acceptodds
Under review as a conference paper at ICLR 2027

Do Pretrained Spectral Subspaces Matter For Adaptation?

Abstract

Do accuracy gains require updates along principal pretrained directions, and what does interaction between spectral regions change? We use fixed-band adapters to separate spectral location from within-band flexibility, then study leading–tail interaction during training and at trained checkpoints. Under fixed recipes, leading, middle, tail and random supports all improve accuracy on RTE and CommonsenseQA. These supports are not interchangeable: tail gives the lowest decoder negative log-likelihood (NLL) within each adapter family. On encoder tasks, an interaction penalty sharply reduces cross-band energy share relative to a roughly norm-matched control. It lowers uncalibrated NLL without a resolved accuracy gain; temperature scaling removes most of the NLL advantage. In a separate decoder study, removing learned cross-region components raises choice NLL relative to shrinking each matrix's update to the same remaining norm. Thus, accuracy gains do not require principal support, while the predictive effects of cross-region interaction depend on the metric and the intervention.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Parameter-efficient fine-tuning, spectral adaptation, singular value decomposition, low-rank adaptation, update geometry, pretrained representations, model adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QgxGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors investigate whether adaptation under parameter-efficient fine-tuning (PEFT) must occur along specific spectral directions by evaluating leading, middle, tail, and random spectral supports, both with and without rotation.

To examine the role of interactions between leading and tail directions, they also compare training with interaction penalty regularization, no regularization, and norm-matched regularization.

Finally, based on the results of removing cross-region components from the learned parameter updates, they conclude that cross-region structure plays an important role in adaptation.

Critical strengths

The authors identify a limitation of prior methods such as PiSSA and MiLoRA, which set the adaptation directions only at initialization, and design experiments that can confine the adaptation directions throughout training.

This allows them to conduct controlled experiments examining the importance of principal directions during the training process.

Critical weaknesses

1. The analytical contribution is limited.

The authors explicitly acknowledge that several results are statistically unresolved, limiting the conclusions that can be drawn about the importance of spectral directions or their interactions.

Furthermore, several findings are difficult to interpret, such as the discrepancy between the trends in NLL and accuracy, and the observation that MIX reduces NLL but loses this advantage after temperature scaling.

As a result, the analysis does not seem to provide readers with sufficiently clear intuition about the underlying mechanisms.

2. The writing lacks clarity and polish.

Several terms are used without proper definitions, such as coefficient/rotated adapter in the Introduction, references to encoder and decoder without specifying the models or tasks being considered, and UNREG/MIX/NORM appearing in Section 3 before being defined.

Additionally, terms such as support, band, and region do not appear to be used consistently, and the transitions between sentences are often awkward.

The motivations behind the experiments and the interpretations of their results also do not seem to be presented in a sufficiently coherent and well-structured manner.

Questions for authors

1. Could you provide results on more recent model families, such as Qwen3, and a broader range of tasks, even if the model sizes remain similar to those currently evaluated?

The current selection of model families and tasks seems insufficient to support the conclusion that principal directions are not necessary for successful adaptation.

2. The interaction experiments rely on regularization as an indirect way of controlling interactions, rather than explicitly preventing the interaction components from being learned.

Would it not be possible to directly disable these components by formulating the parameter update as Δ=U(KLLKLTKTLKTT)V⊤\Delta = U \begin{pmatrix} K_{LL} & K_{LT} \\\\ K_{TL} & K_{TT} \end{pmatrix} V^\top and fixing the cross-region blocks KLTK_{LT} and KTLK_{TL} to zero during training?

I would like to understand why the interaction experiments need to retain the original formulation based on SLS_L and HH.

3. In Section 5.2, you compare completely removing the cross-region components after training with uniformly shrinking the parameter update to match its norm.

However, uniform shrinking preserves the direction of the parameter update, whereas removing the cross-region components can substantially alter its direction, potentially making the comparison unfavorable to cross-region removal.

A fairer comparison might be to compare removing the cross-region components with removing an equal amount of norm from the CLLC_{LL} and CTTC_{TT} components.

Could you provide results from such an experiment?

2 comments

W1. Analytical contribution and mechanisms

With every band tuned, all bands improve accuracy and none has a measurable advantage, while LoRA, which learns its directions, reaches lower NLL than every tuned band (general comment, "Adapter tuning based on learning rates" and "Narrower supports"; Appendix C.3).

We agree the NLL trends in the submission were hard to interpret. They do not persist with tuned learning rates and early stopping, and the revision no longer claims them (Appendices C.1 and C.2).

On the mechanism, tuning appears to compensate for spectral scale: tail coefficients grow to 0.8–6.4 times their singular values, against 0.07–0.15 for leading ones, at comparable accuracy (Appendix C.5).

W2. Clarity

Agreed. The Introduction now defines band, support, the C and R adapters and UNREG/MIX/NORM. The revision uses band in place of region throughout, names models and tasks in the main text instead of "encoder" and "decoder", and leads Section 5.1 with the tuned results (Table 1).

Q1. Qwen3 and more tasks

Yes. We added Qwen3-1.7B on CommonsenseQA and a second task set (ARC-Easy, ARC-Challenge and OpenBookQA) on Qwen2.5-3B (general comment, "Adapter tuning based on learning rates"; Table 1). In both, every band improves accuracy and none has a measurable advantage, so we find no evidence that principal directions are required.

Q2. Why not fix the cross blocks to zero?

The suggested form makes every block a free matrix: even with the cross blocks fixed to zero it trains k² + r² parameters per matrix, about 1.3 million on Qwen2.5-1.5B (d = 1536, r = 512). Ours trains r + 2d, about 3.6 thousand (Eq. 7, Appendix D.1).

We designed the adapter to stay linear in the layer width while still allowing limited interaction. Its diagonal scalers create that interaction, which separates exactly into cross-band blocks for analysis (Eq. 10). These blocks are not parameters, so they cannot be fixed to zero without removing the scalers (Appendix A.7), and we control them with a penalty (Section 3.2).

Q3. Removing an equal amount from the within-band blocks

We ran this control (Section 5.2, Table 3). Removing the cross blocks raises NLL 1.6–3.2 times as much as removing the same norm from C_LL and C_TT, while accuracy differs by less than 0.6 points. The trained model thus makes some use of the cross-band components, but accuracy does not depend on them, as the random-partition control also shows (general comment, "Further control: leading–tail interaction").

Thank you for the detailed response. After carefully considering the rebuttal, I will maintain my original score.

Reviewer_255POpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether parameter-efficient adaptation needs to operate along particular spectral directions of pretrained weight matrices. The authors introduce fixed-support coefficient and rotated adapters to separate spectral location from within-band flexibility, and compare leading, middle, tail, and random supports. They further study interactions between leading and tail spectral regions using an interaction regularizer and post-hoc checkpoint interventions. Experiments on RTE, MRPC, CommonsenseQA, and GSM8K suggest that successful adaptation does not require principal pretrained directions, while the effects of spectral location and cross-region interaction depend on the task, metric, and intervention.

Critical strengths
  1. The paper makes a useful conceptual distinction between spectral location and adaptation flexibility. Existing spectral PEFT methods can simultaneously change initialization, support, and parameterization. By explicitly confining updates to fixed pretrained subspaces and separately varying within-band rotations, this work provides a cleaner experimental framework for studying these factors. The approximation analysis in Proposition 1 also gives a clear interpretation of the two sources of restriction.

  2. The experimental reporting is unusually transparent. The paper reports unresolved comparisons, failed controls, post-hoc studies, calibration issues, and computational limitations rather than hiding them. This substantially improves the interpretability of the empirical evidence.

Critical weaknesses
  1. The experiments do not yet establish a sufficiently general conclusion about the importance of pretrained spectral subspaces for adaptation. The main location study uses relatively wide fixed supports and one common training recipe for the different spectral regions. For example, the selected bands contain 256 dimensions for RoBERTa and 512 dimensions for the decoder models, and the authors explicitly note that different spectral initializations may prefer different learning rates. Showing that several large fixed subspaces can all improve over a frozen model demonstrates that principal directions are not strictly necessary, but it does not show that spectral location is unimportant in the low-capacity regime that motivates PEFT. A rank/support-width sweep, particularly at much narrower supports with separately tuned hyperparameters, seems necessary before drawing broader conclusions about pretrained spectral subspaces.

  2. The cross-region interaction experiments do not cleanly isolate the effect of interaction itself. The central MIX–NORM comparison attempts to distinguish reduced cross-region interaction from simply reducing update magnitude, but the size matching is only approximate on the encoder experiments and fails substantially on CommonsenseQA. In the latter case, NORM has an update norm about 27% larger than MIX. The paper itself acknowledges that these comparisons therefore do not isolate interaction at equal update size, while the checkpoint interventions are post hoc and do not control for rank or distance from the trained update. Since the accuracy differences are also largely unresolved, I find the evidence for the predictive role of cross-region interaction suggestive rather than conclusive.

Questions for authors
  1. How sensitive are the spectral-location conclusions to the available support width? Would leading, middle, tail, and random supports remain similarly capable when the support is reduced to substantially smaller fractions of the weight dimension, closer to commonly used LoRA ranks? A support-width sweep could materially affect my interpretation of the claim that principal pretrained directions are not required.

  2. Can the authors provide a more tightly matched control for the interaction experiments? In particular, can MIX and NORM be matched more closely in achieved update norm, especially on CommonsenseQA, or can cross-region energy be manipulated directly while controlling total update magnitude? This would help determine whether the observed prediction differences arise from interaction structure rather than changes in update size or allocation.

1 comment

W1 and Q1. Tuned rates and narrow supports

We now tune the learning rate of every band, and we ran the support-width sweep: 4, 8, 16 and 64 of 1,536 directions per matrix, tuned for each band and width (general comment, "Adapter tuning based on learning rates" and "Narrower supports"; Table 1, Figure 2, Appendix C.3). At every width, all bands improve accuracy and none has a measurable advantage, so the conclusion also holds in the low-capacity regime. What matters more is learning the directions: LoRA at ranks 4 and 8 is 5.6–5.9 points above the best band of equal width, with about 3,000 times more parameters.

W2 and Q2. Isolating interaction from update size

Thank you for this point. The revision adds both controls you suggested.

  • MIX and NORM at matched size. On Qwen2.5-1.5B (CommonsenseQA), NORM is now within 5% of MIX's update size. MIX has a cross-band share of 32.5% against 41.4%, and accuracy does not differ measurably (general comment, "Further control: leading–tail interaction"; Table 2, Appendix B.3).
  • Direct manipulation at matched distance, size and rank. Removing the cross blocks of a random basis partition matches cross-band removal in all three. Cross-band removal raises NLL 1.3–2.4 times as much, while accuracy differs by less than 0.6 points (same section of the general comment; Section 5.2, Table 3).
  • RoBERTa-base. At the training endpoint, MIX and NORM have similar update sizes, MIX's cross-band share is a fifth (RTE) and a tenth (MRPC) of NORM's, and accuracy does not differ measurably (Table 2). At the early-stopping checkpoint the share is still much lower (16% against 44% on RTE, 5% against 43% on MRPC), but on RTE the update sizes no longer match. We therefore limit the size-matched claim on RoBERTa-base to the training endpoint (Section 5.2, Appendix C.2).
Reviewer_Sq8BOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper analyzes how fine-tuning different spectral regions, or the interactions between different regions, affects performance. The first region separation is done by confining updates to portions of directions based on orderings in terms of singular values (leading, middle, tail or random bands) of the weight matrices in the attention modules. This is done either by a diagonal rescaling of these updates, either by a confined rotation. Experiments use RoBERTa-base on RTE, and Qwen2.5-1.5B, Qwen2.5-3B and Llama-3.1-8B on CommonsenseQA, plus GSM8K on Qwen2.5-1.5B.

To study interactions instead, a separate adapter linking leading and tail directions is trained with an interaction penalty, a norm-matched size penalty, or no penalty. Cross-region components are also removed from trained random-frame and LoRA checkpoints and compared against norm-matched uniform shrinking.

All bands improve accuracy on RTE and CommonsenseQA, concluding no principal directions are required, with tail bands leading to the lowest decoder NLL. For the encoders, the interaction penalty sharply reduces cross-region energy without an accuracy gain. Removing cross-region components raises NLL more than norm-matched shrinking. On CommonsenseQA, the reduction is modest and NLL increases.

Critical strengths

The attempt to separate the factors behind adaptation is interesting, and differs from what is typically shown with PEFT methods, where changes apply altogether, mixing different factors, where no single factor stands out.

Critical weaknesses
  1. Main factors of variations are confounded. First, using the same learning rate for all bands, adapter forms and model sizes. This is critical when affecting leading vs tailing singular values: an additive update leads to similar absolute values variations, but that is a much stronger relative change for tail singular values than it is for leading ones. The paper’s MRPC gate (Appendix B.2) shows its adapters are learning-rate sensitive: 0.05 failed and 0.01 worked. That gate concerns the interaction adapter, but between-band differences may simply reflect learning-rate fit rather than location. Section 7 acknowledges the shared rate, but the abstract’s band comparisons rely on it: a per-band sweep is needed. Secondly, zeroing the leading–tail cross blocks and shrinking the whole update to the same norm are not matched perturbations: with cross-energy share p, removal moves the weights √p ‖Δ‖ while shrinking moves them (1 − √(1−p)) ‖Δ‖. For the reported cross shares (~22% for Llama, ~44% for Qwen2.5-1.5B, Appendix B), removal is a 2.6–4× larger perturbation, therefore expected to hurt performance more, making necessary a distance-matched control.

  2. Paper's generality is weak, and results do not support a solid conclusion. The two large decoder models are only evaluated on CommonsenseQA without explaining this choice or adding more tasks. On this task, patterns are consistent across models. Elsewhere only the decoder NLL pattern holds. On GSM8K no configuration improves performance over the frozen models, with rotations lowering accuracy. On RTE the tail band has the lowest accuracy and highest NLL in both adapter families, but no band contrast is resolved. As noted in point 1, the current setup also cannot distinguish a location effect from an artifact of the untuned shared learning rate. Beyond showing that non-principal bands can improve accuracy, the experiments do not support any solid conclusion on band ranking or interaction. An analysis of convergence speed (as claimed by PiSSA) or forgetting (as claimed by MiLoRA) could have made the study more informative.

Questions for authors
  1. How much does the spectrum change under the confined adapters? Meaning what share does each band hold before and after training. In how many matrices do trained tail singular values overtake middle or leading ones?
  2. Can the authors explain how were the shared learning rates selected? Does the band ordering survive tuning?
  3. How long have the encoder models been trained? Adapted models reach held-out NLL of 1.7-2.3 against 0.68 for head-only training (Table 1), which suggests overfitting. How would validation-based early stopping affect these?
  4. The frozen Qwen2.5-3B reference has choice NLL 2.9985 at 77% accuracy (Table 1), worse than uniform guessing over five choices. What prompt format has been used?
3 comments

W1 and Q2. Learning rate

We now tune the learning rate of every band (general comment, "Adapter tuning based on learning rates"). Every band improves accuracy, and no difference between bands is significant. The band ordering does not survive tuning, which supports the paper's thesis that no pretrained band has a measurable advantage. The original shared rates are in Appendix C.1.

W1. Distance-matched removal

Thank you for the distinction between matching remaining size and matching distance. We added controls matched in both (Section 4, Table 3). At equal distance, cross-band removal raises NLL about half as much as a uniform shrink, and 1.3–2.4 times as much as removing a random part of the update matched in both, with accuracy differences below 0.6 points. We find no evidence that cross-band components play a special role (general comment, "Further control: leading–tail interaction").

W2. Generality

We added a model, a second task set and narrower supports, all at tuned rates, and no band has a measurable advantage in any of them (general comment, "Adapter tuning based on learning rates", "Narrower supports" and "Scope and conclusiveness").

  • Convergence and forgetting, as suggested (Appendix C.4). At tuned rates no band converges measurably faster, and at matched forgetting no band's accuracy gain differs from the leading band's, although 5 of 15 NLL contrasts favor a non-leading band.
  • RTE. With early stopping, no band-versus-tail accuracy difference is significant (Appendix C.2).
  • GSM8K. Rank-8 LoRA (one seed) also falls below the frozen model (Appendix B.10), so this task cannot separate supports.

Q1. Spectrum change

The table gives the tail band's share of each matrix's spectral energy before and after training at tuned rates, and the number of matrices in which a trained tail singular value grows past a middle-band value or into the leading range (coefficient adapters, ranges over seeds; Appendix C.5).

ModelTail share beforeTail share afterTail passes a middle valueTail enters the leading range
Qwen2.5-1.5B0.99%1.68%56 of 56 matrices2–5 of 56 matrices
Qwen2.5-3B0.46%3.20%72 of 7248–51 of 72
Llama-3.1-8B0.053%2.96%64 of 6423–24 of 64

The tail moves the most in relative terms: its singular values change by a median 64–555%, against 7–15% for leading ones, and the leading and middle energy shares rise by at most 1.5 and 3.4 points. Bands therefore fix the directions of an update, not its final singular values, and the tail's larger movement brings no measurable advantage.

Q3. RoBERTa training length

Yes, the endpoint was overtrained. RTE ran 90 epochs and MRPC 30, while loss on the selection split is lowest after 3–18 epochs. Re-evaluating every run at that checkpoint (Appendix C.2), RTE NLL falls from 1.68–2.25 to 0.52–0.62, below head-only training (0.680), and accuracy stays 9–15 points above it. The high NLL was overconfidence from long training, not a failure to adapt. MIX's NLL advantage over NORM also disappears with early stopping, and the revision no longer claims it.

Q4. Frozen Qwen2.5-3B NLL

The prompt is zero-shot through the model's chat template: a system instruction to reply with the option letter only, then the question with options A–E (Appendix B.8). The five letters carry 99.99% of the next-token probability, so the format is not the cause. The high NLL comes from confident errors: the 23% of questions answered wrongly give 99.7% of it, with a median probability of 10⁻⁶ on the correct letter. One temperature fitted on the selection split (T = 6.4) brings it to 0.690 (Appendix C.19), which is why Table 1 reports temperature-scaled NLL.

Thank you for the detailed rebuttal and new experiments, which address my main soundness concerns, and I have raised my soundness score accordingly. However, the paper’s findings are now mostly negative and of unclear practical utility, so I will maintain my overall score.

Thank you for raising the soundness score. We would like to highlight the paper's contribution and practical impact.

Several recent PEFT methods are built on the pretrained spectrum, choosing where to adapt by spectral position. Our novel, controlled adapter design tests this common assumption directly: we find no measurable advantage for any pretrained subspace, at any support width, and no evidence that accuracy depends on the interaction between subspaces (Table 1, Figure 2, Table 3). Task-informed directions, by contrast, do better (Appendix C.3).

We believe this challenges a widely held design principle, shows practitioners where gains are, and are not, to be found, and can guide the design of more effective adaptation methods.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: results of the new experiments

Overall, the new experiments strengthen the paper's main claim: with tuned adapters, we find no measurable advantage for any pretrained spectral subspace. The leading–tail interaction experiments are a further control, with no evidence that the interaction is needed. Reviewer-specific questions are answered in the individual threads.

Adapter tuning based on learning rates

We tune the learning rate (LR) of each band independently, using coefficient (C) and rotated (R) adapters with 512 directions per band. LRs are selected on the selection split from a grid extended until bracketed, then evaluated once on the test set over three seeds (Llama-3.1-8B: two; Appendix C.1). Test accuracy (%), mean ± SE, on CommonsenseQA (CSQA) and on ARC-Easy, ARC-Challenge and OpenBookQA (ARC + OBQA) pooled:

Model, adapter, taskFrozenLeadingMiddleTailRandom
Qwen2.5-1.5B, C, CSQA70.679.1 ± 0.279.1 ± 0.479.2 ± 0.378.7 ± 0.4
Qwen2.5-1.5B, R, CSQA70.679.7 ± 0.179.7 ± 0.579.4 ± 0.379.9 ± 0.4
Qwen2.5-3B, C, CSQA77.083.2 ± 0.384.2 ± 0.183.5 ± 0.283.6 ± 0.05
Qwen2.5-3B, R, CSQA77.083.4 ± 0.183.8 ± 0.283.3 ± 0.583.9 ± 0.3
Llama-3.1-8B, C, CSQA77.083.3 ± 0.383.2 ± 0.183.2 ± 0.0483.7 ± 0.1
Qwen3-1.7B, C, CSQA68.477.6 ± 0.476.6 ± 0.477.1 ± 0.377.2 ± 0.4
Qwen2.5-3B, C, ARC + OBQA86.290.1 ± 0.190.4 ± 0.190.3 ± 0.190.2 ± 0.1

Every band improves over the frozen model, and no band contrast is significant in accuracy (95% paired-bootstrap intervals, Appendix C.1). On ARC + OBQA, the largest test set, the difference between any two bands is bounded within ±1 point. With tuned learning rates, no pretrained spectral band has a measurable advantage.

Narrower supports

We repeat the experiment with 4, 8, 16 and 64 directions (Qwen2.5-1.5B, C, CSQA), tuning each band and width separately. As a reference, we train the same adapter on the top-r singular vectors of each matrix's loss gradient, computed over 256 training examples before training (Section 5.1, Appendix C.3). Test accuracy (%), mean ± SE:

WidthLeadingMiddleTailRandomGradient (reference)
474.1 ± 0.273.6 ± 0.275.0 ± 0.274.6 ± 0.276.7 ± 0.2
874.7 ± 0.374.9 ± 0.175.1 ± 0.274.8 ± 0.175.8 ± 0.9
1676.0 ± 0.175.2 ± 0.275.2 ± 0.0375.7 ± 0.176.9 ± 0.3
6477.5 ± 0.477.6 ± 0.277.1 ± 0.177.3 ± 0.377.8 ± 0.5

No two pretrained bands differ significantly at any width, in accuracy or NLL (Appendix C.3), and all improve over the frozen model (70.6%). The gradient reference does differ: it is ahead of the best band at width 4 (+1.7 points, 95% CI [0.0, 3.4]) and in NLL at widths 4–16, so the setup can detect directional effects. Band choice changes accuracy by at most 1.4 points, whereas LoRA, which learns its directions with about 3,000× more parameters, is 5.6–5.9 points above the best band at ranks 4 and 8 (Appendix C.3). Learning the directions matters more than choosing among pretrained bands.

Further control: leading–tail interaction

We find no evidence that accuracy depends on the leading–tail interaction.

  • During training, suppressing it leaves accuracy unchanged (MIX vs size-matched NORM: +0.05 points, 95% CI [−0.46, +0.57]).
  • After training, removing it changes accuracy by less than 0.6 points compared with removing a random or a within-band part of the update of the same size and distance. NLL rises 1.3–3.2 times as much, so the trained model makes some use of these components, but accuracy does not depend on them.

We thank the reviewers for distinguishing matching by remaining update size from matching by distance moved. Section 4 now defines both. Increase in raw choice NLL per edit (Table 3; intervals in Appendix C.7):

CheckpointCross removalRandom partitionWithin-band removalShrink, same distanceShrink, same size
Qwen2.5-1.5B, C / R+0.226 / +0.190+0.168 / +0.149+0.130 / +0.118+0.443 / +0.401+0.054 / +0.050
Qwen2.5-3B, C / R+0.490 / +0.396+0.308 / +0.247+0.225 / +0.178+0.889 / +0.813+0.097 / +0.072
Llama-3.1-8B, C / R+0.035 / +0.034+0.015 / +0.017+0.011 / +0.013+0.064 / +0.063+0.005 / +0.007

Removing the cross-band components costs modestly more than removing a matched random part of the update, and about half as much as a uniform shrink of the same distance.

Scope and conclusiveness

Across the tested models, tasks, adapter families and widths, no pretrained spectral subspace consistently outperforms the others, whereas task-gradient directions show measurable advantages in some settings. These results challenge the use of pretrained spectral position as a criterion for selecting PEFT subspaces, at least for the 1.5–8B models and multiple-choice tasks studied here.

The revised PDF is uploaded, with changes marked in blue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.