acceptodds
Under review as a conference paper at ICLR 2027

Is Grokking Observable from Representation Dynamics?

Abstract

Prior work has linked grokking to the evolution of hidden representation geometry, yet it remains unclear whether representation geometry can provide a reliable observable of grokking during learning. To address this question, we introduce Mint, a label-free statistic that quantifies pairwise representation overlap, i.e., the overlap of hidden features, and identify geometry-to-behaviour assumptions under which representation overlap serves as a reliable observable of grokking. Although these assumptions need not hold in general, Mint nevertheless captures informative changes in representation overlap throughout training. We further investigate Mint-penalized training as a means of reducing representation overlap. Our experiments show that penalizing the loss by Mint within a small range can advance the onset of generalization in practice. Together, our results clarify the conditions under which representation overlap can serve as an observable of grokking. We also identify representation overlap as a practical target for influencing generalization onset, and establish Mint as an informative descriptor of representation dynamics throughout learning.

Primary area:
learning theory
Keywords:
grokking, representation dynamics, representation geometry, feature overlap, observability, delayed generalization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_V9jjOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether representation geometry can provide useful observables of grokking. The authors introduce M_int, a label-free scalar statistic that measures pairwise overlap between hidden features. They relate M_int to participation-ratio effective rank and introduce target-subspace alignment to formalize whether a representation captures task-relevant directions.

Theoretically, the paper establishes an unobservability result showing that the complete history of M_int, and more generally statistics based only on the feature-side Gram matrix, is insufficient in general to determine the target-relevant geometric transition. Empirically, the authors study the evolution of M_int during grokking on modular tasks, construct an M_int-based detector, and investigate interventions that explicitly penalize representation overlap. The interventions can advance the initial onset of generalization in the tested settings, although this comes with reduced retention and final test performance.

Critical strengths
  1. The general question of whether representation geometry provides useful observables of learning progress during grokking is interesting. If one could identify an internal geometric quantity that reliably precedes generalization, this could potentially provide both mechanistic understanding and a practical diagnostic of learning dynamics.

  2. The paper contains a substantial theoretical component. In particular, the unobservability result clearly characterizes an important limitation of feature-overlap statistics: identical overlap dynamics can in principle correspond to different target-relevant representation dynamics.

  3. I appreciate that the paper includes several controls rather than simply reporting a correlation between M_int and test accuracy. In particular, the experiments investigate cases in which representation overlap changes without successful generalization, as well as interventions that directly modify overlap.

  4. The intervention experiments are interesting. Penalizing representation overlap can substantially change the onset of generalization, suggesting that the measured geometric quantity is not merely a passive statistic of training dynamics.

Critical weaknesses

My main concern is that I do not yet understand the relationship between M_int and grokking, despite this being the central framing of the paper.

First, the different characteristic times associated with M_int do not appear to track the grokking time particularly well. Already in the schematic in Figure 1(b), t_peak, t_detect, and t_grok are substantially separated. This separation is even clearer in the experiments: the M_int peak can occur hundreds of epochs before grokking, and changing hyperparameters can substantially shift the grokking time without producing a comparable shift in the detector time. I therefore do not understand what quantitative relationship is expected between t_peak, t_detect, and t_grok.

More fundamentally, this makes it unclear in what sense M_int should be considered an observable "of grokking", rather than simply an observable of representation dynamics. The theoretical results show that even the complete M_int history is insufficient in general to determine the target-relevant transition, while the empirical detector can also trigger for runs that never grok. If M_int neither reliably determines whether grokking occurs nor predicts when it occurs, I think the paper needs a sharper explanation of what information about grokking M_int actually provides.

Second, I found the geometric meaning of M_int difficult to build intuition for from the current presentation. Figure 1(a) explains how the scalar is computed from the representation matrix, but does not give an intuitive geometric picture of what high versus low M_int representations actually look like. I think simple low-dimensional visualizations would help substantially. For example, the authors could show configurations of feature vectors or representation points with different M_int values, and snapshots of the geometry before, around, and after the M_int peak.

Third, the empirical scope is relatively narrow, and it is unclear how the proposed phenomenon transfers to other settings. Most of the evidence comes from modular tasks and a shallow MLP. The Transformer experiment also seems to behave differently from the characteristic MLP picture. It would strengthen the paper considerably to demonstrate that M_int captures a meaningful phenomenon in deeper networks, different architectures, or tasks beyond the standard modular grokking setting.

Finally, M_int is exactly related to participation-ratio effective rank for fixed feature dimension. This raises a novelty question: if M_int is essentially a monotonic reparameterization of an existing effective-rank statistic, what new information does M_int provide as a representation diagnostic? This question becomes particularly important if the main conclusion is that M_int is an informative descriptor of representation dynamics rather than a reliable predictor of grokking.

Questions for authors
  1. What exactly is the expected relationship between t_peak, t_detect, and t_grok? Is there a quantitative correlation between these quantities across seeds, hyperparameters, architectures, or tasks? In the current experiments, they appear to respond quite differently.

  2. In what precise sense should M_int be considered an observable "of grokking"? The theoretical results suggest that the complete M_int trajectory cannot generally determine the target-relevant transition, and empirically the detector neither determines whether grokking occurs nor closely tracks when it occurs. What information about grokking does M_int provide beyond indicating that representation geometry is changing?

  3. Could the authors provide a more intuitive geometric visualization of M_int? For example, it would be helpful to show simple low-dimensional feature configurations corresponding to high and low M_int, and ideally how this geometry evolves during training.

  4. Since M_int is exactly related to participation-ratio effective rank for fixed K, what is fundamentally new about M_int compared with directly monitoring effective rank? Are there settings in which M_int provides information that participation ratio does not?

  5. How broadly do the empirical observations generalize? Do similar M_int dynamics appear in deeper networks, other architectures, non-modular tasks, or more realistic datasets? The Transformer results seem to already suggest that the characteristic MLP dynamics are not universal.

  6. What is the intended practical use of monitoring M_int? If its peak or detector time does not reliably predict t_grok, what decision could a practitioner make based on observing M_int? The intervention experiments provide one possible application, but advancing the onset also appears to reduce retention and final performance.

2 comments
AuthorsOpenReview ↗

## To Reviewer V9jj (part 1)

W1 — Core relationship between Mint and grokking is unclear; time relationships are not stable

Thank you for this concern about the central framing. The reviewer is right that Mint's characteristic times do not reliably track grokking time. The data show this clearly: the raw Mint peak occurs at epoch ~40, the detector alarm at epoch ~60, and robust onset at epoch ~800 ± 29. Under hyperparameter changes, onset shifts by thousands of epochs while the alarm barely moves (e.g., lr=1e-4: onset ~2440, alarm ~148; wd=0.03: onset ~2200, alarm ~54). There is no stable proportionality.

The relationship between Mint and grokking should be stated as a boundary, not a prediction. Mint measures feature overlap — it does not specify which target directions the representation captures. Theorem 1 formalizes this: two trajectories can share identical overlap histories while differing in whether they capture a fixed target. The negative controls confirm this empirically: overlap changes occur without true-label generalization. The detector is therefore better described as an early representation-change detector, not a grokking predictor.

To alleviate this concern, we will rewrite the introduction to frame Mint as a bounded observable (information limit) rather than a grokking predictor, and remove overstrong positive-observability language from the abstract.


Q1 — Expected relationship between t_peak, t_detect, and t_grok; quantitative correlation

Thank you for this question about quantitative relationships. No stable quantitative relationship exists in our data. The peak-to-onset gap ranges from ~760 epochs (baseline) to ~2290 epochs (lr=1e-4), with no consistent ratio or proportionality. At the current logging resolution (every 10 epochs), the detector alarm shows zero seed-level variation in the baseline (all seeds at epoch 60), so there is no within-condition correlation to report either.

To alleviate this concern, we will provide a per-run table of all three times across conditions and seeds, and state explicitly that no predictive relationship is established.


Q2 — In what sense is Mint an "observable of grokking"?

Thank you for this question about the meaning of observability. Strictly, Mint is not shown to be an observable of behavioral grokking in the predictive sense. Theorem 1 establishes an information limitation: the complete Mint trajectory cannot in general determine whether generalization has occurred. The negative controls confirm this empirically — alarms fire for non-grokking runs. The detector does not predict grokking; it detects representation change.

What Mint does provide beyond indicating geometry is changing: (1) a formal information limit on what any feature-overlap monitoring can achieve (Theorem 1), regardless of monitoring frequency or cleverness; and (2) a controllable intervention handle — when used as β⋅Mint2\beta \cdot M_{\text{int}}^2 in the loss, it changes onset and retention in a structured, measurable way (Table 2).

To alleviate this concern, we will correct the abstract's implication of positive observability and reframe the paper around the information limit.


Q3 — Geometric visualization of high/low Mint

Thank you for this suggestion. We provide a new mathematical schematic (attached as Figure R1: geometry_schematic_v2.png).

Panel (a) plots Mint=∣cos⁡θ∣M_{\text{int}} = |\cos\theta| and rPR=2/(1+cos⁡2θ)r_{\text{PR}} = 2/(1+\cos^2\theta) continuously as the angle θ\theta between two feature columns varies from 0° to 90°. Three key points are annotated: parallel (θ=0°\theta=0°, Mint = 1, fully redundant), θ=45°\theta=45° (Mint ≈ 0.71), and orthogonal (θ=90°\theta=90°, Mint = 0, fully independent). The PR curve on the right axis confirms the exact one-to-one equivalence.

Panels (b) and (c) illustrate the information limit in 3D: two pairs of orthogonal features (both with Mint = 0) span different planes. In panel (b), the target direction lies outside the feature span — the representation cannot capture it. In panel (c), the target lies inside the span — fully captured. Same Mint value, different target alignment. This is the geometric content of Theorem 1: overlap alone does not determine target capture.

All vectors represent feature columns across probe samples, not data points in hidden-coordinate space. The figure is clearly labeled as a mathematical schematic, not experimental data. Showing real training-time geometry snapshots would require recovering the original checkpoint representations and is noted as additional work.


AuthorsOpenReview ↗

## To Reviewer V9jj (part 2)

Q4 — Since Mint equals PR for fixed K, what new information does it provide?

Thank you for this question about informational content. For fixed KK and the same column-normalized matrix, Mint provides no additional scalar information beyond PR — they are exact reparameterizations via Eq. (2), and we should not suggest otherwise.

The value of the Mint formulation is operational, not informational: the cosine formulation makes Mint2M_{\text{int}}^2 a natural differentiable penalty that enters the loss function directly, while PR is more natural for interpreting effective dimensionality. They are the same number; the choice is one of convenience for the application at hand. Mint's specific contribution is the dual role described in our response to W1 of Reviewer UpSV: the same quantity serves as both the diagnostic analyzed theoretically (§3, Theorem 1) and the penalty used for intervention (§4, Table 2). When KK changes during training (e.g., as neurons die), it must be tracked alongside Mint — this is a normalization shift, not additional information content.

To alleviate this concern, we will state the PR equivalence and this operational distinction explicitly in the revision.


Q5 — How broadly do these observations generalize?

Thank you for this question about scope. Our experiments cover a two-layer ReLU MLP on modular addition and multiplication, and a one-layer Transformer on modular addition. The Transformer results show a boundary peak (all 10/10 baselines peak at epoch 0), which does not support the MLP's internal-peak phenomenon. This is an honest scope boundary, not a minor caveat.

We have not tested deeper MLPs, larger Transformers, or non-algorithmic tasks. Whether the onset–retention trade-off and the information limit transfer to these settings are genuine open questions that we cannot resolve with a disclaimer.

To alleviate this concern, we will add an explicit scope paragraph stating what architectures and tasks are covered, what the Transformer boundary peak means for generality, and what remains untested.


Q6 — Practical use of Mint; how to address the onset–retention cost

Thank you for this question about practical implications. In the submitted paper, we do not demonstrate a validated practical policy. The onset–retention trade-off is the paper's main empirical finding: β = 0.17 halves the onset time (from ~800 to ~400 epochs) but cuts retention from 1.000 to 0.047 and final accuracy from 1.000 to 0.885. These costs are central to the result, not secondary caveats.

Our preliminary experiments (not in the submitted version) suggest a path toward resolution: applying the penalty only until generalization onset, then removing it, preserves the earlier onset (epoch ~405 vs. baseline ~840) with retention restored to 1.000 and final accuracy to 1.000 (tested on 2–3 seeds in a local environment). If confirmed at larger scale, this would mean that temporal scheduling of the Mint penalty — apply early, remove after onset — could provide the acceleration without the retention cost.

To alleviate this concern, we will describe this penalty-scheduling direction as future work with appropriate caveats in the revised conclusion, clearly distinguishing it from the submitted results.

Reviewer_UpSVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Prior work has linked the onset of grokking with certain representational geometry signals. The authors introduce the MINT metric as a measure of representational overlap and discuss assumptions where this type of metric can be an indicator of grokking. They also discuss how this metric can be optimized for in order to encourage generalization.

Critical strengths
  1. The authors detail theoretical assumptions on which their metric corresponds to an indicator of generalization behavior. This is a useful framing that can potentially inspire how future theoretical of grokking behavoir might occur.

  2. The authors persent a variety of ablation studies of the Mint method to stress test its utility in different grokking scenarios. This depth provides a useful understanding of the limits of their method. They are very in depth about discussing its limiations and framing when and in what context their metric is useful.

Critical weaknesses
  1. The biggest problem with this paper is that the Mint metric itself can be argued as a generalization of many other metrics that have existed before. Essentially, the metric is the average of the cosine similarity of all features of a sample in a batch. This type of value has potentially strong correlation with other metrics that attempt to measure the overlap of features such as various covariance based metrics as a simple case. However, the biggest problem is that there are many other metrics that are trying to formulate representational quantities to indicate generalization. For example,

Effective rank: The effective rank: a measure of effective dimensionality Alpha-REQ: alpha-ReQ : Assessing Representation Quality in Self-Supervised Learning by measuring eigenspectrum decay LiDAR: LiDAR: Sensing Linear Probing Performance in Joint Embedding SSL Architectures

In the grokking specific literature, that was cited by the authors, there is the ILDR ratio based on inter to intra class distances (similar to cosine similarity averages) and other metrics based on eigenvalues of covariance matrices. All of these metrics do something similar to mint in measuring how features are spreading within a sample matrix. The authors provide a theoretical justification for how Mint indicates generalization behavior, but if Mint is not inherently unique then the theoretical justification is just a rehashing of known theoretical behaviors.

The authors need to compare against other metrics to justify why Mint is especially equipped to be the dynamics behavior we observe rather than one of these other metrics. What makes Mint especially useful, rather than one of these other metrics? This needs to show up in experiments as well where Mint needs to be compared against these other measures and show it is especially better at this task compared to other metrics.

  1. The theoretical justification needs to contrast against existing theoretical analyses of grokking. https://arxiv.org/pdf/2509.21519 serves as a useful foundation. I think the work needs to explain how the theoretical insights of this either fit within the discussions of this paper or contrast against it to disprove the grokking assumptions made in this work. Otherwise, it isn't clear what this paper is adding to the grokking theoretical discussion.

  2. The paper is written in a way that suggests it isn't complete. The authors introduce the Mint method, but make it clear that it isn't a direct indicator of grokking, but just an analysis tool. This doesn't inspire much confidence to the method and the analysis itself. I appreciate the honesty of limitations, but if this metric isn't a good indicator, then it would be better to have a comprehensive analysis of why its better than other intuitive methods as discussed in point 1. Instead, the whole paper is isolated within the context of Mint without proper leveraging of other ideas to contrast against. This point is especially pronounced in figure 2. I see that Mint is maximal at epoch 40, but the grokking starts around epoch 260. I don't understand how this serves as a useful indicator if there is a 200 epoch gap between the two behaviors.

  3. Optimizing for Mint isn't a very novel concept. Mint is an optimization on the cosine similarity of features. Similar ideas have actually been done in large scale pre-training papers such as Dino v3 where additional gram reguralization was needed for localized regularization of papers. It is hard to call this as a novel concept.

Questions for authors

Could you provide more clarification on Figure 5 and Table 3. I don't understand the conclusions of these figures.

8 comments
AuthorsOpenReview ↗

## To Reviewer UpSV (part 1)


W1 — Mint can be argued as a generalization of many existing metrics; comparison with Effective Rank, α-ReQ, LiDAR, ILDR is missing


Thank you for raising this comparison. We agree that Mint shares mathematical connections with existing representational measures — in particular, for fixed KK and the same column-normalized features, Mint and participation-ratio (PR) rank are exact reparameterizations of each other via Eq. (2). However, these metrics differ in important ways, and we believe Mint occupies a distinct role that the current manuscript does not explain clearly enough. We clarify both the common ground and the differences below.

What these metrics share: All five metrics below quantify some aspect of how features spread or concentrate in a representation matrix. In that broad sense, they are related — a representation with highly redundant features will score differently from one with diverse features on all of them.

Where Mint differs: The key distinction is not in the scalar value it computes, but in how it is designed to be used. Mint is formulated as a pairwise cosine quantity specifically so that its square (Mint2M_{\text{int}}^2) can serve directly as a differentiable regularization term in the training loss. This is the role it plays in our intervention study (Table 2): we add β⋅Mint2\beta \cdot M_{\text{int}}^2 to the cross-entropy loss to actively steer the representation during training, and measure the resulting effect on both onset timing and post-onset retention. The other metrics listed by the reviewer are designed as diagnostic tools — they measure properties of a trained or training representation, but are not formulated as loss penalties.

MetricWhat it measuresHow it is usedSimilarity to MintDifference from Mint
Effective rank (Roy & Vetterli)Shannon entropy of normalized singular valuesPost-hoc diagnosticBoth summarize feature diversity in a single scalarSpectral summary, not a pairwise cosine; not formulated as a loss term
α-ReQ (NeurIPS 2022)Power-law decay exponent of centered covariance eigenspectrumRepresentation quality assessmentBoth relate to how spread out the feature spectrum isOperates on centered covariance, not cosine Gram; designed for evaluation, not training
LiDAREntropy-based rank of between-instance vs. within-instance variationProbing without downstream labelsBoth are label-free measures of feature structureRequires multiple views per input (not available in our single-view arithmetic setting); diagnostic only
ILDRInter-class centroid separation vs. intra-class scatterGrokking-specific diagnosticBoth track representation geometry changes during trainingRequires class labels on held-out data; diagnostic, not a penalty
Mint / PRRMS off-diagonal cosine of column-normalized featuresBoth diagnostic and loss-function penalty (β⋅Mint2\beta \cdot M_{\text{int}}^2)Exact equivalence with PR under same normalizationDual role: the same quantity used for observation (§3) and intervention (§4); this enables the onset–retention study

The onset–retention trade-off we report in Table 2 — onset halved while retention drops from 1.000 to 0.047 — is made possible precisely because Mint is not only a measure but also a training objective. This dual role (observe the same quantity that you intervene on) is what connects our information-limit analysis (Theorem 1: what can overlap histories tell you?) to the intervention study (what happens when you actively push overlap down?). The named comparison metrics do not have this interventional component in our setting.

We acknowledge that the submitted manuscript does not make this distinction clear. In the current version, Mint is introduced primarily as a diagnostic, and its role as a loss factor appears only later in §4 without sufficient motivation.

Planned revision: Rewrite §1 and §3.1 to: (1) state the PR equivalence upfront, (2) include the comparison table above, (3) clearly position Mint's contribution as the dual diagnostic-and-penalty role that enables the onset–retention study, rather than claiming standalone metric novelty.


AuthorsOpenReview ↗

## To Reviewer UpSV (part 2)


W2 — Theoretical justification should contrast against Tian's Li² framework (arXiv:2509.21519)

Thank you for this concern about how our theory relates to existing grokking dynamics analysis. Our work and Tian's address different questions: Tian's Li² framework characterizes how features emerge during training — the three-stage gradient dynamics (lazy → independent → interactive feature learning) and provable scaling laws for feature emergence. Our Theorem 1 characterizes what a feature-overlap history can and cannot determine about whether generalization has occurred — an information limitation that holds regardless of which dynamics generated the path.

We were inspired by the shared insight that feature-learning dynamics underlie grokking (Tian, 2025; Kumar et al., 2024; Lyu et al., 2024). However, Tian's framework does not address post-grokking retention: it characterizes when and how features emerge, but not whether the solution is maintained after generalization, or what happens when a penalty is removed. Our onset–retention intervention study (Table 2) directly addresses this gap — showing that a feature-overlap penalty can halve onset time but at a retention cost, and our preliminary results indicate that removing the penalty after onset restores retention.

Additionally, the information limit (Theorem 1) is compatible with Tian's dynamics: his gradient dynamics could generate a particular Gram path, and our theorem would still apply to that path. The two results are complementary, not competing.

To alleviate this concern, we will add a paragraph in §2 explicitly positioning our work relative to Li², making clear that Tian provides the dynamical mechanism of feature emergence while we provide the observation-limit analysis and the retention study.


W3 — The paper feels incomplete; Mint's role is unclear

Thank you for this concern about the paper's direction. The core issue is that Mint's intended role — as a loss-function factor for studying the acceleration–stability trade-off — is not motivated early enough in the manuscript.

The missing context is: recent work shows that faster grokking does not always mean stable grokking. Janati et al. (2026) find that Muon-accelerated grokking is uniformly unstable across all nine configurations tested; Zhou et al. (NeuralGrok, 2025) report that even standard training with weight decay can oscillate between ~100% and <5% test accuracy after generalization; Verma (2026) finds that higher weight decay (which accelerates grokking) leads to more seeds losing generalization at 20,000 epochs. Our Table 2 adds to this pattern: the Mint penalty accelerates onset from ~800 to ~400 but degrades retention from 1.000 to 0.047.

Mint's intended role going forward is not as a standalone diagnostic but as a controllable penalty whose temporal schedule (when to apply, when to remove) determines whether the onset–retention trade-off can be resolved. Our preliminary experiments show that applying the penalty only until onset, then removing it, preserves the acceleration while restoring full retention — reframing Mint2M_{\text{int}}^2 as a practical tool for studying this trade-off.

Regarding the epoch-40 peak vs. epoch-~800 generalization gap in Figure 2: this wide separation is itself a result — overlap dynamics and generalization dynamics are not synchronous — not evidence of accurate prediction. We will clarify this and remove any implication that an early peak alone predicts the later transition.

To alleviate this concern, we will add a motivation paragraph in §1 citing the post-grokking instability literature and reframe the paper's arc as: information limit → negative controls → intervention → onset–retention trade-off → (future) penalty scheduling to resolve the trade-off.


W4 — Optimizing Mint lacks novelty; similar to DINOv3 Gram regularization

Thank you for this concern about novelty. We agree that regularizing cosine-similarity or Gram-based quantities is not new as a technique. The novelty of our work is not the penalty form but the experimental finding it produces: to our knowledge, no prior work reports that a single representation penalty can halve grokking onset while cutting retention from 1.000 to 0.047, nor that temporal scheduling of the same penalty resolves this trade-off. This finding is enabled by the controlled experimental design: five paired seeds with gradient-norm-matched controls (v1.1, matching error < 3 × 10⁻⁸), which allows the specific decomposition of the penalty's effect on onset timing versus post-onset stability.

To alleviate this concern, we will cite DINOv3 and other Gram regularization precedents explicitly in the revision, and frame the contribution as the experimental finding (the onset–retention trade-off and its potential resolution through scheduling), not as the penalty form itself.

AuthorsOpenReview ↗

## To Reviewer UpSV (part 3)

Q1 — Figure 5 and Table 3 clarification

Thank you for this question about the figures' conclusions. Each panel of Figure 5 answers one specific question:

  • Panel (a): Are the Mint peak, detector alarm, and generalization onset the same event? No — the raw Mint peak is at epoch ~40, the detector alarm at epoch ~60, and the robust onset at epoch ~800. These are separated by hundreds of epochs.
  • Panel (b): Does an alarm imply generalization? No — permuted-target and low-data controls both trigger the alarm without ever grokking. The alarm detects representation change, not generalization.
  • Panel (c): Do alarm times shift proportionally with onset times? No — changing the learning rate to 1e-4 moves onset to ~2440 but the alarm to only ~148; changing weight decay to 0.03 moves onset to ~2200 while the alarm stays at ~54.

The conclusion from all three panels is that the frozen detector transfers as an early representation-change detector, not as a reliable grokking predictor. Table 3 reports the alarm/onset pairs across all conditions; 75 transfer rows use the single frozen detector, with 5 calibration rows listed separately.

To alleviate this concern, we will rewrite the Figure 5 caption to state each panel's question and the negative conclusion directly, and expand the Table 3 legend to explain the row categories.


Reviewer_ZTHeOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Due to formatting issue, I am pausing reviews according to AC guidance.

Critical strengths

Due to formatting issue, I am pausing reviews according to AC guidance.

Critical weaknesses

Due to formatting issue, I am pausing reviews according to AC guidance.

Questions for authors

Due to formatting issue, I am pausing reviews according to AC guidance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.