acceptodds
Under review as a conference paper at ICLR 2027

Narrowing the Aperture for the Tail: Controllable Spatial Evidence for Long-Tailed Recognition

Abstract

Multi-expert architectures have emerged as a prominent paradigm for long-tailed visual recognition by distributing specialized classifiers across the class frequency spectrum. However, current frameworks uniformly employ Global Average Pooling (GAP) for spatial feature aggregation, implicitly enforcing an identical spatial observation scale for all classes and experts. In contrast to frequent classes that often benefit from broader contextual integration, rare categories tend to rely more heavily on compact, discriminative local cues. To address this architectural bottleneck, we propose Spatial Moment, a framework that endows controllable multi-expert networks with preference-conditioned evidence apertures. By coupling the hypernetwork preference state with a generalized-mean pooling ladder, our method enables tail-oriented experts to adaptively contract their spatial apertures, concentrating aggregation weights onto local, high-magnitude evidence. To complement this spatial mechanism, we further introduce a bounded one-sided geometric offset derived from within-expert classifier radii to restore radial magnitude statistics canceled by cosine normalization, alongside a class-orthogonal directional regularizer that discourages probability mass from concentrating on competing classes. During inference, experts are aggregated via a parameter-free, decisiveness-aware consensus route that maintains stable ensemble behavior. Extensive evaluations on long-tailed benchmarks and test-prior agnostic scenarios demonstrate that our approach achieves competitive performance against strong multi-expert baselines. Component removals and spatial-flow analyses indicate that adaptive aperture modulation is the primary driver of performance gains, mitigating head-bias classification errors on minority-class samples.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-Tailed Learning, Vision-Language Models

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_HK1gOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Spatial Moment for long-tailed recognition. The main idea is to use different pooling orders for different experts. Head-focused experts use more global information, while tail-focused experts focus more on local features. The method also uses a logit correction term and a regularization loss to improve classification. At test time, the experts are combined based on how confident their predictions are. Experiments on several long-tailed datasets show that the method achieves competitive results and improves tail-class accuracy.

Critical strengths
  1. The paper presents an interesting idea of using different pooling orders for different experts in long-tailed recognition.
  2. The experiments cover several long-tailed datasets and show competitive results, especially for tail classes.
Critical weaknesses
  1. The core motivation is not well supported. The paper assumes that head classes rely more on global context, while tail classes rely more on local features. However, class frequency does not necessarily determine the spatial scale of useful features. The paper does not provide direct evidence to support this assumption.

  2. The link between high activations and useful features is unclear. GeM pooling focuses on regions with high activations. However, these regions may contain background noise or spurious features rather than useful information. The related work Shortcut-Resistant CAM Distillation for Long-Tailed Recognition (ICML 2026) has studied similar issues in long-tailed recognition. The paper does not discuss this work in enough detail or clearly explain how its method differs from existing methods in handling background bias and spurious correlations.

  3. The need for preference-conditioned modulation is unclear. Table 5 shows that a fixed GeM ladder (1, 2, 3) already achieves 53.4% overall accuracy, while the proposed modulation only increases it to 53.8%. This suggests that most of the gains come from using different pooling orders across experts, rather than from preference modulation. It is unclear whether the added complexity is necessary.

  4. Limited novelty and too many hand-tuned settings. The method mainly combines existing techniques, such as GeM pooling, cosine classifiers, logit calibration, margin losses, and confidence-based expert weighting. It also uses many manually chosen hyperparameters, including the modulation coefficient, Dirichlet concentration, radial correction strength, margin coefficient, regularization weight, and expert weighting ratio. The paper provides limited analysis of these choices. Overall, the method seems more like a combination of existing techniques than a new approach that offers deeper insights into long-tailed recognition.

Questions for authors

See weaknesses.

2 comments

Response to Reviewer HK1g

Thank you for the comments on motivation, shortcuts, and complexity. Our supplementary analyses focus on the architectural question: whether spatial aggregation can usefully participate in expert specialization before classification.

1. Frequency objectives and aggregation sensitivity

The design asks whether experts with different frequency objectives benefit from different aggregation sensitivities. This formulation accommodates varied object sizes and spatial layouts; frequency defines the learning objective rather than a fixed assumption about object geometry.

In a separate supplementary CIFAR-100-LT IF=100IF=100 evaluation—ResNet-32, 200 epochs, seed 1—we froze the model and compared its deployed orders with uniform GAP, uniform p=2p=2, uniform p=3p=3, fixed (1,2,3)(1,2,3), and reversed deployed orders.

The deployed configuration achieves 53.04% Overall / 32.80% Few. The five alternatives range from 50.06% to 52.77% Overall and 31.47% to 32.47% Few. Reversing the deployed orders gives 52.34% Overall / 32.03% Few.

These controls directly establish that aggregation and expert–order assignment affect recognition in the trained model. They support studying spatial aggregation as an architectural variable, with the comparison scoped to fixed-checkpoint interventions.

2. High activations and background shortcuts

Activation magnitude and foreground relevance are distinct properties. GeM provides explicit control over how strongly different activations contribute to the pooled representation, and the new controls show that this choice affects recognition.

SRCD addresses evidence selection through teacher maps constructed from head-proxy CAMs and explicit distillation. Spatial Moment operates through aggregation before classification, using no such CAM supervision.

This distinction clarifies the scope of our evidence: the measurements support control of activation concentration and its effect on recognition. Semantic localization and shortcut resistance remain separate properties from concentration itself.

3. Conditioning and deployment complexity

A fixed ladder implements the spatial hierarchy, while shared conditioning supplies bounded adjustments and connects pooling to classifier generation during training. At deployment, q⋆=softmax⁡(a)q^\star=\operatorname{softmax}(a) is fixed, so classifier weights and pooling orders can be precomputed and cached; consensus weighting remains sample-dependent. This fixed-parameter deployment form follows directly from the formulation.

4. Novelty and hyperparameters

GeM, cosine classifiers, and class-aware losses provide established building blocks. Our contribution is to introduce expert diversity at the spatial aggregation stage and study its effect through an explicit sensitivity mechanism and controlled comparisons. The pooling branch reuses the existing preference coordinates and adds no independent learnable pooling parameters.

The supplementary pooling comparisons hold all auxiliary coefficients fixed, directly testing aggregation changes without joint retuning. A matched-training-recipe comparison also gives 53.04% versus 52.04% Overall, and 32.80% versus 31.17% Few, for Spatial Moment and its reference baseline.

We position the auxiliary components as supporting design choices. The central contribution is the concrete architectural insight of spatial expert specialization, supported by the observed aggregation behavior and recognition benefits.

Sincerely,

The Authors

Thank you for your response. However, my main concerns remain. The new experiments only change pooling orders in a frozen model and do not fully support the core motivation. Also, high activations do not necessarily indicate useful features, and the need for preference modulation remains unclear given the strong results of the fixed GeM ladder. The concerns about limited novelty and many manually chosen hyperparameters are also not fully addressed. Therefore, I will keep my original score.

Reviewer_C5rtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spatial Moment, a multi-expert framework for long-tailed visual recognition that addresses the uniform use of global average pooling across experts. The authors argue that tail classes rely more heavily on localized discriminative cues, whereas head classes benefit from broader contextual information. To address this, the method introduces preference-conditioned generalized-mean pooling, assigning higher pooling orders to tail-oriented experts to emphasize local high-response features. Additionally, the framework incorporates geometry-derived logit calibration, rarity-conditioned margin and directional regularization, and decisiveness-aware expert aggregation. Extensive experiments validate the effectiveness of the proposed method.

Critical strengths
  • The method explores an underexplored perspective in long-tailed recognition. By introducing preference-conditioned spatial pooling, Spatial Moment enables different experts to focus on different levels of spatial evidence, particularly emphasizing local discriminative features for tail classes. This demonstrates a certain degree of novelty.
  • The experimental evaluation is comprehensive, covering major long-tailed recognition benchmarks such as ImageNet-LT, Places-LT, and iNaturalist, as well as scenarios with unknown test distributions. The proposed method achieves competitive performance across these settings, matching or outperforming most established baselines under comparable experimental protocols.
Critical weaknesses
  • The method is built on the assumption that tail classes benefit more from localized discriminative features, whereas head classes rely more on global contextual information. However, the paper provides limited direct empirical evidence supporting this distinction. It remains unclear whether this assumption consistently holds across different datasets and visual categories.

  • Several components of the proposed method, particularly the Geometry-Derived One-Sided Offset and Rarity-Conditioned Margin and Directional Regularization, appear somewhat heuristic. Although these designs are intuitively reasonable, the paper lacks sufficient theoretical justification for their specific formulations and does not clearly explain why these particular choices are preferred over alternative designs. Further explanations and empirical analyses would help strengthen the motivation and validity of these components.

Questions for authors
  • The proposed method is motivated by the assumption that tail classes rely more on localized discriminative features, while head classes benefit from broader contextual information. Could the authors provide more direct empirical evidence supporting this assumption?
  • The Geometry-Derived One-Sided Offset and Rarity-Conditioned Margin involve several specific design choices. Could the authors provide stronger theoretical or empirical justification for these formulations, particularly why they are preferable to conventional logit adjustment or existing class-aware margin losses?
2 comments

Response to Reviewer C5rt

Thank you for recognizing the value of the spatial aggregation perspective. The supplementary pooling comparison directly tests the architectural variable behind our motivation, while the radial intervention clarifies the contribution of the geometric correction.

These evaluations use existing checkpoints from a separate CIFAR-100-LT IF=100IF=100 configuration: ResNet-32, 200 epochs, seed 1, and FP32 evaluation on the complete test set.

1. Direct evidence for spatial aggregation

Our motivation is that experts trained with different frequency objectives may benefit from different sensitivities to spatial activation magnitudes. This concerns how evidence is aggregated and allows for objects with varied sizes and spatial layouts.

To test this variable directly, we held all learned parameters and auxiliary coefficients fixed, changed the pooling orders on the same feature maps, and recomputed consensus from the resulting raw logits.

Pooling configurationOverall (%)Few (%)
Deployed orders53.0432.80
Uniform GAP50.0631.47
Uniform p=2p=252.6632.47
Uniform p=3p=352.7731.63
Fixed (1,2,3)(1,2,3)52.4432.17
Reversed deployed orders52.3432.03

The deployed orders are approximately (1.378,2.063,3.059)(1.378,2.063,3.059). Their performance across these controls shows that both aggregation order and its assignment to experts matter to the trained model. These are fixed-checkpoint interventions.

The spatial measurements also confirm the intended concentration pattern. Relative to p=1p=1 on the same feature maps, the three deployed experts retain approximately 81.5%, 60.8%, and 42.8% of the mean effective spatial support. This follows the sensitivity relationship

=\left(\frac{F_i}{F_j}\right)^{p-1},$$ for positive activations above the numerical floor. Together, these results connect the design to measurable changes in activation concentration and show that recognition responds to the aggregation configuration. The support statistic characterizes concentration at the feature level. **2. Why this geometric offset?** The offset combines within-expert scale normalization, bounded magnitude, and one-sided correction. When numerical clipping is inactive, standardizing log norms removes a common multiplicative classifier scale. The $\tanh$ bounds the code, and the one-sided operation gives below-reference classifiers a nonnegative adjustment of at most $s\beta=0.15$. If classifier norms approximately follow a positive power law, $r_{e,c}\approx K_e n_c^{\nu_e}$, standardization gives an approximate log-frequency interpretation. This provides an interpretable connection between the geometric signal and frequency-based calibration under that assumption. Conventional logit adjustment can use known training frequencies, as in [Menon et al.](https://www.sanjivk.com/long_tail_learning_ICLR2021.pdf). Our correction uses expert-specific norm statistics already available from the classifiers. The new fixed-checkpoint ablation makes its practical role explicit: Overall/Few changes from **52.90%/31.40% to 53.04%/32.80%**, with the gain concentrated on Few classes and a tradeoff in Many and Medium accuracy. **3. Rarity margins and directional regularization** The margin $$m_{e,y}=\eta(e/E)\rho_y$$ provides a bounded target-class constraint that varies with normalized rarity and expert role while retaining a shared scale. The directional term $$D_e=1-p_{e,y}/\|\mathbf p_e\|_2$$ complements this constraint. At fixed target probability, reducing $D_e$ discourages the remaining probability mass from concentrating on a single rival. The two terms therefore act on target separation and the distribution of competing probabilities, respectively. Their controlled scales and explicit functional roles motivate these supporting choices, alongside the principal architectural contribution of expert-specific spatial aggregation. Sincerely, The Authors

Thank you to the authors for their detailed and thoughtful response. Most of my concerns have been adequately addressed, and I have decided to maintain my positive rating.

Reviewer_XX7dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Spatial Moment, a multi-expert framework for long-tailed recognition. Its main idea is to associate head-, medium-, and tail-oriented experts with progressively higher generalized-mean (GeM) pooling orders, coupled to the preference state of a classifier-generating hypernetwork. Additional components include a bounded classifier-norm-based logit offset, rarity-dependent margins, directional regularization, and decisiveness-weighted inference. Experiments cover four long-tailed benchmarks and eleven test-prior configurations on CIFAR-100-LT. The paper reports improved tail accuracy and competitive overall performance without iterative test-time optimization.

Critical strengths
  1. Zero-parameter, zero-FLOPs resolution of the spatial aggregation bottleneck: Prior multi-expert methods (e.g., RIDE, SADE, PRL) focus on losses, classifier heads, or test-time routing while overlooking how uniform GAP (p=1p=1) weights all spatial locations equally. Using the gradient sensitivity ratio ∂ge/∂Fi∂ge/∂Fj=(Fi/Fj)pe−1\frac{\partial g_e / \partial F_i}{\partial g_e / \partial F_j} = (F_i / F_j)^{p_e - 1}, the authors clearly show that assigning increasing pooling orders (p1<p2<p3p_1 < p_2 < p_3) lets the tail expert quadratically suppress diffuse background noise and focus on high-response local features—matching a standard 3-expert ResNet-32 in both parameters (3.85M3.85\text{M}) and compute (0.22 GFLOPs0.22\text{ GFLOPs}).
  2. Clear multi-dimensional diagnostic analyses: Section 4.2 effectively quantifies how the three experts progressively compress effective spatial support (Inverse Participation Ratio, Fig. 2) from >500>500 to <200<200, while Figs. 3 and 4 trace the evolution of class norm encodings γe,c\gamma_{e,c} and long-tailed margins across stages to clarify each module's physical role.
Critical weaknesses
  1. "Controllable modulation" degenerates into static constants: Despite the title and abstract emphasizing preference-coupled pooling (pe=be+0.5qe(e)p_e = b_e + 0.5 q_e^{(e)}), at inference (δ(e)=0\delta^{(e)}=0) the preference distribution q(e)=softmax(a)q^{(e)} = \text{softmax}(a) collapses to a fixed near-uniform (1/3,1/3,1/3)(1/3, 1/3, 1/3), fixing pooling orders at (1.17,2.17,3.17)(1.17, 2.17, 3.17) for all inputs. Table 5 exposes this directly: removing hypernetwork modulation entirely and using a static integer ladder p=(1,2,3)p=(1,2,3) already achieves 53.4%53.4\% Overall / 33.8%33.8\% Few-shot, whereas adding the full preference modulation yields only +0.4%+0.4\% Overall.
  2. Artificial simplex collapse from softmax(a+Dirichlet)\text{softmax}(a + \text{Dirichlet}): In Eq. (1),

δ(e)∼Dirichlet(0.5×1E)\delta^{(e)} \sim \text{Dirichlet}(0.5 \times \mathbf{1}_E)

already lies on the probability simplex ΔE−1\Delta^{E-1} (δi∈[0,1]\delta_i \in [0,1], ∑δi=1\sum \delta_i = 1). While baseline PRL uses Dirichlet vectors directly as simplex coordinates, this paper wraps δ(e)∈[0,1]3\delta^{(e)} \in [0,1]^3 in an unscaled softmax. Mathematically, passing [0,1]3[0,1]^3 vectors through a standard softmax caps the maximum possible output entry at ≤e1e1+2e0≈0.576\le \frac{e^1}{e^1 + 2e^0} \approx 0.576, making the simplex vertices unreachable. This directly explains why PRL spans the simplex vertices in Fig. 6 (Neff≈1.2N_{\text{eff}} \approx 1.2) while the proposed method is squeezed into a small central triangle (Neff≈2.7N_{\text{eff}} \approx 2.7); likewise, the "low variance and zero flip rate under Dirichlet perturbations" in Fig. 5 is an artifact of compressing the input perturbation range via a redundant softmax rather than learning a more robust feature manifold.

  1. Low module coupling and incremental ablations: Beyond GeM pooling, the method stacks radial offset Rrad\mathcal{R}_{\text{rad}},

composite rarity loss Lrar\mathcal{L}_{\text{rar}}

(itself combining 4 separate heuristics: rarity margin η=0.035\eta=0.035, cosine probability regularization λd=0.1\lambda_d=0.1, dynamic branch weights ωe\omega_e, and stop-gradient scaling κ=2\kappa=2), and consensus routing Ccon\mathcal{C}_{\text{con}}. In Table 4, these three modules add only +0.5%+0.5\%, +0.3%+0.3\%, and +0.3%+0.3\% Overall on CIFAR-100-LT (just 30–50 test images), and all experiments use a single seed (seed=1, App. F) without mean/std across runs.

  1. Minimal gains over SOTA baselines: Overall accuracy merely ties PRL on ImageNet-LT (60.8%60.8\%), slightly trails 3-expert ICL on iNaturalist 2018 (75.8%75.8\% vs. 75.9%75.9\%), and edges out PRL by just +0.2%+0.2\% on CIFAR-100-LT at IF=50IF=50 (57.5%57.5\% vs. 57.3%57.3\%) and IF=10IF=10 (65.8%65.8\% vs. 65.6%65.6\%).
Questions for authors
  1. Mathematical compression of softmax(a+Dirichlet)\text{softmax}(a + \text{Dirichlet}) and necessity of the hypernetwork (Eq. 1 & Figs. 5–6): Since δ(e)∼Dirichlet(0.5×1E)\delta^{(e)} \sim \text{Dirichlet}(0.5 \times \mathbf{1}_E)

already resides on the probability simplex, why apply an additional softmax to a+δ(e)a + \delta^{(e)} in Eq. (1), which mathematically restricts q(e)q^{(e)} to ≤0.58\le 0.58 and prevents reaching simplex boundaries? If the exact same softmax compression is applied to PRL's preference inputs, does PRL exhibit similar perturbation variance/drift (Fig. 5) and simplex concentration (NeffN_{\text{eff}}, Fig. 6)? Moreover, given that δ(e)=0\delta^{(e)}=0 at test time and static p=(1,2,3)p=(1,2,3) already reaches 53.4%53.4\% (Table 5), what justifies keeping the hypernetwork classifier generator?

  1. Multi-seed statistical significance and training alignment (Tables 1, 4, 5): Since

Rrad\mathcal{R}_{\text{rad}},

Lrar\mathcal{L}_{\text{rar}}, and

Ccon\mathcal{C}_{\text{con}}

each contribute only 0.3%∼0.5%0.3\% \sim 0.5\% in Table 4, please report the mean and standard deviation over 3–5 random seeds for Tables 1 (CIFAR-100-LT), 4, and 5, and confirm whether data augmentation and learning-rate schedules on CIFAR-100-LT are strictly identical to PRL.

  1. True effective magnitude of the one-sided radial offset (Eq. 5 & Fig. 4): Because γe,c∈(−1,1)\gamma_{e,c} \in (-1, 1) and sβ=30×0.005=0.15s\beta = 30 \times 0.005 = 0.15 in Eq. (5), the maximum positive bias for tail classes is only 0.150.15 logits. In Fig. 4, where target–competitor margins span [−10,+7][-10, +7], the few-shot mean margin shifts by just +0.07+0.07 from the Moment stage (−2.10-2.10) to the Radial stage (−2.03-2.03), whereas the jump to −1.10-1.10 in the Consensus stage stems from incorporating the log-prior τelog⁡πc\tau_e \log \pi_c. In a leave-one-out ablation removing only

Rrad\mathcal{R}_{\text{rad}}

from the full model (while retaining

Lrar\mathcal{L}_{\text{rar}}

and Ccon\mathcal{C}_{\text{con}}), how much does accuracy actually drop?

1 comment

Response to Reviewer XX7d

Thank you for the detailed questions. Our supplementary analyses directly examine the radial term’s contribution, the effect of preference normalization, and the framework’s performance under a matched training recipe.

The supplementary Spatial Moment results below use existing checkpoints from a separate CIFAR-100-LT configuration: IF=100IF=100, ResNet-32, 200 epochs, seed 1, and FP32 evaluation on all 10,000 test images. We report these separately from the main-table configuration.

1. Direct full-minus-radial evaluation

We removed only Rrad\mathcal R_{\mathrm{rad}}, keeping the checkpoint, pooling, and classifiers fixed. Consensus weights also remained unchanged because they are computed from the raw logits.

Prediction configurationOverall (%)Few (%)
Full minus radial offset52.9031.40
Full53.0432.80
Difference (percentage points)+0.14+1.40

Within the Few group, the correction fixed 43 predictions and introduced one new error. Its benefit is concentrated on tail classes, with a modest tradeoff in Many and Medium accuracy.

This directly answers whether the bounded radial term affects the complete prediction pipeline. Its effect is consistent with targeted calibration near decision boundaries, where a small logit adjustment can change a borderline prediction. We therefore present it as an auxiliary mechanism for improving tail recognition.

2. Matched training protocol and paired uncertainty

The supplementary method–baseline comparison uses identical augmentation, optimization, and training budgets, including learning-rate drops at epochs 160/180, weight decay 5×10−45\times10^{-4}, and Mixup α=0.4\alpha=0.4.

Spatial Moment achieves 53.04% Overall / 32.80% Few, compared with 52.04% / 31.17% for the reference baseline. Across 20,000 class-stratified paired bootstrap resamples, the 95% intervals for the differences are [0.27, 1.73] percentage points Overall and [0.23, 3.03] points Few.

This provides a controlled comparison of the complete framework under a shared recipe. The intervals describe test-sample uncertainty for the fixed checkpoints; the smaller cumulative ablation gains remain configuration-specific point estimates.

3. What the softmax control establishes

Softmax provides a shared normalized state for classifier generation and pooling:

a\in\mathbb R^E,\quad \delta\in\Delta^{E-1}.$$ It bounds pooling modulation to $b_e<p_e<b_e+0.5$. The three-expert bound $q_{\max}\leq \mathrm e/(\mathrm e+2)$ applies when the entries of $a$ are equal. Our new control compares both mappings on the **same original PRL checkpoint**, using 1,000 fixed test probes and 24 perturbations per setting. At raw perturbation magnitude 1.0, softmax reduces effective input displacement (RMS) from **0.677 to 0.183** and mean logit variance from **$9.85\times10^{-7}$ to $4.43\times10^{-7}$**. At matched effective displacement, the variance advantage is absent. This clarifies input compression as an important factor in the preference diagnostics and gives a more precise interpretation of the mapping’s behavior. **4. The hypernetwork’s role and fixed deployment** The hypernetwork provides a shared training interface for classifier generation and bounded pooling modulation. A fixed ladder defines the spatial hierarchy, and the preference state adjusts the orders within that structure. In the supplementary fixed-checkpoint intervention, the deployed orders achieve **53.04% Overall / 32.80% Few**, compared with **52.44% / 32.17%** for fixed $(1,2,3)$. This directly demonstrates the effect of aggregation configuration in the trained model. At inference, $q^\star=\operatorname{softmax}(a)$ is fixed, so both $W_e^\star=h_e(q^\star)$ and the pooling orders can be precomputed and cached. Sample-dependent adaptation remains in consensus weighting, giving the shared training interface a straightforward deployment form. Sincerely, The Authors

Comments on the paper

1 comment

Dear Reviewers,

The authors have provided their rebuttals to the reviewers' comments. Please take a moment to review the responses and continue the discussion on any remaining concerns, unresolved issues, or points that may require further clarification.

Your further input and discussion are highly appreciated.

Regards,
AC for this paper

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.