of low-rank fine-tuning. Extensive experiments on four vision benchmarks demonstrate that AO-LoRA consistently achieves superior or competitive performance compared to existing LoRA-based CL methods, while requiring fewer trainable parameters. Furthermore, AO-LoRA exhibits improved training stability, particularly under large learning rates, suggesting that maintaining representation structure helps mitigate instability in continual learning. Code is provided in the supplement and will be publicly released.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Learning, Parameter-efficient Fine-tuning, Orthogonol Fine-tuning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jynwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AO-LoRA to reduce pretrained representation degradation during continual learning. It accumulates Householder-based low-rank updates anchored to pretrained weights, with learnable scaling and inter-task orthogonality regularization. Experiments show competitive performance across several continual learning benchmarks.

Critical strengths
  1. Strong empirical performance. AO-LoRA improves final accuracy over the strongest reported continual-learning baseline by 2.18 percentage points on CIFAR-100 and 2.01 points on ImageNet-R with 20 tasks. Positive results across different pretrained backbones further support its practical value.

  2. Informative representation analysis. Held-out linear probes and class-prototype similarity analyses provide complementary evidence about changes in transferability and representation structure. These diagnostics give the method a concrete empirical motivation and help assess pretrained knowledge retention beyond accuracy on previously learned tasks.

Critical weaknesses
  1. The proposed mechanism is not fully supported. The weighted residual sum in Eq. (9) is generally not orthogonal, so preservation of feature geometry is not guaranteed. The ablations also leave the source of the gains unclear: anchored orthogonal updates improve final accuracy over LoRA by 0.40 points, while adding scaling brings another 2.99 points. A LoRA baseline with the same scaling and comparable regularization is needed to establish the benefit of the proposed design.

  2. The results are sensitive to shuffling on ImageNet-A. Enabling shuffling reduces final accuracy from 64.44% to 58.66% and average accuracy from 74.97% to 67.21% (Tables 8 and 10). These changes are much larger than those of CL-LoRA and SD-LoRA. Although AO-LoRA still has the highest final accuracy, this sensitivity reduces confidence in the consistency of its gains and requires further explanation.

Questions for authors
  1. What geometric property does Eq. (9) preserve, given that the weighted residual sum is generally not orthogonal?

  2. Under matched training settings and parameter budgets, does AO-LoRA still outperform LoRA with the same learnable scaling and comparable orthogonality regularization?

5 comments

We would like to express our sincere gratitude to Reviewer jynw for the positive comments on the strengths of our work and for providing us with the valuable opportunity to further clarify its contributions. Below, we address each point raised by the reviewer and describe the corresponding revisions made to the manuscript, which are highlighted in blue. We hope that our explanations can address the Reviewer's concerns, and we sincerely welcome any further comments or suggestions.

(Weakness 1 and Question 1 & 2) Geometric Properties and the Contributions of AO-LoRA

We sincerely thank the reviewer for raising this important concern regarding the underlying geometric properties and the contributions of AO-LoRA. We address the concern from the following three aspects:

  • The geometric properties targeted by Eq. (9).
  • The sources of performance gains in the ablation study (Table 4).
  • A controlled comparison with LoRA under matched experimental settings.

(1) Geometric properties and design motivation of Eq. (9)

We agree that the scaled residual accumulation in Eq. (9) does not generally constitute an orthogonal transformation. This design choice is motivated by the key consideration: Enforcing strict orthogonality transformation constraints may limit the model’s ability to adapt to new tasks.

This observation is supported by our preliminary experimental results, as reported in Table 1 (OFT-based). These results suggest that strict orthogonality constraints may impose excessive restrictions on parameter updates in continual learning scenarios, thereby hindering the model’s adaptability to new tasks.

Table 1. Comparison between LoRA-based methods (InfLoRA, SD-LoRA, CL-LoRA), Strict Orthogonal Transformation (OFT-based), and Ours (AO-LoRA) on ImageNet-R with 20 tasks. All results were obtained using the same random seed.

MethodA‾20\overline{A}_{20} (↑\uparrow)Avg. (↑\uparrow)Stability (↑\uparrow)Plasticity (↑\uparrow)Forgetting (↓\downarrow)
InfLoRA70.7776.4275.1177.177.58
SD-LoRA75.9381.1680.7981.836.60
CL-LoRA76.0782.5181.7583.659.26
OFT-based62.0868.5967.8368.226.46
AO-LoRA (Ours)78.9283.3682.3483.935.75

Therefore, compared with LoRA-based methods, Eq. (9) is designed to mitigate arbitrary parameter changes induced by unconstrained additive updates. To preserve sufficient plasticity, AO-LoRA does not enforce strict orthogonality between the accumulated weights and the pretrained weights. Instead, it uses orthogonality as a structural constraint on task-specific update directions. This design helps mitigate the degradation of pretrained representations caused by unconstrained updates, while preserving sufficient flexibility for adaptation to new tasks.

(2) Explanation of improvement in the ablation study

To further disentangle the contributions of different components in AO-LoRA, we extend the ablation study in Table 2 (Numbered as Table 4 in the revised manuscript) by additionally reporting Stability, Plasticity, and Forgetting. These metrics provide further insights into the sources of performance improvement.

Table 2. Ablation study on ImageNet-R with 20 tasks. The first row shows the LoRA baseline. OFT denotes the results obtained by applying orthogonal fine-tuning alone.

OFTA-Orth.Scal.I-Orth.A‾20\overline{A}_{20}Avg.StabilityPlasticityForgetting (↓\downarrow)
−-−-−-−-74.1381.1180.2481.708.15
✓\checkmark62.0868.5967.8368.226.46
✓\checkmark✓\checkmark74.5382.8282.2181.087.00
✓\checkmark✓\checkmark✓\checkmark77.52‾\underline{77.52}83.9382.4983.37‾\underline{83.37}6.36‾\underline{6.36}
✓\checkmark✓\checkmark✓\checkmark✓\checkmark78.8583.37‾\underline{83.37}82.34‾\underline{82.34}83.935.75

Although anchored orthogonal adaptation (A-Orth.) achieves a final-task accuracy (A‾20\overline{A}_{20}) of 74.53% comparable to the 74.13% achieved by the LoRA baseline, it exhibits different stability and plasticity. Specifically, A-Orth. considerably improves model stability, suggesting that its structured low-rank updates provide a more stable adaptation mechanism than conventional LoRA updates. Furthermore, incorporating the task-specific scaling factor (Scal.) introduces additional flexibility in controlling the magnitude of task-specific updates, thereby improving the model's ability to adapt to new tasks and enhancing its plasticity.

(Weakness 1 and Question 1 & 2)

(3) Controlled comparison with LoRA under matched experimental settings

Following the reviewer’s valuable suggestion, we introduce Scaled LoRA and LoRA with I-Orth. as controlled baselines to investigate the contributions of task-specific scaling and inter-task orthogonality, respectively, while retaining the standard additive low-rank update formulation of LoRA.

Scaled LoRA: Scaled LoRA adopts the same task-specific bounded scaling mechanism as AO-LoRA. Specifically, its task-specific update is defined as:

ΔWScaled-LoRA(t)=αtB(t)A(t)\Delta W_{\mathrm{Scaled\text{-}LoRA}}^{(t)} = \alpha_t B^{(t)}A^{(t)}

where αt\alpha_t is a learnable, task-specific scaling factor with the same parameterization and initialization as in AO-LoRA.

The results are summarized in Table 3. The results suggest that the effect of scaling depends on the underlying update mechanism. For the LoRA baseline, the scaling factor modulates the unconstrained low-rank update determined by the factors A(t)A^{(t)} and B(t)B^{(t)}. In particular, it does not impose additional structural constraints on the update directions of LoRA. In contrast, for A-Orth., the scaling factor modulates the contribution of a structurally parameterized update. It provides an additional task-specific degree of freedom for controlling the strength of the orthogonally induced adaptation.

Table 3. Comparison of LoRA baseline (LoRA), LoRA with scaling (Scaled LoRA), Anchored Orthogonal Adaptation (A-Orth.), and A-Orth. with scaling (Scaled A-Orth.) on ImageNet-R with 20 tasks.

MethodUpdate mechanismA‾20\overline{A}_{20}Avg.StabilityPlasticityForgetting (↓\downarrow)
LoRAB(t)A(t)B^{(t)}A^{(t)}74.1381.1180.2481.708.15
Scaled-LoRAαtB(t)A(t)\alpha_t B^{(t)}A^{(t)}72.6580.3579.4480.448.48
A-Orth.Ht(W)−W\mathcal{H}^{t} (W) - W74.5382.8282.2181.087.00
Scaled A-Orth.αt(Ht(W)−W)\alpha_t (\mathcal{H}^{t} (W) - W)77.5283.9382.4983.376.36

LoRA with I-Orth.: Regarding inter-task orthogonality, AO-LoRA constrains the overlap between the low-rank directions projected through the pretrained weight matrix. Since this constraint is specifically designed for the Householder-based parameterization, it cannot be directly applied to the LoRA baseline. Therefore, when constructing the inter-task orthogonality regularizer for LoRA, we follow the design of O-LoRA [1]. Specifically, we define the subspace spanned by the columns of A(t)A^{(t)} as the update subspace of task tt: Ut=span⁡(A(t))\mathcal{U}_t = \operatorname{span}(A^{(t)}).

To encourage the update subspace of the current task to be orthogonal to those of previous tasks, we impose the following constraint:

A(i)⊤A(t)=0,i<tA^{(i)\top}A^{(t)}=0,\quad i<t

Accordingly, we define the inter-task orthogonality loss as:

\left|\left|A^{(i)\top}A^{(t)}\right|\right|_F^2$$ The final training objective is therefore formulated as: $$L^{t} = L_{ce}^{(t)} + \lambda L_{orth}^{(t)}$$ where $\lambda$ controls the strength of the inter-task orthogonality regularization. The results are summarized in **Table 4**. **Inter-task orthogonality is also effective when applied to the LoRA baseline**, suggesting that encouraging task-specific update subspaces to have minimal overlap is an effective way to mitigate forgetting in continual learning. **Table 4.** Comparison of LoRA baseline (LoRA), LoRA with inter-task orthogonality (LoRA with I-Orth.), Anchored Orthogonal Adaptation (A-Orth.), and A-Orth. with inter-task orthogonality (A-Orth. with I-Orth.) on ImageNet-R with 20 tasks. Method | $\overline{A}_{20}$ | Avg. | Stability | Plasticity | Forgetting ($\downarrow$) -------|-----------|-----------|-----------|-----------|----------- LoRA | 74.13 | 81.11 | 80.24 | 81.70 | 8.15 LoRA with I-Orth. | **75.47** | 82.37 | 82.31 | **82.01** | 7.52 A-Orth. | 74.53 | 82.82 | 82.21 | 81.08 | 7.00 A-Orth. with I-Orth. | 75.21 | **83.10** | **82.49** | 81.75 | **6.22** [1] Orthogonal subspace learning for language model continual learning, EMNLP 2023. --- We sincerely thank the reviewer for this insightful question. In response, we have - **1)** Revised the explanation **(Section 3.2: Line 294-298)** to make this distinction explicit. - **2)** Further discussed the contributions of AO-LoRA in the extended ablation study **(Section 4.3: Line 463-467 and Table 4)**. - **3)** Included controlled evaluations of scaling and regularization using LoRA baselines in both the main text and the appendix **(Section 4.3: Line 481-485, Section D.7: Line 1322-1379)**. We sincerely appreciate the reviewer’s opportunity to clarify these concerns.

(Weakness 2) Performance on ImageNet-A under class shuffling

We sincerely thank the reviewer for helping us identify this issue. We acknowledge that AO-LoRA exhibits sensitivity to class shuffling on ImageNet-A.

To better understand this phenomenon, we revisited the training logs and conducted additional analyses across 5 random seeds under both shuffling settings. As shown in Figure 1. We observe that, without shuffling, the class order and task composition remain fixed across different seeds, resulting in relatively consistent training behavior. In contrast, enabling shuffling changes the class composition of each incremental task, potentially leading to substantial variations in task difficulty.

Under the shuffled setting, the model tends to maintain relatively high training accuracy while exhibiting lower test accuracy. We conjecture that this behavior may be partly attributed to the limited and non-uniform number of samples per class in ImageNet-A, which may make the model more susceptible to overfitting under certain task compositions.

Figure 1. Training and test performance under shuffle is True or False across 5 random seeds on ImageNet-A.

We referenced the experimental configurations provided in the official code of related works [1-3] for ImageNet-A, we noticed that these methods use different learning rates and numbers of training epochs. Motivated by this observation, we tried different training parameter settings to see if they could alleviate the problem shown in the figure. Our results show that, although a performance gap remains, appropriately reducing the number of training epochs can partially mitigate the performance degradation of AO-LoRA under the shuffle=True setting, as shown in Figure 2.

Figure 2. Training and test performance with different number of training epochs on ImageNet-A.

[1] CL-LoRA: Continual low-rank adaptation for rehearsal-free class-incremental learning, CVPR 2025.

[2] Sd-lora: Scalable decoupled low-rank adaptation for class incremental learning, ICLR 2025.

[3] Revisiting weight regularization for low-rank continual learning, ICLR 2026.


We thank the reviewer again for raising this important question, which provides valuable insights and directions for our future research and we will pay closer attention to this issue. We have clarified this limitation and incorporated the additional analysis into the revised manuscript (Section D.1: Line 1095-1100).

Reviewer_ZZWFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes AO-LoRA for parameter-efficient continual learning. It accumulates scaled Householder-based updates anchored to pretrained weights and applies an inter-task orthogonality regularizer. Experiments show improved accuracy on several benchmarks and better tolerance to large learning rates.

Critical strengths
  • The method achieves meaningful gains across several datasets and pretrained backbones with relatively few trainable parameters per task.
  • The held-out linear probes provide useful evidence about how continual adaptation affects transferability beyond the training tasks.
Critical weaknesses
  • Although each Householder transformation is orthogonal, the scaled residual sum in Equation (9) generally is not. It therefore does not automatically preserve feature norms or angles. This leaves the claimed connection between orthogonality, representation preservation, and training stability insufficiently supported.
  • In Table 3, anchored orthogonal adaptation improves over LoRA by only 0.40 points; larger gains appear after adding scaling and regularization. A LoRA baseline with comparable scaling and regularization is needed to establish the value of the proposed parameterization.
Questions for authors
  • What geometric property does Equation (9) preserve, and what evidence connects that property to reduced forgetting?
  • Does the advantage remain when LoRA uses comparable scaling, regularization, and adaptation locations?
  • Holding out 20 ImageNet-R classes and using 20 classes per task implies nine training tasks. Why does Figure 1(a) show ten?
  • What explains the drop on ImageNet-A when shuffling is enabled, particularly the average accuracy falling from 74.97% to 67.21%?
6 comments

We would like to sincerely thank the Reviewer ZZWF for the time spent on our work and for the encouraging comments. We greatly appreciate the opportunity to address the concerns and questions raised. Below, we address each point raised by the reviewer and describe the corresponding revisions made to the manuscript, which are highlighted in blue. We hope that our explanations can address the Reviewer's concerns, and we sincerely welcome any further comments or suggestions.

(Weakness 1 and Question 1) Clarification of Eq.(9) and the connection to forgetting mitigation

We thank the reviewer for raising this important point. This comment provides us with a valuable opportunity to clarify the properties of AO-LoRA and their relationship to forgetting mitigation. We would like to further clarify the design motivation behind AO-LoRA and the key properties it aims to achieve.

(1) Geometric properties and design motivation of Eq. (9)

We agree that the scaled residual accumulation in Eq. (9) does not generally constitute an orthogonal transformation. This design choice is motivated by our preliminary experimental results: Enforcing strict orthogonality transformation constraints may limit the model’s ability to adapt to new tasks.

As reported in Table 1. These results suggest that strict orthogonality constraints may impose excessive restrictions on parameter updates in continual learning scenarios, thereby hindering the model’s adaptability to new tasks.

Table 1. Comparison between LoRA-based methods, Strict Orthogonal Transformation (OFT-based), and Ours (AO-LoRA) on ImageNet-R with 20 tasks. All results were obtained using the same random seed.

MethodA‾20\overline{A}_{20} (↑\uparrow)Avg. (↑\uparrow)Stability (↑\uparrow)Plasticity (↑\uparrow)Forgetting (↓\downarrow)
InfLoRA70.7776.4275.1177.177.58
SD-LoRA75.9381.1680.7981.836.60
CL-LoRA76.0782.5181.7583.659.26
OFT-based62.0868.5967.8368.226.46
AO-LoRA (Ours)78.9283.3682.3483.935.75

Therefore, compared with LoRA-based methods, Eq. (9) is designed to mitigate arbitrary parameter changes induced by unconstrained additive updates. To preserve sufficient plasticity, AO-LoRA does not enforce strict orthogonality between the accumulated weights and the pretrained weights. Instead, it uses orthogonality as a structural constraint on task-specific update directions. This design helps mitigate the degradation of pretrained representations caused by unconstrained updates, while preserving sufficient flexibility for adaptation to new tasks.

(2) Connection to forgetting mitigation

To better understand which components contribute to reducing forgetting, we further report the Stability, Plasticity, and Forgetting metrics for the ablation experiments in Table 2 (Numbered as Table 4 in the revised manuscript).

Table 2. Ablation study on ImageNet-R with 20 tasks. The first row shows the LoRA baseline. OFT denotes the results obtained by applying orthogonal fine-tuning alone.

OFTA-Orth.Scal.I-Orth.A‾20\overline{A}_{20}Avg.StabilityPlasticityForgetting (↓\downarrow)
−-−-−-−-74.1381.1180.2481.708.15
✓\checkmark62.0868.5967.8368.226.46
✓\checkmark✓\checkmark74.5382.8282.2181.087.00
✓\checkmark✓\checkmark✓\checkmark77.52‾\underline{77.52}83.9382.4983.37‾\underline{83.37}6.36‾\underline{6.36}
✓\checkmark✓\checkmark✓\checkmark✓\checkmark78.8583.37‾\underline{83.37}82.34‾\underline{82.34}83.935.75

As shown in the table, imposing strict orthogonality (OFT) reduces forgetting, but leads to relatively poor performance across the other metrics. With the introduction of anchored residual accumulation, the model's ability to learn new tasks improves substantially. Although its Forgetting score increases slightly compared with OFT, it remains considerably better than those of LoRA. Another key contributor to reducing forgetting is inter-task orthogonality (I-Orth.). By encouraging reduced overlap between task-specific directions, this regularizer helps mitigate potential interference among different tasks, thereby contributing to reduced forgetting.


We thank the reviewer for helping us identify this distinction, which has enabled us to better present and clarify our contributions. We have revised the explanation (Section 3.2: Line 294-298) to make this distinction explicit. We have also extended the ablation study and analysis (Section 4.3: Line 459-461 and Table 4). We believe these clarifications provide a more precise interpretation of AO-LoRA.

(Weakness 2 and Question 2) Explanation of Table 3 and controlled comparison with LoRA

We sincerely thank the reviewer for giving us this valuable opportunity to clarify this point. This insightful observation has helped us better disentangle the contributions of individual components and provide a clearer understanding of their respective roles in the overall performance improvements.

To avoid any confusion, we would like to clarify that Table 3 in the original manuscript is numbered as Table 4 in the revised manuscript.

(1) Explanation of the contribution of different components in Table 3

For the reviewer’s convenience, we reproduce the ablation table below, including the additional evaluation metrics of Stability, Plasticity, and Forgetting. As discussed in our response to Weakness 1, these metrics provide further insights into the contributions of different components to the overall performance improvements.

Table 3. Ablation study on ImageNet-R with 20 tasks. The first row shows the LoRA baseline. OFT denotes the results obtained by applying orthogonal fine-tuning alone.

OFTA-Orth.Scal.I-Orth.A‾20\overline{A}_{20}Avg.StabilityPlasticityForgetting (↓\downarrow)
−-−-−-−-74.1381.1180.2481.708.15
✓\checkmark62.0868.5967.8368.226.46
✓\checkmark✓\checkmark74.5382.8282.2181.087.00
✓\checkmark✓\checkmark✓\checkmark77.52‾\underline{77.52}83.9382.4983.37‾\underline{83.37}6.36‾\underline{6.36}
✓\checkmark✓\checkmark✓\checkmark✓\checkmark78.8583.37‾\underline{83.37}82.34‾\underline{82.34}83.935.75

Although anchored orthogonal adaptation (A-Orth.) achieves a final-task accuracy (A‾20\overline{A}_{20}) of 74.53% comparable to the 74.13% achieved by the LoRA baseline, it exhibits different stability and plasticity. Specifically, A-Orth. considerably improves model stability, suggesting that its structured low-rank updates provide a more stable adaptation mechanism than conventional LoRA updates. Furthermore, incorporating the task-specific scaling factor (Scal.) introduces additional flexibility in controlling the magnitude of task-specific updates, thereby improving the model's ability to adapt to new tasks and enhancing its plasticity.

(Weakness 2 and Question 2)

(2) Controlled comparison with LoRA under matched experimental settings

Following the reviewer’s suggestion, we introduce Scaled LoRA and LoRA with I-Orth. as controlled baselines to investigate the contributions of task-specific scaling and inter-task orthogonality, respectively, while retaining the standard additive low-rank update formulation of LoRA.

Scaled LoRA: Scaled LoRA adopts the same task-specific bounded scaling mechanism as AO-LoRA. Specifically, its task-specific update is defined as:

ΔWScaled-LoRA(t)=αtB(t)A(t)\Delta W_{\mathrm{Scaled\text{-}LoRA}}^{(t)} = \alpha_t B^{(t)}A^{(t)}

where αt\alpha_t is a learnable, task-specific scaling factor with the same parameterization and initialization as in AO-LoRA.

The results are summarized in Table 4. The results suggest that the effect of scaling depends on the underlying update mechanism. For the LoRA baseline, the scaling factor modulates the unconstrained low-rank update determined by the factors A(t)A^{(t)} and B(t)B^{(t)}. In particular, it does not impose additional structural constraints on the update directions of LoRA. In contrast, for A-Orth., the scaling factor modulates the contribution of a structurally parameterized update. It provides an additional task-specific degree of freedom for controlling the strength of the orthogonally induced adaptation.

Table 4. Comparison of LoRA baseline (LoRA), LoRA with scaling (Scaled LoRA), Anchored Orthogonal Adaptation (A-Orth.), and A-Orth. with scaling (Scaled A-Orth.) on ImageNet-R with 20 tasks.

MethodUpdate mechanismA‾20\overline{A}_{20}Avg.StabilityPlasticityForgetting (↓\downarrow)
LoRAB(t)A(t)B^{(t)}A^{(t)}74.1381.1180.2481.70‾\underline{81.70}8.15
Scaled LoRAαtB(t)A(t)\alpha_t B^{(t)}A^{(t)}72.6580.3579.4480.448.48
A-Orth.Ht(W)−W\mathcal{H}^{t} (W) - W74.53‾\underline{74.53}82.82‾\underline{82.82}82.21‾\underline{82.21}81.087.00‾\underline{7.00}
Scaled A-Orth.αt(Ht(W)−W)\alpha_t (\mathcal{H}^{t} (W) - W)77.5283.9382.4983.376.36

LoRA with I-Orth.: Regarding inter-task orthogonality, AO-LoRA constrains the overlap between the low-rank directions projected through the pretrained weight matrix. Since this constraint is specifically designed for the Householder-based parameterization, it cannot be directly applied to the LoRA baseline. Therefore, when constructing the inter-task orthogonality regularizer for LoRA, we follow the design of O-LoRA [1]. Specifically, we define the subspace spanned by the columns of A(t)A^{(t)} as the update subspace of task tt: Ut=span⁡(A(t))\mathcal{U}_t = \operatorname{span}(A^{(t)}).

To encourage the update subspace of the current task to be orthogonal to those of previous tasks, we impose the following constraint:

A(i)⊤A(t)=0,i<tA^{(i)\top}A^{(t)}=0,\quad i<t

Accordingly, we define the inter-task orthogonality loss as:

\left|\left|A^{(i)\top}A^{(t)}\right|\right|_F^2$$ The final training objective is therefore formulated as: $$L^{t} = L_{ce}^{(t)} + \lambda L_{orth}^{(t)}$$ where $\lambda$ controls the strength of the inter-task orthogonality regularization. The results are summarized in **Table 5**. **Inter-task orthogonality is also effective when applied to the LoRA baseline**, suggesting that encouraging task-specific update subspaces to have minimal overlap is an effective way to mitigate forgetting in continual learning. **Table 5.** Comparison of LoRA baseline (LoRA), LoRA with inter-task orthogonality (LoRA with I-Orth.), Anchored Orthogonal Adaptation (A-Orth.), and A-Orth. with inter-task orthogonality (A-Orth. with I-Orth.) on ImageNet-R with 20 tasks. Method | $\overline{A}_{20}$ | Avg. | Stability | Plasticity | Forgetting ($\downarrow$) -------|-----------|-----------|-----------|-----------|----------- LoRA | 74.13 | 81.11 | 80.24 | 81.70 | 8.15 LoRA with I-Orth. | **75.47** | 82.37 | 82.31 | **82.01** | 7.52 A-Orth. | 74.53 | 82.82 | 82.21 | 81.08 | 7.00 A-Orth. with I-Orth. | 75.21 | **83.10** | **82.49** | 81.75 | **6.22** [1] Orthogonal subspace learning for language model continual learning, EMNLP 2023. --- We sincerely thank the reviewer for this insightful and technically valuable question. In response, we have further discussed the contributions of AO-LoRA in the revised manuscript **(Section 4.3: Line 463-467 and Table 4)**. We have also included controlled evaluations of scaling and regularization using LoRA baselines in both the main text and the appendix **(Section 4.3: Line 481-485, Section D.7: Line 1322-1379)**. These additional analyses have provided valuable insights that have helped us further improve the quality of the manuscript.
Reviewer_pnu8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AO-LoRA for parameter-efficient continual learning. It anchors task-specific Householder updates to pretrained weights and uses learnable scaling and inter-task orthogonality to reduce interference. Results across vision and language benchmarks are competitive.

Critical strengths
  1. The paper addresses an important problem: preserving pretrained knowledge during continual adaptation.
  2. The method is simple and achieves strong results across several datasets and pretrained backbones.
  3. The ablations and representation analyses provide useful insights into its behavior.
Critical weaknesses
  1. The geometric explanation is incomplete. Individual Householder transformations preserve norms and angles, but their scaled residual sum is generally not orthogonal. Moreover, orthogonality before a linear projection does not guarantee preservation of output feature geometry. The paper should clarify which properties the complete method actually preserves.

  2. The source of improvement remains unclear. The small, bounded scaling factor may explain part of the performance and learning-rate robustness. LoRA with the same scaling mechanism is needed as a controlled baseline.

  3. Representation analysis is mainly correlational. Higher cosine similarity between class prototypes does not necessarily imply poorer classification. Quantitative measures of class separation would better support the proposed explanation.

  4. Experimental details need clarification. The relationship between the training classifier and prototype cla4ssifier is unclear. Some baseline results also differ substantially between the main tables and Table 13.

  5. Efficiency claims need fuller accounting. New adapters are stored for every task, so cumulative storage grows with task count. Reporting total storage and training time would complement the trainable-parameter comparison.

Questions for authors
  1. Does AO-LoRA remain more stable when LoRA uses the same bounded scaling?
  2. How are old-class prototypes handled as representations change?
  3. What causes the baseline discrepancies between the main results and Table 13?
4 comments

We sincerely thank the Reviewer pnu8 for taking the time to carefully evaluate our work and provide such valuable feedback and constructive suggestions. We greatly appreciate the opportunity to revise our work in response to these comments. Below, we address each point raised by the reviewer and describe the corresponding revisions made to the manuscript, which are highlighted in blue. We hope that our explanations can address the Reviewer's concerns, and we sincerely welcome any further comments or suggestions.

(Weakness 1) Clarification of the design and properties of AO-LoRA

We thank the reviewer for highlighting this important distinction. This provides valuable insights for clarifying the properties and key design choice of AO-LoRA. We agree that the scaled residual sum in AO-LoRA does not keep the same geometric structure as the pretrained weight matrix. This design choice was motivated by our experimental observations. In the following, we provide further clarification:

(1) Design motivation of AO-LoRA

Compared with LoRA-based methods, AO-LoRA is designed to mitigate arbitrary parameter changes induced by unconstrained additive updates. Compared with OFT-based methods, AO-LoRA aims to provide parameter plasticity to support continual adaptation to new tasks.

AO-LoRA does not impose a strict orthogonality constraint between the final continual learning model and the pretrained model. This design choice is motivated by our preliminary experimental results: Enforcing strict orthogonality transformation constraints may limit the model’s ability to adapt to new tasks.

As reported in Table 1, these results suggest that strict orthogonality constraints may impose excessive restrictions on parameter updates in continual learning scenarios, thereby hindering the model’s adaptability to new tasks.

Table 1. Comparison between LoRA-based methods, Strict Orthogonal Transformation (OFT-based), and Ours on ImageNet-R with 20 tasks. All results were obtained using the same random seed.

MethodA‾20\overline{A}_{20} (↑\uparrow)Avg. (↑\uparrow)Stability (↑\uparrow)Plasticity (↑\uparrow)Forgetting (↓\downarrow)
InfLoRA70.7776.4275.1177.177.58
SD-LoRA75.9381.1680.7981.836.60
CL-LoRA76.0782.5181.7583.659.26
OFT-based62.0868.5967.8368.226.46
AO-LoRA (Ours)78.9283.3682.3483.935.75

Therefore, to preserve sufficient plasticity, AO-LoRA does not enforce strict orthogonality between the accumulated weights and the pretrained weights. Instead, it uses orthogonality as a structural constraint on task-specific update directions. This design helps mitigate the degradation of pretrained representations caused by unconstrained updates, while preserving sufficient flexibility for adaptation to new tasks.

(2) Design mechanisms and their properties in AO-LoRA

Based on the design motivation discussed above, we would like to further clarify the properties that AO-LoRA aims to achieve. For individual task updates, Householder reflection provides an explicit structured low-rank update to mitigate changes from unconstrained updates. In continual learning, anchored residual accumulation integrates these task-specific updates in a common reference space to integrate knowledge from different tasks, and inter-task orthogonality encourages reduced overlap between task-specific directions in the output space. The three mechanisms together mitigate unnecessary disruption to pretrained representations structure while retaining the flexibility required for continual adaptation.

To be more specific:

  1. Householder Orthogonal Transformation: For individual task updates, the weight update can be written as ΔW=WH−W=−2Wuu⊤\Delta W=WH-W=-2Wuu^{\top}. Thus, AO-LoRA constructs a structured low-rank update induced by orthogonal transformations. This property provides an explicit geometric constraint on task-specific adaptation in the input space.

  2. Anchored Residual Accumulation: Each task-specific residual ΔW(t)\Delta W^{(t)} is defined with respect to the same pretrained weight WW, and the residuals are then accumulated through learnable scaling coefficients in a common reference space. This provides flexibility in integrating knowledge from different tasks.

  3. Inter-task Orthogonality: For the projected directions, the regularizer encourages reduced overlap between the projected low-rank directions Vr(t)=WUr(t)V_{r}^{(t)} = WU_{r}^{(t)} associated with different tasks. This helps mitigate potential task interference.


We appreciate this comment, which helps us make the claims more precise. We have revised the explanation (Section 3.2: Line 294-298) to make this distinction explicit. We believe these clarifications offer a more clearer interpretation of AO-LoRA and better characterize its preservation properties.

(Weakness 2 and Question 1) Controlled LoRA comparison and explanation of improvement

We thank the reviewer for raising this important concern. This observation helps us better disentangle the contribution of the scaling mechanism from that of the proposed structured update mechanism.

(1) Controlled LoRA comparison

Following the reviewer’s suggestion, we introduce Scaled LoRA, a controlled baseline that retains the standard additive low-rank update of LoRA while adopting the same task-specific bounded scaling mechanism as AO-LoRA. The results are summarized in Table 2, while the corresponding learning rate sensitivity is illustrated in Figure 1. Specifically, its task-specific update is defined as: ΔWScaled-LoRA(t)=αtB(t)A(t),\Delta W_{\mathrm{Scaled\text{-}LoRA}}^{(t)} = \alpha_t B^{(t)}A^{(t)},

Table 2. Comparison of LoRA baseline (LoRA), LoRA with scaling (Scaled LoRA), Anchored Orthogonal Adaptation (A-Orth.), and A-Orth. with scaling (Scaled A-Orth.) on ImageNet-R with 20 tasks.

MethodUpdate mechanismA‾20\overline{A}_{20}Avg.StabilityPlasticityForgetting (↓\downarrow)
LoRAB(t)A(t)B^{(t)}A^{(t)}74.1381.1180.2481.70‾\underline{81.70}8.15
Scaled LoRAαtB(t)A(t)\alpha_t B^{(t)}A^{(t)}72.6580.3579.4480.448.48
A-Orth.Ht(W)−W\mathcal{H}^{t} (W) - W74.53‾\underline{74.53}82.82‾\underline{82.82}82.21‾\underline{82.21}81.087.00‾\underline{7.00}
Scaled A-Orth.αt(Ht(W)−W)\alpha_t (\mathcal{H}^{t} (W) - W)77.5283.9382.4983.376.36

Figure 1. Learning rate sensitivity of LoRA, Scaled LoRA, A-Orth., and Scaled A-Orth.

The results suggest that the effect of scaling depends on the underlying update mechanism. For the LoRA baseline, the scaling factor modulates the unconstrained low-rank update determined by the factors A(t)A^{(t)} and B(t)B^{(t)}. In particular, it does not impose additional structural constraints on the update directions of LoRA. In contrast, for A-Orth., the scaling factor modulates the contribution of a structurally parameterized update. It provides an additional task-specific degree of freedom for controlling the strength of the orthogonally induced adaptation. As shown in Figure 1, A-Orth. exhibits substantially greater learning rate robustness than LoRA, indicating that the robustness of AO-LoRA cannot be attributed solely to the bounded scaling mechanism.

(2) Explanation of improvement

To further disentangle the contributions of different components in AO-LoRA, we extend the ablation study by additionally reporting Stability, Plasticity, and Forgetting. These metrics provide further insights into the sources of performance improvement. The results are reported in Table 3 (Numbered as Table 4 in the revised manuscript).

Table 3. Ablation study on ImageNet-R with 20 tasks. The first row shows the LoRA baseline. OFT denotes the results obtained by applying orthogonal fine-tuning alone.

OFTA-Orth.Scal.I-Orth.A‾20\overline{A}_{20}Avg.StabilityPlasticityForgetting (↓\downarrow)
−-−-−-−-74.1381.1180.2481.708.15
✓\checkmark62.0868.5967.8368.226.46
✓\checkmark✓\checkmark74.5382.8282.2181.087.00
✓\checkmark✓\checkmark✓\checkmark77.52‾\underline{77.52}83.9382.4983.37‾\underline{83.37}6.36‾\underline{6.36}
✓\checkmark✓\checkmark✓\checkmark✓\checkmark78.8583.37‾\underline{83.37}82.34‾\underline{82.34}83.935.75

Although anchored orthogonal adaptation (A-Orth.) achieves a final-task accuracy (A‾20\overline{A}_{20}) of 74.53% comparable to the 74.13% achieved by the LoRA baseline, it exhibits different stability and plasticity. Specifically, A-Orth. considerably improves model stability, suggesting that its structured low-rank updates provide a more stable adaptation mechanism than conventional LoRA updates. Furthermore, incorporating the task-specific scaling factor (Scal.) introduces additional flexibility in controlling the magnitude of task-specific updates, thereby improving the model's ability to adapt to new tasks and enhancing its plasticity.


We appreciate this comment. In response, we have extended the ablation study in the revised manuscript (Section 4.3: Line 463-467 and Table 4) by additionally reporting Stability, Plasticity, and Forgetting. We have also included the results and analysis of the controlled LoRA baseline experiments in both main text and appendix (Section 4.3: Line 481-485, Section D.7: Line 1322-1379). We believe these revisions provide a clearer characterization of the effectiveness and contributions of the proposed components.

(Weakness 3) Additional quantitative evidence of class separability

We thank the reviewer for this insightful comment. To address this limitation, we provide additional quantitative evidence by measuring class separability and evaluating the resulting classification performance, as detailed below.

(1) Quantitative measurement of class separation

We employ the Fisher discriminant ratio to jointly consider inter-class scatter and intra-class scatter. JFisher=Tr⁡(Sinter)Tr⁡(Sintra)+ϵ,J_{\mathrm{Fisher}}=\frac{\operatorname{Tr}(S_{\text{inter}})}{\operatorname{Tr}(S_{\text{intra}})+\epsilon}, where SinterS_{\mathrm{inter}} and SintraS_{\mathrm{intra}} denote the inter-class and intra-class scatter matrices, respectively. ϵ\epsilon is a small constant for numerical stability. A higher Fisher ratio indicates better class separability, as it reflects greater inter-class dispersion relative to intra-class variation. SinterS_{\mathrm{inter}} and SintraS_{\mathrm{intra}} are defined as: Sinter=∑c=1CNc(μc−μ)(μc−μ)⊤,S_{\mathrm{inter}} = \sum_{c=1}^{C}N_c (\boldsymbol{\mu}_c-\boldsymbol{\mu}) (\boldsymbol{\mu}_c-\boldsymbol{\mu})^\top,

(\mathbf{z}_i-\boldsymbol{\mu}_c) (\mathbf{z}_i-\boldsymbol{\mu}_c)^\top,$$ where $C$ is the number of classes, $N_c$ is the number of samples in class $c$, $\mathbf{z}_i$ denotes the feature representation of the $i$-th sample, $\boldsymbol{\mu}_c$ is the mean feature representation of class $c$, and $\boldsymbol{\mu}$ is the global mean feature representation across all evaluation samples. We compute this metric on the same held-out classes used in the motivation experiments presented in the original Introduction. For a more interpretable comparison of changes during continual learning, we normalize the Fisher ratio by its value before continual learning: $J_{\mathrm{Fisher}}^{(t)} / J_{\mathrm{Fisher}}^{(0)}$ (the superscript 0 denotes the pretrained model). The resulting curves are presented in **Figure 2**. [**Figure 2.** Normalized Fisher discriminant ratio on ImageNet-R over 10 continual learning tasks, with 20 classes introduced per task.](https://postimg.cc/bdwYz5GJ) As shown in Figure 2, compared with LoRA, AO-LoRA maintains a relatively stable normalized Fisher ratio throughout continual learning. LoRA exhibits an overall decreasing trend, dropping to 96.39% by the final task. This provides additional quantitative evidence supporting the effectiveness of AO-LoRA in mitigating representation degradation and preserving class separability. ## (2) Complementary evidence from task-agnostic inference **Table 15 in the Appendix** provides complementary empirical evidence regarding the discriminative capability of learned representations. As shown in the table, AO-LoRA achieves the highest CIL accuracy (80.94%) and Task-ID accuracy (84.90%) among all compared methods. Moreover, it yields the smallest task-incremental and class-incremental performance gap (9.86%). This indicates that AO-LoRA better preserves both class-discrimination and task-discrimination capabilities, leading to more effective task-agnostic inference. **Table 15.** Task-agnostic inference decomposition on reproduced ImageNet-R (N=10). CIL Acc. is final task-agnostic accuracy ($\overline{A}_{10}$), Task-ID Acc. measures whether the predicted class belongs to the correct task, and TIL Acc. uses the oracle task ID. Methods | CIL Acc. | Task-ID Acc. | TIL Acc. | **TIL-CIL Gap** ($\downarrow$) -------|-----------|-----------|-----------|----------- InfLoRA | 75.04 | 78.86 | 87.63 | 12.59 SD-LoRA | 71.00 | 74.99 | 85.24 | 14.24 EWC-LoRA | 72.68 | 76.03 | 87.07 | 14.38 CL-LoRA | $\underline{79.55}$ | $\underline{82.58}$ | **91.03** | $\underline{11.48}$ AO-LoRA (Ours) | **80.94** | **84.90** | $\underline{90.80}$ | **9.86** --- We sincerely thank the reviewer again for their valuable feedback. We have incorporated the above analyses and corresponding results into the Appendix **(Section D.6: Line 1276-1315)** to provide a more comprehensive assessment of class separability and discriminative capability.

(Weakness 4 and Question 3) Clarification of experimental details and baseline results

We thank the reviewer for pointing out the ambiguity in our experimental description. In the revised manuscript, we have clarified the relationship between the training classifier and the prototype classifier (Section C.2: Line 919-923). We have also highlighted explanations in the Main text (Section 4.2: Line 394-395) and Appendix (Section D.3: Line 1181-1183) to clarify the different sources of the reported results.

(1) Relationship between the training and prototype classifiers

During the training stage of each task, we use a proxy classifier to perform classification for the current task. Specifically, it is implemented as a cosine linear classifier, which computes the cosine similarity between the input feature zz and the class weight wcw_c to obtain the classification score: sc=z⊤wc∣∣z∣∣2∣∣wc∣∣2s_c = \frac{z^{\top} w_c}{|| z ||_2 || w_c ||_2} After completing the training of the current task, we discard the proxy classifier and extract features from the current task's training samples using the learned model. We then compute the class prototypes and combine them with the prototypes from all previously learned tasks to construct a global prototype-based cosine classifier for final inference.

(2) Discrepancies between Table 1 and Table 13

The discrepancies between Table 1 and Table 13 (Numbered as Table 15 in the revised manuscript) arise from the different sources of the reported results. In Table 1, we directly adopt the results reported in the original papers for performance comparison. In Table 15, we reproduce all baselines because this analysis requires access to model logits, as explained in Section D.3 of the manuscript (Line 1181-1183).

Therefore, the results of InfLoRA, SD-LoRA, and EWC-LoRA differ between the two tables because Table 1 uses the results reported in their original papers, while Table 13 uses our reproduced results. For CL-LoRA, the results in both tables are reproduced by us, as the original paper does not report results under this experimental setting.

(Weakness 5) Total storage and training time of AO-LoRA

We sincerely appreciate the reviewer’s valuable comment. This constructive suggestion has been particularly helpful in improving the completeness of our evaluation. Following your suggestion, we have added the resource consumption results of our method to Section 4.2 (Line 438-442).

AO-LoRA requires storing the task-specific adapter, scaling factor (α\alpha), and prototype classifier for each task. To provide a more comprehensive assessment of its efficiency, we report the average training time per task along with its standard deviation, as well as the GPU memory consumption.

In Table 4 (Table 3 in the revised manuscript), we compare LoRA-based methods and AO-LoRA on CIFAR-100 and ImageNet-R, both with 10 tasks. For a fair comparison, all methods are evaluated under the same hardware and training settings. The results show that AO-LoRA incurs moderate computational overhead. Notably, its training time on ImageNet-R is the lowest among the compared methods, demonstrating its favorable computational efficiency.

Table 4. Comparison of different methods in terms of average training time per task and GPU memory consumption.

MethodsCIFAR-100 (N=10)ImageNet-R (N=10)Peak Memory
Vanilla LoRA10m37s ±\pm 12s13m21s ±\pm 1m37s∼\sim 18 GB
InfLoRA12m17s ±\pm 16s14m53s ±\pm 1m33s∼\sim 20 GB
SD-LoRA13m4s ±\pm 53s20m46s ±\pm 3m29s∼\sim 37 GB
CL-LoRA12m56s ±\pm 40s14m57s ±\pm 2m8s∼\sim 28 GB
EWC-LoRA11m19s ±\pm 21s13m32s ±\pm 1m30s∼\sim 24 GB
AO-LoRA (Ours)11m10s ±\pm 15s11m7s ±\pm 1m23s∼\sim 20 GB

(Question 2) Clarification of old-class prototype

We thank the reviewer for giving us the opportunity to clarify this point. In our implementation, class prototypes are computed when their corresponding classes are first introduced and stored in the prototype classifier. As new tasks arrive and the feature representations evolve, we keep the prototypes of previously learned classes fixed. This design avoids revisiting old data to recompute their prototypes. We have clarified this strategy in the revised manuscript (Section C.2: Line 923-924).

Comments on the paper

2 comments

Dear Area Chair and Reviewers,

We would like to sincerely thank you once again for the generous time, patience, and insightful feedback you have devoted to our paper during this busy discussion period. Your guidance has helped us substantially improve the clarity and rigor of our manuscript.

Following our point-by-point responses, we have incorporated all experiments, clarifications, and discussions into the newly uploaded revised manuscript, with all changes highlighted in blue. To facilitate your review, we summarize below the main concerns raised by the reviewers and our corresponding responses:

  1. Clarification of the design and properties of AO-LoRA (for Reviewers pnu8, ZZWF, and jwnw)
  • To preserve sufficient plasticity, AO-LoRA does not impose strict orthogonality between the accumulated weights and the pretrained weights. Instead, it employs orthogonality as a structural constraint on task-specific update directions. This design mitigates the degradation of pretrained representations caused by unconstrained updates while preserving the flexibility needed to adapt to new tasks.
  • We have revised the explanation in Section 3.2 (Lines 294-298) to explicitly clarify this distinction.
  1. Explanation of the improvement in the ablation study (for Reviewers pnu8, ZZWF, jwnw)
  • Although anchored orthogonal adaptation (A-Orth.) achieves final accuracy comparable to that of the LoRA baseline, it exhibits greater stability on previously learned tasks. This suggests that its structured low-rank updates provide more stable adaptation than conventional LoRA updates. Furthermore, incorporating the scaling factor (Scal.) introduces additional flexibility in controlling the magnitude of task-specific updates, thereby improving the model's ability to adapt to new tasks and enhancing its plasticity.
  • We have reported the Stability, Plasticity, and Forgetting metrics in Section 4.3 and Table 4 (Lines 463-467).
  1. Controlled LoRA comparison (for Reviewers pnu8, ZZWF, jwnw)
  • The effectiveness of scaling depends on the underlying update mechanism. Directly applying scaling to the LoRA baseline leads to a decrease in overall performance. Meanwhile, inter-task orthogonality also proves effective when applied to the LoRA baseline, suggesting that encouraging task-specific update subspaces to have minimal overlap is an effective way to mitigate forgetting in continual learning.
  • We have included the results and analysis of the controlled LoRA baseline experiments in both the main text and the appendix (Section 4.3: Lines 481-485; Section D.7: Lines 1322-1379).
  1. Quantitative evidence of class separability (for Reviewers pnu8)
  • We employ the Fisher discriminant ratio, which jointly considers inter-class scatter and intra-class scatter, to quantitatively evaluate class separability. We have incorporated the corresponding analysis into the appendix (Section D.6: Lines 1276-1315) to provide a more comprehensive assessment of class separability and discriminative capability.
  1. Total storage requirements and training time of AO-LoRA (for Reviewers pnu8)
  • We have added the resource consumption results of AO-LoRA to Section 4.2 (Line 438-442).
  1. Performance on ImageNet-A under class shuffling (for Reviewers ZZWF, jwnw)
  • We have conducted additional experiments to investigate the potential reasons behind AO-LoRA's sensitivity to class shuffling on ImageNet-A. The reviewers' insights have provided valuable directions for our future research. We have also clarified this limitation and incorporated the corresponding analysis into the revised manuscript (Section D.1: Lines 1095-1100).

Once again, we sincerely appreciate your time, constructive feedback, and valuable suggestions. We would truly appreciate it if you might kindly take these updates into account in your final assessment. Please let us know if anything remains unclear or if further analysis would be helpful and we would be more than happy to address it promptly!

Thank you for your consideration.

Best regards,

The Authors

Dear Reviewer,

Thank you once again for your time and constructive feedback on our paper. We have uploaded the revised version of the manuscript (with all changes highlighted in blue), along with detailed response and additional experimental results. As the discussion period is approaching its end, please let us know if any points require further clarification. We would highly appreciate your feedback and would be grateful if you could reconsider your assessment in light of the updated paper.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.