acceptodds
Under review as a conference paper at ICLR 2027

SAFE-Align: Alignment-Conditioned Evidence Routing for Multimodal Manipulation Detection and Grounding

Abstract

Multimodal manipulation detection requires reasoning over image–text pairs whose semantic correspondence can vary substantially across samples. Existing methods typically use multimodal fusion mechanisms whose aggregation is not explicitly conditioned on the sample-level image–text alignment regime. We challenge this assumption and propose SAFE-Align, an alignment-conditioned framework that treats image–text correspondence as a sample-level routing prior rather than an authenticity score. SAFE-Align combines a discrete alignment state, obtained by quantizing global image–text similarity and represented by a learned state embedding, with the original continuous similarity to characterize the correspondence regime, and uses this alignment prior to route complementary global-interaction, local-spatial, and text-only evidence. To preserve fine-grained manipulation cues, it further models bidirectional patch–token correspondence and captures local agreement and discrepancy cues for image and text grounding. On the official DGM4 split, SAFE-Align achieves 95.63% AUC, 88.74% accuracy, and 89.63% mAP, improving over ASAP by 1.25, 1.03, and 1.10 percentage points, respectively. Alignment-conditioned routing also improves AUC by 1.55 points over uniform fusion and by 1.97 points over a learned gate without the alignment prior. These results support explicitly conditioning multimodal evidence routing on cross-modal alignment rather than relying on a fixed fusion pattern.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Manipulation Detection, Multimodal Grounding, Cross-Modal Alignment, Evidence Routing, 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_H4UaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SAFE-Align uses image-text similarity to condition a gate over global, local, and text-only evidence for multimodal manipulation detection. A bidirectional patch-token branch handles grounding. The method reports improved detection and manipulation-type recognition on DGM4.

Critical strengths
  1. Using image-text alignment to guide evidence fusion is a reasonable idea, motivated by the observation that semantic consistency alone does not establish authenticity.

  2. The method improves detection and manipulation-type recognition over the reported baselines on DGM4. Ablations against uniform fusion and an unconditioned gate provide evidence that alignment conditioning is useful.

Critical weaknesses
  1. The evaluation is limited to DGM4. Although the paper covers four tasks, they all use the same dataset. There is no evaluation on another dataset or unseen manipulation sources, leaving the generality of the approach untested. Results are averaged over three runs, but no standard deviations are reported.

  2. The paper lacks comparisons with the latest methods, such as MaLSF [1] (CVPR 2026). Without these comparisons, SAFE-Align’s competitiveness against recent advances remains unclear.

  3. The main motivation is not directly tested. The gate ablations support an overall benefit, but the paper does not report performance separately for weakly aligned authentic pairs and strongly aligned manipulated pairs, or show how routing weights change across alignment levels.

  4. The grounding results are mixed. Text F1 is lower than ASAP (75.73% vs. 76.52%), and image [email protected] is lower than UFAFormer (76.38% vs. 79.20%). Although the grounding components improve over their ablated variants, their advantage over existing methods is less convincing than the detection gains.

[1] Zizhao Chen, Ping Wei, Ziyang Ren, Huan Li, and Xiangru Yin. Bridging Pixels and Words: Mask-Aware Local Semantic Fusion for Multimodal Media Verification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 26561-26571, 2026.

Questions for authors
  1. Can you provide results on another dataset or unseen manipulation sources to assess generalization beyond DGM4?
  2. How does SAFE-Align compare with recent methods?
  3. Does alignment-conditioned routing improve performance specifically on weakly aligned authentic pairs and strongly aligned manipulated pairs? Can you show how routing weights vary across alignment levels?
5 comments
AuthorsOpenReview ↗

Response to Reviewer H4Ua

We thank the reviewer for recognizing the motivation for alignment-guided fusion. The revision adds recent baselines, three-run standard deviations, direct tests of the motivating alignment regimes, controlled routing ablations, and grounding-error analysis. These additions strengthen the evidence for sample-dependent evidence aggregation.

Q1: Generalization beyond DGM4 and run-to-run variability.

A1: The revised evaluation strengthens the DGM4 evidence through three independent runs across all 12 metrics and targeted analysis of different correspondence regimes. Table 1 reports AUC 95.63±0.69%, ACC 88.74±1.91%, mAP 89.63±0.41%, and IoU50 91.73±1.50%. Table 5 additionally uses identical seeds (333, 555, 777) across four controlled configurations.

The alignment-stratified experiments test challenging cases within DGM4, including authentic pairs with weak correspondence and manipulated pairs with strong correspondence. Full improves performance in both cases, as detailed below. Cross-dataset and unseen-source generalization remain to be evaluated; the revised conclusions identify DGM4 as the current empirical scope.

Q2: Comparison with recent methods, particularly MaLSF.

A2: We have added MaLSF and EMSF to Table 1:

MethodAUC (%)CF1 (%)IoU50 (%)
EMSF95.1183.6088.35
MaLSF, variant 195.5683.8488.83
MaLSF, variant 295.6083.9388.95
SAFE-Align95.63±0.6984.88±0.5391.73±1.50

SAFE-Align achieves the highest reported AUC, CF1, and IoU50 among these methods. Compared with the stronger MaLSF value for each metric, its CF1 and IoU50 improve by 0.95 and 2.78 percentage points, respectively. Against EMSF, the gains are 0.52 AUC, 1.28 CF1, and 3.38 IoU50 points.

These results establish competitiveness against recent advances, particularly in class-wise manipulation recognition and moderate-overlap localization. Table 1 also reports the remaining metrics: EMSF achieves higher mAP, and MaLSF achieves higher strict grounding accuracy. Baseline results retain their original-paper provenance.

Q3: Does routing improve the specific alignment regimes motivating the method?

A3: Yes. Table 7 directly evaluates these cases in one matched run:

ConfigurationWeakly aligned authentic TNR (%)Strongly aligned manipulated TPR (%)
Neither73.3641.35
Direct-only73.1340.38
Routing-only76.0957.69
Full80.5068.27

With the direct alignment-feature path retained, routing improves TNR by 7.37 points and TPR by 27.89 points. Without direct injection, routing also improves both metrics. This directly supports the motivation: correspondence strength informs evidence interpretation without determining authenticity.

Full also achieves the highest AUC in every alignment stratum:

Alignment levelNeitherFullGain
Level 094.0995.06+0.97
Level 193.1094.50+1.40
Level 281.1286.74+5.62
Level 379.3784.45+5.08

The strongest gains occur in Levels 2–3. The strongly aligned manipulated subset contains 104 samples.

Table 5 provides complementary three-seed evidence: routing improves mean AUC by 0.67 points without direct injection and 0.97 points with it. Table 8 reports branch-wise discrimination by alignment level. These diagnostics evaluate branch effectiveness; routing-weight summaries remain a separate analysis.

Table 10 further tests the alignment representation. Combining continuous similarity and discrete states achieves AUC/ACC/mAP of 95.13/90.06/89.90%, improving over the respective stronger single-signal control by 0.90/0.50/0.65 points.

Grounding performance and error analysis.

SAFE-Align’s IoU50 exceeds UFAFormer by 6.34 points (91.73% vs. 85.39%), demonstrating higher localization success at moderate overlap. Table 9 examines 3,427 predictions below IoU 0.75: small-region errors and partial coverage occur in 65.80% and 46.13%, respectively, with overlapping categories. Figure 12 illustrates how incomplete coverage and boundary displacement reduce strict overlap.

We also corrected text F1: Table 1 now reports precision 76.51±0.34%, recall 74.48±1.51%, and F1 75.48±0.60%. The revised conclusions distinguish detection and moderate-overlap localization strengths from the remaining challenges in boundary precision and text F1.

We thank the reviewer again. To support reproduction, we will release the implementation, configurations, and trained weights after the double-blind review period, with the repository link in the final manuscript.

Area_Chair_VaEfOpenReview ↗

The authors' rebuttal has been submitted

Dear reviewer, this is a kind reminder that the authors have posted their responses to your reviews. Please review their rebuttal, engage in any necessary follow-up discussion with the authors, and confirm whether your concerns have been addressed by updating your review/scores. Thank you for your colaboration.

Thanks for the detailed response and the great ablations for Q3! However, is there really no way for the authors to add just one more dataset, such as Weibo17 or Weibo21? I totally understand they don't satisfy all the requirements of your full framework, but they can still serve as a nice validation experiment, right? For example, you could simply remove the bbox detector&loss and just test the core routing/detection part. Honestly, relying on a single dataset makes it really tough to justify accepting this paper to the main conference.

Reviewer_WrmLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method (SAFE-Align) for multimodal manipulation detection and grounding. It uses discrete alignment levels and continuous image–text similarity to condition a gate over global, local, and text-only evidence. A bidirectional patch–token branch provides additional grounding features. On DGM4, the method improves detection and manipulation-type recognition, while grounding results are mixed.

Critical strengths

The paper addresses a meaningful distinction between semantic alignment and authenticity, motivating sample-dependent evidence aggregation rather than treating similarity as a direct manipulation score.

The method reports improvements across several classification metrics and includes component and fusion ablations averaged over three runs. The paper also acknowledges its weaker results on stricter grounding metrics.

Critical weaknesses

The central routing claim is not sufficiently isolated. The alignment representation both conditions the gate and enters the final classification representations directly through Eq. 12. The ablations need to distinguish gains from routing from those due to adding alignment features to the classifier.

The construction and interpretation of the alignment signal are insufficiently validated. Projection initialization, threshold estimation, and stability under end-to-end training are unclear. No alignment-stratified analysis demonstrates that the router selects more reliable evidence under different correspondence regimes.

The methodological contribution mainly combines established components. Stronger controlled comparisons and routing diagnostics are needed to establish the specific value of alignment-conditioned aggregation.

Evaluation is restricted to DGM4. Although three-run averages are reported, variability and computational costs are missing. Grounding improvements are not consistent across metrics, and the full model’s reported text F1 appears inconsistent with its precision and recall.

Questions for authors

Does alignment-conditioned routing still help when all variants retain the same direct alignment-feature path and comparable parameter counts?

How are similarity projections initialized and thresholds estimated, and do alignment levels remain semantically stable during training?

Can you report routing weights and branch effectiveness by alignment level, together with continuous-only and discrete-only controls?

Can you clarify the text-F1 discrepancy and provide per-run variability and evaluation under a shifted data or manipulation distribution?

I would like to raise my score if these questions resolved.

5 comments
AuthorsOpenReview ↗

Response to Reviewer WrmL

We thank the reviewer for recognizing the distinction between alignment and authenticity. The revision adds a matched three-seed routing/direct-injection ablation, alignment-stratified and branch-wise diagnostics, continuous/discrete controls, standard deviations, and complexity and timing measurements. We also correct text F1 and clarify alignment construction.

Q1: Does routing help when the direct alignment-feature path is retained?

A1: Yes. Table 5 independently controls routing and direct injection using identical seeds (333, 555, 777) and training settings. Results are mean ± sample standard deviation:

ConfigurationAUC (%)mAP (%)
Neither94.25±0.6788.69±0.43
Direct-only94.66±0.6387.93±0.40
Routing-only94.92±0.5889.01±0.37
Full95.63±0.6989.63±0.41

With the direct path retained, routing improves AUC by 0.97 points and mAP by 1.70 points. Without direct injection, it improves AUC by 0.67 points. Direct injection alone increases AUC but decreases mAP. These results support routing’s contribution beyond adding alignment features to the classifier.

The ablation controls the two functional paths under a common protocol. Table 11 reports full-model complexity; individual-variant parameter counts have not been measured.

Q2: Alignment construction, thresholds, and training stability.

A2: A frozen pretrained vision-language encoder precomputes cosine similarities. Quantile thresholds are estimated exclusively from training-set similarities before optimization and fixed thereafter. The same encoder and thresholds assign validation and test samples to alignment levels.

Similarity values and level assignments remain fixed throughout training. Level embeddings and the downstream projection are optimized jointly with task objectives, learning how to represent and use these consistent correspondence regimes. The projection comprises a linear layer, layer normalization, and GELU. Its initialization remains an implementation detail to specify.

The levels characterize correspondence strength independently of authenticity, preserving weakly aligned authentic and strongly aligned manipulated cases.

Q3: Alignment-stratified diagnostics and representation controls.

A3: Table 7 evaluates these two motivating cases in one matched run:

ConfigurationWeakly aligned authentic TNR (%)Strongly aligned manipulated TPR (%)
Neither73.3641.35
Direct-only73.1340.38
Routing-only76.0957.69
Full80.5068.27

Routing improves both metrics with or without direct injection. With the direct path retained, it increases TNR by 7.37 points and TPR by 27.89 points. Full also achieves the highest AUC across all four alignment strata. The strongly aligned manipulated subset contains 104 samples.

Table 8 reports branch-wise discrimination by alignment level, examining each source’s informativeness. Explicit routing-weight summaries remain a separate analysis.

Table 10 provides the requested single-run controls:

SignalAUC (%)ACC (%)mAP (%)
Continuous-only93.7389.5688.71
Discrete-only94.2388.8289.25
Both95.1390.0689.90

Combining both improves AUC, ACC, and mAP by 0.90, 0.50, and 0.65 points over the respective stronger single-signal variant. These controls support conditioning aggregation on the correspondence regime.

Q4: Text F1, variability, computational cost, and distribution shift.

A4: We corrected the text-grounding inconsistency. Table 1 reports precision 76.51±0.34%, recall 74.48±1.51%, and F1 75.48±0.60%. The displayed precision and recall yield a harmonic mean of approximately 75.48%. Table 2’s Full row is also corrected.

Table 1 now reports three-run standard deviations for all 12 metrics. New Table 11 quantifies complexity and inference efficiency:

Total parametersTrainable parametersForward latency
482.33M269.47M17.85 ms/sample

Latency is measured on one RTX 4090, batch size 1, over 200 iterations after 50 warm-up runs, excluding loading, tokenization, and input transfer.

The evaluation establishes effectiveness on DGM4; additional datasets and unseen manipulation sources are the next generalization tests. Table 9 and Figure 12 identify small-region errors and partial coverage as the most frequent grounding failures. The revised conclusion distinguishes moderate-overlap localization gains from precise boundary recovery.

We thank the reviewer again. We will publicly release the implementation, configurations, and trained weights after the double-blind review period, with the repository link in the final manuscript.

Area_Chair_VaEfOpenReview ↗

The authors' rebuttal has been submitted

Dear reviewer, this is a kind reminder that the authors have posted their responses to your reviews. Please review their rebuttal, engage in any necessary follow-up discussion with the authors, and confirm whether your concerns have been addressed by updating your review/scores. Thank you for your colaboration.

Thank you for the detaild responses. The additional experiments address most of my question, but some weaknesses are still existing. Two main points remain:

  1. The evaluation is still limited to DGM4.

  2. Route weights are still not report which are key design of this paper.

Please clarify if there is any misunderstanding.

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

SAFE-Align uses global image-text alignment to condition a gate over global, local, and text-only evidence for manipulation detection. A bidirectional patch-token component provides grounding cues. Experiments on the DGM4 split report detection, manipulation classification, and image/text grounding measures.

Critical strengths
  1. The alignment-conditioned routing has a concrete, testable motivation: a globally aligned pair can still contain a local manipulation.

  2. Table 4 isolates the alignment prior: AUC is 95.63 versus 94.08 for uniform fusion and 93.66 for a learned gate without the prior.

Critical weaknesses
  1. Table 1 imports baseline results from prior papers rather than rerunning the strongest competitors under matched training conditions, which weakens the interpretation of small cross-method margins.

  2. Detection and recognition gains do not extend uniformly to grounding: mIoU 77.53 is below UFAFormer 78.33, and mIoU75 76.38 is below 79.20. The conclusions should distinguish these tasks.

Questions for authors
  1. Can the authors provide a matched-protocol rerun or a precise comparison audit for ASAP and UFAFormer, together with variation across the three runs?

  2. What error categories explain the weaker high-IoU image-grounding results?

4 comments
AuthorsOpenReview ↗

Response to Reviewer gkee

We thank the reviewer for recognizing the motivation for alignment-conditioned routing and its supporting ablations. Building on the original manuscript, the revision adds: (i) three-run standard deviations and recent baselines EMSF/MaLSF; (ii) a matched three-seed routing/direct-injection factorial ablation; (iii) alignment-stratified, branch-wise, and continuous/discrete representation analyses; and (iv) quantitative grounding-error analysis with visual examples. We address the concerns below.

Q1: Baseline comparability and statistical reliability.

A1: ASAP, UFAFormer, and SAFE-Align are evaluated on the same official DGM4 benchmark, public splits, tasks, and metrics. The publicly specified training settings are consistent with our protocol. This supports benchmark-level comparison, while we distinguish it from retraining every baseline in our implementation.

Faithful reproduction faces specific obstacles. ASAP requires per-sample large-model-assisted refinement, but the released resources do not fully provide its associated data-construction pipeline. Reconstructing it would require substantial processing and introduce unspecified implementation choices. For UFAFormer, the available implementation resources and reported hyperparameters are insufficient to reconstruct the complete training configuration faithfully. We therefore report the original papers’ results rather than label an approximate reproduction as a matched experiment. Their provenance is explicit in revised Table 1.

The original manuscript reported three-run means; the revision additionally reports standard deviations:

MethodAUC (%)ACC (%)mAP (%)
UFAFormer (published)93.8186.8087.85
ASAP (published)94.3887.7188.53
SAFE-Align95.63±0.6988.74±1.9189.63±0.41

The reported gains over ASAP are 1.25 AUC, 1.03 ACC, and 1.10 mAP points. These are benchmark-level mean differences.

Crucially, the routing contribution is now tested under controlled conditions. Table 5 uses the same three seeds (333, 555, 777) and training protocol to separate alignment-conditioned routing from direct alignment-feature injection:

ConfigurationAUC (%)mAP (%)
Neither94.25±0.6788.69±0.43
Direct-only94.66±0.6387.93±0.40
Routing-only94.92±0.5889.01±0.37
Full95.63±0.6989.63±0.41

Routing improves mean AUC both without direct injection (+0.67 points) and with it (+0.97 points), supporting its contribution beyond simply supplying alignment features to the classifier.

Additional matched-run diagnostics in Table 7 show that Full increases TNR on weakly aligned authentic pairs from 73.36% to 80.50%, and TPR on strongly aligned manipulated pairs from 41.35% to 68.27%, relative to Neither. Table 8 reports branch-wise discrimination across alignment levels. Table 10 further shows that combining continuous similarity and discrete states improves AUC by 0.90 points over the stronger single-signal variant.

Q2: Why is high-IoU image grounding weaker?

A2: SAFE-Align achieves IoU50 of 91.73±1.50%, exceeding UFAFormer’s 85.39% by 6.34 points. Thus, more manipulated regions are successfully localized at moderate overlap. Strict boundary recovery is a separate requirement: our mIoU and IoU75 are 77.53±0.62% and 76.38±0.29%, versus 78.33% and 79.20%.

Appendix Table 9 analyzes 3,427 predictions with IoU below 0.75:

Error categoryCountShare (%)
Small-region error2,25565.80
Partial coverage1,58146.13
Center offset91626.73
Incorrect localization78422.88
Loose boundary40511.82
Other/unclassified2126.19

Categories are non-exclusive. Small regions, occupying at most 2% of image area, are the most frequent condition, followed by incomplete coverage. Oversized predictions occur but are not dominant.

For small regions, modest boundary displacement can substantially reduce IoU. Partial coverage and loose boundaries can preserve overlap above 0.5 while failing the 0.75 threshold; center-offset cases have particularly low mean IoU (0.0471). Figure 12 illustrates these patterns. These diagnostics explain our localization limitations without attributing the cross-method gap to unobserved UFAFormer predictions.

The results identify a clear strength: strong manipulation detection combined with higher moderate-overlap localization success. The remaining challenge concerns precise box extent and boundary recovery.

We thank the reviewer again for helping strengthen the evidence and clarify the task-specific conclusions. To facilitate reproduction, we will publicly release the implementation, experiment configurations, and trained model weights after the double-blind review period, and provide the repository link in the final manuscript.

Area_Chair_VaEfOpenReview ↗

The authors has submitted their rebuttal

Dear reviewer, this is a kind reminder that the authors have posted their responses to your reviews. Please review their rebuttal, engage in any necessary follow-up discussion with the authors, and confirm whether your concerns have been addressed by updating your review/scores. Thank you for you colaboration.

Thank you for the additional experiments and clarifications. I have also read the other reviewers’ comments and the authors’ responses. The ablations and routing-weight analysis strengthen the evidence for alignment-conditioned routing, and the grounding-error analysis addresses my question about localization limitations. However, validation beyond DGM4 is still pending. This limitation leaves uncertainty about the robustness and generality of the reported gains. I am therefore keeping my current rating.

AuthorsOpenReview ↗

Response to Reviewer gkee

We thank the reviewer for carefully examining our additional experiments and responses, and for acknowledging that the controlled ablations, routing-weight analysis, and grounding-error analysis have addressed the corresponding concerns.

We understand that the remaining uncertainty regarding evaluation beyond DGM4 was the primary reason for maintaining your rating. We would like to provide an important update: we have now completed the additional experiments on Weibo17 and incorporated the quantitative results and corresponding analysis into our latest revised manuscript.

1. Additional Validation on Weibo17

Following the suggestions raised during the discussion, we extended SAFE-Align to Weibo17, a Chinese multimodal fake-news detection benchmark with a different data distribution and task setting from DGM4.

Since Weibo17 does not provide manipulation-type, bounding-box, or token-level grounding annotations, we adapted SAFE-Align to binary authenticity detection. The alignment prior, Global/Local/Text evidence branches, and alignment-conditioned routing mechanism were retained, while task-specific prediction heads and losses requiring unavailable annotations were removed.

The additional comparison is summarized below.

MethodACC (%)Fake F1 (%)Real F1 (%)
FND-CLIP (ICME'23)90.790.890.7
BMR (AAAI'23)91.891.490.4
COOLANT (MM'23)92.392.592.0
MIMoE-FND (WWW'25)92.892.892.7
RaCMC (AAAI'25)91.591.791.4
FSRU (AAAI'24)90.190.689.5
MaLSF (CVPR'26)93.493.393.5
SAFE-Align (Ours)94.293.593.7

SAFE-Align achieves 94.2% accuracy, outperforming the strongest reported baseline, MaLSF (93.4%), by 0.8 percentage points. It also improves the fake-news and real-news F1 scores over MaLSF by 0.2 percentage points each.

These results provide additional evidence that the task-adapted SAFE-Align framework achieves competitive detection performance beyond the DGM4 benchmark. Importantly, this is an evaluation on a second dataset, rather than zero-shot transfer from DGM4. We therefore do not claim that this experiment establishes robustness to arbitrary distribution shifts or generalization of the complete four-task framework.

2. Evidence Supporting the Core Mechanism

We appreciate your recognition that our additional analyses have strengthened the evidence for alignment-conditioned routing.

In particular, the matched three-seed factorial ablation demonstrates AUC improvements of 0.67 and 0.97 percentage points from enabling routing without and with direct alignment-feature injection, respectively.

The learned routing-weight analysis further shows that, across four alignment levels, the Text weight decreases from 30.97% to 10.07%, while the Local weight increases from 11.32% to 23.02%. These results directly characterize the learned evidence-allocation behavior and complement the controlled performance ablations.

Together with the newly completed Weibo17 evaluation, the revised manuscript now provides evidence from both controlled mechanism analysis on DGM4 and additional detection experiments on a separate dataset.

3. Request for Reconsideration

In your latest comment, you indicated that the absence of validation beyond DGM4 was the principal remaining concern underlying your decision to maintain the rating. This experiment has now been completed, and its results are included in the revised manuscript rather than remaining pending future work.

We acknowledge that broader robustness evaluation remains valuable. Nevertheless, we believe the additional Weibo17 results directly address the specific limitation you identified, while the previously completed ablations and grounding-error analysis address your other technical concerns.

Given that the principal remaining concern has now been addressed with new experimental evidence, we respectfully ask whether you would reconsider your assessment and rating in light of the completed revision.

We sincerely appreciate your constructive feedback and the careful attention you have given to our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.