acceptodds
Under review as a conference paper at ICLR 2027

RELY: Beyond Deterministic Graphs for Uncertainty-Aware Crime Forecasting

Abstract

Crime forecasting relies on dependencies among urban regions, but sparse event histories provide ambiguous evidence about which dependencies are reliable. A single deterministic graph commits to one structural estimate without explicitly representing this ambiguity, potentially allowing unreliable relations to influence predictions. We introduce RELY, a probabilistic structure learning framework that uses uncertainty in urban relations to guide spatiotemporal crime forecasting. RELY encodes geographic, functional, and historical information as probabilistic priors over pairwise links and regional group memberships. It refines these probabilities using each observation window and learns the resulting structures jointly with the forecasting model through differentiable sampling. To connect structural uncertainty with prediction, RELY derives reliability estimates for different relation channels and uses them to guide separate routing mechanisms for count intensity and excess-zero probability in a zero-inflated negative binomial decoder. This enables the two distributional components to draw on different combinations of relational evidence according to its estimated reliability. Experiments on Chicago and New York datasets show MAE reductions of 6.9% and 4.5%, respectively, over the strongest evaluated baselines, alongside improvements in RMSE and MAPE.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph learning, Spatio-temporal forecasting, Crime Prediction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SXdzOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RELY is a crime forecasting framework with probabilistic graph and hypergraph structures. It refines urban relation priors using recent observations. A reliability score summarizes structural support and uncertainty. Separate routers select relational information for the negative-binomial mean and excess-zero components of a ZINB distribution. Experiments cover Chicago and New York with eight baselines and five random seeds. The authors report MAE improvements of 6.9% and 4.5%. Additional analyses examine component contributions, sparse regimes, reliability–error relationships, and computational efficiency.

Critical strengths
  1. The method has a clear task-specific motivation. Sparse crime counts require appropriate output distributions. Different urban relations may provide different predictive information. RELY addresses these properties through probabilistic relation refinement and separate routing for the two ZINB components. This combination offers a plausible approach to heterogeneous crime forecasting.

  2. The results provide evidence of practical value. RELY reports the lowest mean aggregate errors on both datasets. The experiments include five random seeds and several component ablations. Removing the reliability bias increases Chicago MAE from 0.803 to 0.837. Sharing one router increases MAE to 0.830. These results support the usefulness of the proposed routing design under the evaluated setting.

Critical weaknesses
  1. The evaluation does not establish generalization to future periods. Section 4.1 randomly splits forecast windows into training, validation, and test sets. This can place later observations in training than some test forecast origins. Overlapping windows may create additional dependence between the partitions. The paper does not report a chronological evaluation. This limits confidence in the claimed forecasting gains under deployment conditions. The exact overlap and any resulting information leakage require clarification.

  2. The interpretation of structural reliability needs stronger validation. Section 4.5 shows lower errors for samples with higher reliability. The authors acknowledge a possible confound from crime count levels. Quieter regions may have lower absolute errors regardless of structural uncertainty. The current analysis does not establish whether reliability provides information beyond this effect. The ablations demonstrate a useful routing bias. They do not establish that the score reliably measures the quality of inferred relations. This distinction matters because structural reliability is a central contribution.

Questions for authors
  1. Clarification: What is the window stride? Can observations from a test target date appear in a training window? Please specify the dataset dates and the data used to fit preprocessing statistics.

  2. Essential validation: Does RELY retain its advantage under a strictly chronological train/validation/test split? A comparison against the strongest existing baselines would materially affect my assessment.

  3. Essential validation: Does reliability still predict error within groups with similar historical count levels and occurrence frequencies? Does it provide additional information beyond simple count-based scores or reliability computed from the priors alone? These controls would help establish the meaning of the learned reliability scores.

3 comments
AuthorsOpenReview ↗

Response to Reviewer SXdz (1/2)

W1 & Q1 & Q2: The window stride is one day, and observations from test target dates do not appear in any training window. RELY retains its advantage under a strictly chronological train/validation/test split (Table 1). The table below provides the date ranges or versions of each data source and specifies the data used to estimate preprocessing statistics. Upon checking, we confirmed that an early implementation used random window splits, whereas all results ultimately reported in the paper were obtained using training, validation, and test splits ordered strictly by forecast target date. Thus, Table 1 shows that RELY’s gains hold when no training window occurs after a held-out forecast. The manuscript retained the description of the earlier implementation and therefore did not accurately reflect the final experimental setup. We thank the reviewer for identifying this oversight and will correct the description in the revised manuscript.

Data typeDate range or versionPreprocessingData used to estimate preprocessing statistics
Crime counts2023-01-01 to 2023-12-31Apply log⁡(1+x)\log(1+x) to historical inputs; retain raw counts as prediction targetsNo estimation of means, standard deviations, or other statistics is required
311 service requests2023-01-01 to 2023-12-31Aggregate by day, region, and request type; apply log⁡(1+x)\log(1+x) to historical inputsNo global statistics need to be estimated
Weather2023-01-01 to 2023-12-31Apply z-score standardisation separately to each weather variableVariable-wise means and standard deviations estimated from the training set
Socioeconomic attributes2024 ACS five-year estimatesAggregate to regions; impute missing values using the city-specific median; then apply feature-wise z-score standardisationAll regions within each city: 79 in Chicago and 148 in New York; compute feature-wise medians, means, and standard deviations
POIsOpenStreetMap (OSM) snapshot dated 2023-01-01Count POIs by category within each region; apply log⁡(1+x)\log(1+x), then scale each category by its column maximumCategory-wise maxima of log-transformed counts across all regions within the corresponding city
AuthorsOpenReview ↗

Response to Reviewer SXdz (2/2)

W2 & Q3: Reliability still predicts error within groups with similar occurrence frequencies, and beyond reliability computed from the priors alone it provides additional information that helps reduce forecast error. We agree that the available evidence is insufficient to directly validate the latent structural quality, and we therefore limit our conclusions to the score’s in-model diagnostic value for forecasting and its incremental routing utility, which we support with the frequency-stratified analysis and the comparison against prior reliability.

To further examine the effect of occurrence frequency, we have added an analysis stratified by crime occurrence frequency. The table below reports MAE, RMSE, and MAPE across reliability groups within each occurrence-frequency bin, where Q1 denotes the lowest reliability group and Q5 denotes the highest.

Frequency0-0.250.25-0.50.5-0.750.75-1
ReliabilityMAERMSEMAPEMAERMSEMAPEMAERMSEMAPEMAERMSEMAPE
Q1(Lowest)0.1750.5530.0800.5340.9450.2480.8731.2640.3731.5532.1190.438
Q20.1580.4550.0780.5070.9240.2360.8321.2100.3621.4421.9760.435
Q30.1180.4130.0570.4890.9030.2280.8351.1860.3561.3641.8390.433
Q40.0990.4010.0460.4660.8860.2160.8281.1810.3521.3211.8180.424
Q5(Highest)0.0840.3280.0420.4580.8790.2130.7551.0820.3361.2051.6490.412

All four occurrence-frequency bins show an overall trend of decreasing error as reliability increases. From Q1 to Q5, MAE decreases by approximately 51.9%, 14.2%, 13.5%, and 22.4%, respectively. This indicates that the reliability–error relationship persists among samples with similar occurrence frequencies, providing further support for the score’s diagnostic value for forecasting.

To address the question of additional information raised in Q3, we have added a comparison against reliability computed from the priors alone to examine the incremental predictive value of learned reliability. Both variants retain structure refinement and message propagation based on the learned structures, and both use the same reliability function. The full model supplies the routers with c(Qt)c(Q_t), derived from the learned structural distribution, whereas the control variant supplies c(Pˉ0)c(\bar P_0), derived from the completed prior. This comparison differs from an ablation that directly removes structure refinement: it assesses the predictive utility of the two reliability sources while retaining structure learning.

The results within different occurrence-frequency groups are reported below. “Learned Reliability” denotes the full model, and “Prior Reliability” denotes the variant that uses prior reliability for routing.

MAEMAERMSERMSEMAPEMAPE
frequencyLearned ReliabilityPrior ReliabilityLearned ReliabilityPrior ReliabilityLearned ReliabilityPrior Reliability
[0,0.25)0.126±0.00020.126±0.00080.423±0.00030.424±0.00090.061±0.00010.061±0.0004
[0.25,0.5)0.489±0.00680.501±0.01450.895±0.00610.899±0.00780.228±0.00350.234±0.0078
[0.5,0.75)0.819±0.01480.835±0.02621.183±0.02621.212±0.05290.353±0.00670.360±0.0104
[0.75,1]1.431±0.02591.571±0.15251.988±0.04582.275±0.30060.422±0.00480.432±0.0099

In the three intervals with occurrence frequencies of at least 0.25, the full model achieves lower MAE, RMSE, and MAPE than the prior reliability variant. In the highest-frequency group, the three metrics decrease by approximately 8.9%, 12.6%, and 2.3%, respectively. In the lowest-frequency group, the two variants have essentially the same mean errors. In addition, the full model has smaller standard deviations on all three metrics across all four frequency groups, indicating lower variability across random seeds.

Structural reliability is a model-internal confidence indicator derived from the support strength and decisiveness of the relation posteriors, used to assist the allocation of relational information and forecast diagnosis. The two analyses above support, respectively, the association between the score and forecast error and its incremental predictive utility over the prior score, but they do not directly test whether the latent relations are in fact correct. We therefore limit the corresponding conclusions to the score’s diagnostic and routing value within this model, and do not interpret it as direct validation of true structural quality or of probability calibration.

The author's experiments addressed my concerns to some extent, so I have decided to increase my score.

Reviewer_Tn5aOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a new method to forecast crimes in predefined regions, which are represented as graphs. The method models many information while generating the forecast, such as the history of crime counts over different categories, auxiliary urban activities, weather, socioeconomic attributes, POI counts, and spatial adjacency. The core idea of the paper is to model this information as relationships to capture partial or direct correlations through uncertain relations and make it more flexible for forecast errors, which is especially strong when the sparsity of events exists. The paper models the prediction through a zero-inflated negative binomial (ZINB) distribution and predicts the median. The experiments shows statistically significant improvements.

Critical strengths
  • The paper is like a library of techniques for researchers. There are many intricate details and design problems that are addressed with deliberate techniques.
  • The paper actually focuses on sparsity and uncertainty problem in crime forecasting which is the main reason most of the models fall short in regions with sparse event history.
  • The offered refinement techniques shows the reasons which goes to the final model prediction, therefore, it somewhat offers interpretibility.
  • The experiments provide statistically significant improvements.
Critical weaknesses
  • The paper's presentation is its weakest point. There are many details one needs to grasp and there are many design decisions that were explained with one sentences or never explained.
  • The presentation of the results is not good. For example, in Figure 3, instead of mean absolute change, relative change could be used. And how many steps of refinement were performed?
  • The method trains the MAE term on the mean but forecasts with the median. That’s a small train/inference mismatch which needs further discussion.
  • There are many hyperparameters and definitions; the system is more of an engineering effort with many intricate details and low presentability.
  • The relationship between prior and posterior can be explained through the Bayesian formula
Questions for authors
  • Can equation 5 be represented as a Bayesian formula? What do you call this posterior?
  • What does the denominator represent in the calculation of d(q) in equation 6
  • "During training, the differentiable conditional mean 'Y^mean_t = (1−π_t)⊙µ_t' is used in the MAE term of the prediction loss." This can be shown mathematically: reducing the mean also reduces the prediction loss.
3 comments
AuthorsOpenReview ↗

Response to Reviewer Tn5a (1/2)

W1: Thank you for pointing out the presentation issues. We have made our code publicly available, and the design details are reflected in the implementation. We also recognise that the explanations of our design decisions in the paper are insufficient. We will provide more detailed explanations of these decisions in the revised manuscript.

W2: All experiments reported in the paper use a single refinement step (K=1K=1). In Figure 3 of the revised manuscript, we will retain the mean completed prior probability and mean posterior probability and replace the mean absolute change with the signed relative change: Mean⁡(Qtr)−Mean⁡(Pˉ0,tr)Mean⁡(Pˉ0,tr)×100\frac{\operatorname{Mean}(Q_t^{r})-\operatorname{Mean}(\bar P_{0,t}^{r})}{\operatorname{Mean}(\bar P_{0,t}^{r})}\times 100 \\%. Our framework supports multiple steps of structure refinement. We use a single refinement step primarily for the following reasons. First, our model learns a residual correction to the prior log-odds, allowing the network to predict the overall correction directly from the current window without requiring multiple small updates. Second, a single step limits potential error accumulation across repeated structure updates and reduces additional computational overhead. We therefore adopt a single refinement step as a design choice that balances window adaptivity, prior constraints, and computational efficiency.

W3: Our choice of the predictive median for point forecasting follows Adler et al., Beyond Accuracy: Are Time Series Foundation Models Well-Calibrated? (ICLR 2026). In §3.1, Eq. (4), that study uses the predictive median to compute the point forecast metric MASE. Following this evaluation practice, we likewise use the median of the predictive distribution as the point forecast. However, the median of a discrete distribution is a piecewise constant function of the distribution parameters, making it difficult to obtain useful gradients through standard backpropagation. We therefore use the differentiable conditional mean to construct an auxiliary training objective.

W4: We acknowledge that RELY involves many hyperparameters and symbol definitions, and that the concentration of implementation details in the original manuscript obscures the core contribution.

The novelty lies not in adding modules or hyperparameters, but in explicitly propagating structural uncertainty into the allocation of predictive evidence. Specifically, RELY learns window-conditioned relation posteriors constrained by priors derived from urban evidence, derives structural reliability from these posteriors, and uses separate reliability-guided routers to organise evidence for predicting crime intensity and the excess-zero probability. These three mechanisms address the same research question through a connected design, rather than simply stacking modules.

The hyperparameters control model capacity, the scope of structure search, prior and sparsity constraints, and structure sampling, each with a clear modelling role. The sensitivity experiments in Appendix A.5 support the model’s robustness within the tested ranges. In the revision, we will highlight these core connections, simplify the symbol definitions, move nonessential details to the appendix, and group the hyperparameters by function with explanations of their settings to improve readability and reproducibility.

SymbolMeaning and role
dg,dhd_g, d_hGraph and hypergraph hidden dimensions, controlling representation capacity.
k0k_0Prior neighborhood size, controlling the initial membership of anchor-based hyperedges.
ksk_sSemantic neighborhood size, controlling the expansion of candidate relations.
pnewp_{\mathrm{new}}Weak prior probability assigned to newly admitted candidates without observational support.
τ\tauBinary-Concrete temperature, controlling the degree of relaxation in structure sampling.
λKL\lambda_{\mathrm{KL}}Weight of the prior-preserving term, limiting posterior deviation from the completed prior.
λsparse\lambda_{\mathrm{sparse}}Weight of the sparsity term, penalising high posterior probabilities on unsupported new candidates.
λdiv\lambda_{\mathrm{div}}Weight of the diversity term, discouraging structural collapse across relation channels.
ηaux\eta_{\mathrm{aux}}Weight of the auxiliary urban-activity prediction loss, supervising the excess-zero branch.
AuthorsOpenReview ↗

Response to Reviewer Tn5a (2/2)

W5 & Q1: Equation (5) can be equivalently rewritten as a Bayesian odds update, but it is not an exact Bayesian posterior derived from an explicit generative likelihood. The term “posterior” in our paper refers to the structural distribution obtained by updating the urban relation priors using information from the current historical window. It describes which relations may exist and their corresponding probabilities of existence.

Specifically, for non-fixed relations within the candidate set, let the prior probability be p=Pˉ0,ijrp=\bar P^r_{0,ij} and the updated probability be q=Qt,ijrq=Q^r_{t,ij}. Equation (5) can be written as q=σ!(logit⁡(p)+Δ)=p,eΔ(1−p)+p,eΔq=\sigma\\!\left(\operatorname{logit}(p)+\Delta\right)=\frac{p\\,e^\Delta}{(1-p)+p\\,e^\Delta}, and therefore, q1−q=p1−p,eΔ\frac{q}{1-q}=\frac{p}{1-p}\\,e^\Delta. This has the same form as the Bayesian update “posterior odds = prior odds × likelihood ratio”. When Δ\Delta corresponds to the log-likelihood ratio of the current-window evidence under the hypotheses that the relation exists and does not exist, respectively, this expression is the standard Bayesian update. Strictly speaking, this likelihood ratio should be conditioned on the information used to construct the prior.

In RELY, Δ\Delta is computed by a neural network from the current-window representations and learned jointly through the prediction objective and the KL prior constraint. We do not explicitly specify a generative likelihood for this observational evidence and therefore do not directly interpret Δ\Delta as a validated log-likelihood ratio. Equation (5) instead uses the Bayesian odds update as a parameterisation for data-driven structure refinement under prior constraints: Δ=0\Delta=0 preserves the prior, while positive and negative corrections respectively increase and decrease the odds that a relation exists.

Q2: The quantity ∑jqj\sum_j q_j in the denominator is the total posterior relation mass, that is, the sum of the posterior relation probabilities; it is equivalently the expected number of active relations associated with the region. When this mass exceeds ϵ\epsilon, it serves as the normalising factor for the weights wj=qj/∑kqkw_j=q_j/\sum_k q_k, making d(q)d(q) a probability-weighted average of normalised posterior decisiveness. This separates the support strength a(q)a(q) from the average decisiveness d(q)d(q) of that support. The ϵ\epsilon safeguard ensures numerical stability when the posterior relation mass is zero or very small.

Q3: Reducing the predictive mean does not generally reduce the prediction loss. We provide the following proof.

Let y^=(1−π)μ\hat{y}=(1-\pi)\mu, let yy denote the observed count, and let the prediction loss for a single observation be ℓpred=ℓNLL+λMAE∣y^−y∣\ell_{\mathrm{pred}}=\ell_{\mathrm{NLL}}+\lambda_{\mathrm{MAE}}|\hat{y}-y|, where λMAE>0\lambda_{\mathrm{MAE}}>0. Then, ∂ℓpred∂∣y^−y∣=λMAE>0\frac{\partial\ell_{\mathrm{pred}}}{\partial|\hat{y}-y|}=\lambda_{\mathrm{MAE}}>0 always holds, so reducing ∣y^−y∣|\hat{y}-y| also reduces the loss. However, for y^≠y\hat{y}\neq y, ∂∣y^−y∣∂y^=sign⁡(y^−y)\frac{\partial|\hat{y}-y|}{\partial\hat{y}}=\operatorname{sign}(\hat{y}-y). Therefore, reducing y^\hat{y} decreases the loss when correcting an overestimate, but increases the loss when the prediction already underestimates the observed count.

I thank the authors for the responses to my questions. I believe the paper can be improved with better presentation, but I am satisfied with the answers given to my questions; therefore, I am keeping my already positive score.

Reviewer_6idTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RELY learns window-specific probabilities for several graph and hypergraph relations between urban regions. It computes channel reliability from these probabilities and uses separate routers for the intensity and excess-zero parameters of a zero-inflated negative-binomial forecast.

Critical strengths

Giving count intensity and excess zeros separate relation mixtures is a useful modeling choice

Critical weaknesses
  1. Sec. 4.1 randomly splits overlapping seven-day windows, allowing training windows later than test forecasts, so the main results do not establish performance on future periods.
  2. Sec. 4.5 does not control the reliability-error relationship for count level, and no calibration results establish that the score measures structural uncertainty.
Questions for authors
  1. Do the gains hold when training, validation, and test targets occupy successive time periods and no training window occurs after a held-out forecast?
  2. Does reliability still predict error within regions and count-level groups, and does it improve predictive likelihood or interval coverage over the deterministic variant?
  3. Are gradients stopped through the reliability function in Eq. (6)?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 6idT

W1 & Q1: All results reported in the paper were obtained using chronological data splits. The results in Table 1 show that RELY’s gains hold when no training window occurs after a held-out forecast. Upon checking, we confirmed that an early implementation did randomly split forecast windows. However, the final protocol assigns training, validation, and test targets to successive time periods according to their forecast target dates, and all reported results were obtained under this chronological protocol. Section 4.1 retained the description of the earlier implementation and therefore did not accurately reflect the final experimental setup. We will correct this description and thank the reviewer for bringing this oversight to our attention.

W2 & Q2: The reliability score produced by RELY still predicts error when regions are grouped by crime occurrence frequency. RELY also improves predictive likelihood over the deterministic variant.

First, to assess RELY’s probabilistic forecasting quality relative to the deterministic variant, we have added a test negative log-likelihood (NLL) evaluation to complement the ablation study in Figure 2. The complete results are as follows:

VariantNLL
Ours1.225 ± 0.0070
w/o Stochastic Sampling1.259 ± 0.0080
w/o Posterior Refinement1.258 ± 0.0058
w/o Reliability1.265 ± 0.0163
w/o Separate Routing1.258 ± 0.0058
w/o Hypergraph1.257 ± 0.0050
w/o Heterogeneous Graph1.271 ± 0.0096
w/o 311 Auxiliary Loss1.265 ± 0.0054
w/o ZINB Decoder1.262 ± 0.0046
w/o HF1.258 ± 0.0025
w/o HE1.264 ± 0.0090
w/o Temporal1.258 ± 0.0073
w/o KL1.258 ± 0.0132
w/o Sparsity1.256 ± 0.0070
w/o Diversity1.264 ± 0.0114

The full model achieves an NLL of 1.225 ± 0.0070, compared with 1.259 ± 0.0080 for the deterministic variant, w/o Stochastic Sampling, corresponding to an approximately 2.7% reduction in mean NLL. This provides evidence of improved predictive likelihood beyond point forecast error metrics.

Second, to examine whether the reliability–error relationship also holds among regions with similar crime occurrence frequencies, we divide regions into four groups based on historical crime occurrence frequency and compare MAE, RMSE, and MAPE across reliability levels within each group. Q1 denotes the lowest reliability quantile, and Q5 denotes the highest. The results are as follows:

Frequency0-0.250.25-0.50.5-0.750.75-1
ReliabilityMAERMSEMAPEMAERMSEMAPEMAERMSEMAPEMAERMSEMAPE
Q1(Lowest)0.1750.5530.0800.5340.9450.2480.8731.2640.3731.5532.1190.438
Q20.1580.4550.0780.5070.9240.2360.8321.2100.3621.4421.9760.435
Q30.1180.4130.0570.4890.9030.2280.8351.1860.3561.3641.8390.433
Q40.0990.4010.0460.4660.8860.2160.8281.1810.3521.3211.8180.424
Q5(Highest)0.0840.3280.0420.4580.8790.2130.7551.0820.3361.2051.6490.412

All four frequency groups show an overall trend of lower error at higher reliability, with the highest reliability quantile, Q5, achieving the lowest error on all three metrics. Relative to Q1, Q5 reduces MAE by approximately 51.9%, 14.2%, 13.5%, and 22.4%, respectively. These results indicate that the reliability–error relationship is not solely an aggregate effect arising from pooling regions with different crime occurrence frequencies: reliability retains its ability to distinguish forecast error levels within groups of regions with similar occurrence frequencies.

Q3: No. Gradients are not stopped through the reliability function in Eq. (6). Gradients from the prediction loss can backpropagate through the routing computation in Eq. (7) and the reliability function in Eq. (6) to the network that produces the structural posteriors.

Our intended meaning was that reliability is computed by a fixed function of the posterior’s support strength and posterior decisiveness, with no separate trainable parameters or additional prediction network. This does not imply that reliability is unaffected by the training loss. We thank the reviewer for pointing out this imprecise wording and will revise the statement “it cannot be optimised to fit the loss” in the Introduction accordingly.

Dear Reviewer, thank you very much for your valuable comments on our manuscript. We have submitted a point-by-point response, but have not yet received your further feedback. We fully understand that you are busy, and would be very grateful if you could respond at your convenience to let us know whether you need any additional clarification or materials from us. Thank you sincerely for your support.

Dear reviewer, thank you for your valuable comments. We have submitted a point-by-point response but have not yet received further feedback. Please let us know at your convenience whether you need any additional clarification or materials. Thank you sincerely for your support.

Dear reviewer, We have submitted our point-by-point response but have not yet received your further feedback. We fully understand that you are busy. If it is convenient for you, we would be very grateful if you could reply to let us know whether we need to provide any additional clarification or materials. Thank you sincerely for your support.

Comments on the paper

1 comment

Dear reviewers,

The authors have submitted their response to your review comments. Please kindly take a moment to read their response and assess whether your concerns have been adequately addressed. Further discussion with the authors to reach a deeper mutual understanding would be greatly appreciated.

Thanks, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.