acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Geometric Priors in Novel View Synthesis through Multi-Axis Reliability Fields

Abstract

Geometric priors are widely used in novel view synthesis (NVS) to incorporate structural guidance. However, recent results show that this conditioning often provides marginal gains under strong backbones. Does this mean that geometric priors are no longer important once the model can already infer cross-view correspondence on its own? Our controlled corruption experiments show otherwise: geometry remains useful, but its value is under-utilized when reliable and unreliable guidance are not distinguished. Motivated by this finding, we introduce Reliability-Calibrated Geometric Attention (RCGA), which models geometric reliability as a multi-axis field to control how explicit geometric evidence modifies the backbone's intrinsic correspondence. The field identifies where and under which model states geometry contributes complementary information. Specifically, the partition-preserving product-of-experts uses the predicted field to calibrate geometric evidence in joint attention for selective correspondence refinement. A comparative oracle supervises the field by measuring the incremental utility of geometry beyond the backbone during training. Across RealEstate10K and ScanNet, RCGA consistently improves epipolar, depth-projection, and coordinate priors over a strong diffusion backbone, achieving state-of-the-art performance while suppressing unreliable geometric guidance.

Primary area:
generative models
Keywords:
Novel View Synthesis, Diffusion Models, Cross-View Image Generation, Conditional Generation, Geometric Priors

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_Qy9KOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RCGA, a learned gate that balances the frozen VIVID backbone's own correspondence against external geometric priors. The gate controls how much the geometric prior reshapes the backbone attention and is supervised by an oracle that compares geometric and backbone correspondence. RCGA improves over VIVID on RealEstate10K and ScanNet and is more robust to corrupted geometry than fixed injection.

Critical strengths
  1. The motivation and idea are clear. Explicit geometry often adds a small improvement to strong NVS backbones, and judging its value relative to the backbone's own correspondence is a reasonable fix.
  2. As a lightweight adapter on a frozen backbone, RCGA consistently outperforms VIVID and degrades much more gracefully than fixed injection when the geometry is corrupted. This shows that the gating is useful in practice.
Critical weaknesses

The paper's main claim is contradicted by its experiment. It attributes the limited gain of geometry to reliable and unreliable cues offsetting each other, yet with clean ground-truth geometry, naive injection still barely beats pose-only while RCGA clearly does, so the gain of RCGA may come from its update rule rather than from modeling reliability.

Questions for authors
  1. With clean ground-truth geometry, naive injection barely improves over pose-only. What causes this, if not unreliable cues? Does RCGA with a constant gate, i.e. the same update rule without reliability modeling, already recover most of the gain?
  2. How is the scale of the Depth Anything V2 depth aligned with the RealEstate10K poses before reprojection?
5 comments

Thank you for this insightful and constructive commentt, which encouraged us to further investigate and better understand our experimental findings.

Response to Weakness and Question 1

To directly investigate whether RCGA's gains arise from its update rule or reliability modeling, we added a constant-gate PoE baseline to the controlled corruption analysis (Sec. 4.2, Fig. 6). It uses the same update rule as RCGA, differing only in that its reliability is fixed rather than dynamically predicted. The results are as follows:

Method0%25%50%75%
Pose-only20.9420.9620.9520.94
Naive Depth21.1820.8920.4119.72
Constant-gate PoE21.6421.3420.8020.21
RCGA-Depth22.3022.0821.7021.33

The results show that RCGA's gains do not come solely from reliability modeling, but also from how geometric evidence is integrated with backbone attention. The contribution of the update rule remains relatively stable across corruption levels, whereas the additional gains from dynamic reliability increase as corruption becomes more severe. Importantly, even when all geometric correspondences are correct, dynamic reliability provides additional benefits.

This also explains why naive injection provides only marginal gains even with clean ground-truth geometry. Geometric correctness does not necessarily imply geometric utility. Even when all geometric correspondences are correct, they may provide redundant information that the backbone has already implicitly learned, rather than additional complementary evidence. A fixed injection strategy cannot distinguish between these cases or adapt the geometric influence to the varying requirements of different spatial queries and denoising stages. As illustrated by the reliability visualizations in Fig. 8, RCGA predicts substantially different geometric influence across spatial queries and denoising stages.

Therefore, it reflects two distinct limitations of native injection:

(1) Under heterogeneous geometric quality, the benefits of reliable correspondences can be offset by the negative effects of corrupted ones.

(2) Even with clean geometry, a fixed gate cannot fully exploit complementary geometric evidence according to its varying utility across queries and denoising states.

Response to Question 2

For RCGA-Depth, we follow VIVID's depth preprocessing procedure. Specifically, the predicted depth is inverted, normalized by its per-image maximum, and standardized using fixed mean and standard deviation statistics. As in VIVID, no additional metric-scale alignment with RealEstate10K camera translations is performed, since the monocular depth predictions are not up to scale with the dataset poses. The resulting depth is therefore used as an approximate geometric prior rather than metrically calibrated geometry, whose inherent uncertainty motivates our reliability-aware integration.

We sincerely thank the reviewer again for raising these important questions. They prompted us to conduct an additional controlled experiment that more clearly disentangles the contributions of geometric integration and reliability calibration, as well as to clarify the geometric assumptions underlying our depth-based correspondence construction. We believe these revisions provide a more precise explanation of the motivation and effectiveness of RCGA.

I would like to thank the authors for their detailed rebuttal and additional experiments. Most of concerns are addressed. I have small question about features chosen as input: what motivates the choice of these seven statistics, e.g., what does Δtop\Delta_{top} add beyond H(Pi)H(P_i)? Maybe those seven statistics provide redundant information. (No experiments required)

We thank the reviewer for this important question. Our central motivation is to determine whether geometric evidence can provide useful information beyond what the backbone has already learned. Accordingly, the seven statistics are selected to characterize complementary aspects of geometric reliability from three perspectives: the intrinsic correspondence state of the backbone, the characteristics of geometric evidence, and the agreement between the two.

Specifically, H(P_i) and Delta_top capture different aspects of backbone correspondence uncertainty: the former measures the overall dispersion of the correspondence distribution, while the latter measures the discriminability between the top two matching candidates. Similarly, H(P_g) characterizes the concentration of geometric correspondence, whereas C_g provides complementary information about geometric support. Finally, JS(P_i || P_g) measures the distributional discrepancy between backbone and geometric correspondence, while O_ig captures their support overlap. M_src describes the backbone's source-attention allocation, which is particularly relevant to our partition-preserving update.

We have recognized that these statistics are not necessarily statistically independent. To examine their complementary contributions, we conducted a leave-one-out ablation by individually masking each statistic while keeping the field prediction function unchanged (Appendix Table 6). Removing each individual statistic degrades performance, suggesting that none is entirely redundant given the remaining six. Notably, removing either agreement statistic causes a relatively larger degradation, highlighting the importance of explicitly modeling the relationship between backbone and geometric correspondence.

We thank the reviewer again for this helpful suggestion and the constructive discussion throughout the review process.

Reviewer_TCDTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work argues that explicit geometric priors yield only marginal gains in strong generative NVS backbones because reliable and unreliable geometric guidance are integrated uniformly. It proposes Reliability-Calibrated Geometric Attention (RCGA). RCGA expresses epipolar, depth-projection and coordinate cues as correspondence distributions over source keys and fuses them into the frozen backbone’s source attention through a partition-preserving product-of-experts update. The fusion is weighted by a reliability field predicted per layer, head, query and noise level from statistics comparing backbone and geometric correspondence. During training, this field is supervised by a comparative oracle based on clean-target features. Built on a frozen VIVID backbone, RCGA reports state-of-the-art PSNR and FID on RealEstate10K and ScanNet.

Critical strengths
  1. The motivation is sound. Asking whether geometry still contributes once a strong backbone already forms its own cross-view correspondence is a genuinely interesting question. Calibrating geometric influence at the attention level, using statistics that compare backbone and geometric correspondence, is a principled way to address it.

  2. The method is well constructed. Representing all three priors through a common correspondence-distribution interface lets one fusion mechanism serve every geometric cue. The partition-preserving PoE update reduces exactly to the backbone when reliability is zero, which the matching pose-only rows of Table 2 confirm (21.11 vs. 21.10). The oracle supervision requires no extra inputs at inference, and the component ablations (Table 4) show a consistent progression as reliability modeling becomes finer.

Critical weaknesses
  1. “Naive injection” is not clearly defined in the paper. This possibly makes the central controlled comparison a straw man. Table 2, the paper’s main evidence that conventional injection yields marginal gains, compares RCGA only against this undefined baseline. Its +Depth result (21.29 PSNR / 2.99 FID) is identical to the “Fixed geometry” row of Table 4. This suggests it is a fixed-weight log-prior added to the frozen backbone’s attention, with no stated tuning of that weight. Established geometry-guided methods such as GenWarp and epipolar-attention approaches instead learn to use geometry through training, so the claim that conventional injection is ineffective is not tested against conventional injection.

  2. It is unclear whether the improvements over baselines come from a stronger backbone or a genuinely better architecture. RCGA trains only lightweight parameters on top of a frozen VIVID model, which itself outperforms every other baseline by a wide margin. On RealEstate10K mid-range, 3.93 dB of RCGA’s 4.91 dB advantage over GenWarp is already present in VIVID alone (Table 1). The only backbone-matched comparison is against VIVID (+0.98 dB). Geometry-guided and adaptive-integration methods (e.g., WAVE, NVS-Solver, MoAI) are not reimplemented on the same backbone, so the architectural contribution relative to existing geometry-injection designs remains unestablished.

  3. The paper’s own data undercuts its cancellation explanation. The central claim is that geometry yields marginal gains because the benefit of reliable correspondence is cancelled by unreliable correspondence. Yet in Fig. 6(c), at 0% corruption with ground-truth geometry, naive injection improves only about 0.15 dB over pose-only, while RCGA improves about 1.3 dB. If cancellation were the cause, naive injection should already work when all geometry is reliable. Table 3 points the same way: on clean correspondences, naive injection achieves a higher correspondence benefit than RCGA (+0.0581 vs. +0.0491), yet RCGA generates substantially better images. RCGA’s gain therefore appears to stem from how geometry is combined with backbone attention, not from separating reliable from unreliable guidance as the paper claims.

  4. The manuscript appears to have been prepared in haste:

  • Citations use \cite instead of \citep throughout, so references run into the text (e.g., “the target viewpoint Wiles et al. (2020)”).
  • Several citations do not support their claims. A visible-infrared UAV detection dataset (Liu et al., 2026) is cited for geometric unreliability, and a satellite-to-street synthesis paper (Li et al., 2027) is cited for depth-based NVS projection.
  • The “Inside RCGA” panel of Fig. 2 appears to be AI-generated.
Questions for authors
  1. How exactly is “naive injection” implemented, and how was its geometric weight chosen? Can the authors compare RCGA against a trained geometry-conditioning baseline (e.g., GenWarp-style or epipolar-attention conditioning) on the same VIVID backbone?

  2. How much of the improvement over Table 1 baselines is attributable to the VIVID backbone? Can the authors evaluate existing geometry-injection or adaptive-integration methods on the same backbone to isolate RCGA’s architectural contribution?

  3. Why does naive injection yield only marginal gains even with clean ground-truth geometry (Fig. 6(c), 0% corruption), and why does it achieve higher clean correspondence benefit than RCGA despite worse generation? How do the authors reconcile this with the cancellation explanation?

  4. Can the authors clarify whether generative AI tools were used to create Fig. 2, and correct the citation format and the citations that do not support their claims?

7 comments

Response to Weakness 3 and Question 3

Thank you for identifying this important issue. We have added a baseline to the controlled corruption analysis to better disentangle the respective contributions of the geometric update rule and reliability calibration. We have also expanded our explanation of why naive injection provides only limited gains under 0% corruption.

To directly investigate whether RCGA's gains arise from its update rule or reliability modeling, we added a Constant-gate PoE baseline to the controlled corruption analysis (Sec. 4.2, Fig. 6). It uses the same update rule as RCGA, differing only in that its reliability is fixed rather than dynamically predicted. The results are as follows:

Method0%25%50%75%
Pose-only20.9420.9620.9520.94
Naive Depth21.1820.8920.4119.72
Constant-gate PoE21.6421.3420.8020.21
RCGA-Depth22.3022.0821.7021.33

The results show that RCGA's gains do not come solely from reliability modeling, but also from how geometric evidence is integrated with backbone attention. The contribution of the update rule remains relatively stable across corruption levels (0.39–0.49 dB over naive injection), whereas the additional gains from dynamic reliability increase with corruption severity (0.66–1.12 dB over Constant-gate PoE). Importantly, even when all geometric correspondences are correct, dynamic reliability provides additional benefits.

This helps explain why naive injection yields only marginal gains even with clean ground-truth geometry (Fig. 6(c), 0% corruption), supporting our central motivation for comparative reliability supervision: geometric correctness does not necessarily imply geometric utility. As illustrated by the reliability visualizations in Fig. 8, the predicted geometric influence varies across spatial queries and denoising stages. This is consistent with prior studies, such as Geometry-Free View Synthesis: Transformers and No 3D Priors and Scene Representation Transformer, which demonstrate that attention-based models can implicitly learn geometric structures. Consequently, clean ground-truth geometry does not automatically guarantee substantial improvements under fixed geometric injection or even Constant-gate PoE.

Therefore, naive injection has two limitations:

  1. Cancellation under heterogeneous geometric quality. When geometric quality varies, the benefits of reliable correspondences can be offset by the negative effects of corrupted ones. Under 50% corruption (Table 3), naive injection provides a +0.0581 benefit on clean queries but incurs a −0.0565 penalty on corrupted queries. In contrast, RCGA achieves +0.0491 and −0.0147, respectively, substantially reducing the harmful influence of corrupted correspondences.

  2. Limited utilization of correct geometric evidence. Even when geometry is clean, naive injection cannot fully exploit complementary geometric evidence throughout the generation process, because different spatial queries and denoising states may benefit from different levels of geometric guidance.

We have added the corresponding experimental results and analysis to the revised manuscript. We sincerely thank the reviewer for this insightful comment, which encouraged us to further investigate and better understand our experimental findings.

Response to Weakness 1 and Question 1

Thank you for raising this important concern. We agree that the original experimental comparisons could be further improved to better isolate the contributions of different geometry-conditioning mechanisms.

For clarity, we have renamed the left-hand column of the original Table 2 as Geometry-encoding baselines, which directly adopts the published VIVID Base-Ablation results. These results serve as external references rather than an independently implemented naive-injection baseline.

For the fixed geometry In Table 4, It uses the same depth-projection prior as RCGA, converts it into a centered geometric log-prior , and directly adds it to the backbone's source-attention logits, without partition preservation.

To provide a more rigorous and controlled comparison, we further introduce Constant-gate PoE, which retains the same partition-preserving PoE update as Full RCGA but replaces its query-, head-, layer-, and noise-dependent geometric modulation with a single shared scalar gate. This comparison isolates the additional benefit of dynamic reliability modeling from that of the partition-preserving update. The revised results are reported below.

VariantQuery-adaptiveHead/LayerNoise-awarePartition preservePSNR ↑FID ↓Brier ↓
Pose baseline××××21.103.02–
Fixed log-prior injection××××21.292.970.3889
Constant-gate PoE×××✓21.352.880.3624
Static calibration×✓×✓21.582.900.2088
Non-spatial dynamic×✓✓✓21.692.820.2053
No explicit noise✓✓×✓22.352.780.1074
No partition✓✓✓×20.743.040.2182
Full RCGA✓✓✓✓22.522.650.1055

These results show that a constant geometric gate provides only limited improvements, whereas progressively more adaptive reliability modeling yields better generation quality and calibration. In particular, query-adaptive reliability provides the main improvement, while noise conditioning brings an additional consistent gain. The comparison between Constant-gate PoE and Full RCGA further demonstrates the benefit of dynamically estimating the relative utility of geometric evidence.

Furthermore, we implemented the Epipolar Attention Bias described in VIVID (Appendix B.2) on the same frozen VIVID backbone. We train its learnable per-head epipolar mixing parameters without introducing RCGA's reliability prediction or comparative-oracle supervision.

MethodPSNR ↑FID ↓
RCGA-Epipolar22.332.58
VIVID-style Epipolar Attention21.642.88

The results demonstrate that reliability-calibrated geometric integration provides additional benefits over conventional learned geometry conditioning, supporting the effectiveness of RCGA in selectively exploiting geometric evidence.

Response to Weakness 2 and Question 2

Thank you for raising this important concern. We agree that the performance gap over previous baselines should be carefully distinguished from the improvement introduced by RCGA itself.

First, we acknowledge that a substantial part of the overall improvement over GenWarp comes from the stronger VIVID backbone.We do not attribute the entire gap to RCGA. Rather, our objective is to investigate how explicit geometric conditioning can be more effectively exploited by a strong pretrained generative backbone.

Second, we have added WAVE as an additional baseline. Its guidance mechanisms are adapted to the frozen VIVID Base-Final backbone, and the resulting model is evaluated under the same RealEstate10K sequence protocol, using identical source images, target camera poses, depth inputs, and evaluation metrics. The results are summarized below:

MethodPSNR ↑FID ↓SSIM ↑LPIPS ↓LPIPS-next ↓CLIPSIM-next ↑
VIVID Base-Final21.103.020.7420.2140.5210.882
VIVID + WAVE21.162.880.7450.2160.4660.906
VIVID + RCGA22.242.630.7620.1850.5180.889

WAVE primarily improves cross-view consistency, whereas RCGA achieves larger gains in target-view reconstruction quality through reliability-aware geometric integration. We emphasize that these methods address complementary aspects of novel view synthesis rather than exactly the same objective.

We further clarify why other methods mentioned by the reviewer are not directly adopted as same-backbone baselines. WAVE is a training-free warp-guided method and can be adapted to VIVID's sampling and attention mechanisms. In contrast, NVS-Solver modulates the sampling process of pretrained video diffusion models and relies on their temporal modeling interfaces. MoAI employs parallel image and geometry generation branches with cross-modal attention distillation. Directly transferring NVS-Solver or MoAI to VIVID's pixel-space NVS architecture would require substantial changes to their original generation frameworks and could alter their core mechanisms, making the resulting comparisons less representative of the original methods.

We selected VIVID because its pixel-space diffusion architecture and source-target attention provide a relatively general and controlled framework for investigating the utilization of explicit geometric conditioning, without introducing additional video-temporal modeling or parallel geometry-generation branches.

Response to Weakness 4 and Question 4

Thank you for your careful examination. We sincerely apologize for the errors in the citations. We have carefully rechecked the manuscript and corrected these issues throughout. Regarding Fig. 2, we acknowledge that the small feature icons and their corresponding formulas in the "Inside RCGA" panel were generated using OpenAI. We apologize for taking this shortcut, which may have given readers a negative impression. We have replaced these elements in the revised manuscript.

Finally, we sincerely appreciate the careful and critical review of our manuscript. Your comments helped us recognize both technical and presentation-related issues and prompted us to re-examine the manuscript more thoroughly. We are grateful for the opportunity to improve the clarity, rigor, and transparency of this work.

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

The paper investigates the limited gains from explicit geometric priors when using a strong generative backbone and proposes Reliability-Calibrated Geometric Attention (RCGA). The method represents epipolar, depth-projection, and coordinate priors as correspondence distributions over source positions. Based on the confidence of the backbone’s intrinsic matching, geometric support, and their agreement, it predicts reliability weights that vary across layers, attention heads, query positions, and diffusion noise levels. These weights control a product-of-experts attention update.

Critical strengths

1.The method achieves substantial performance improvements on the final task across two datasets. Compared with VIVID, RCGA improves PSNR from 17.36 to 18.34 and from 15.21 to 16.96 on the mid-range and long-range settings of RealEstate10K, respectively, and from 16.72 to 17.84 on ScanNet. FID also decreases in all three settings. 2.The paper is carefully written, and its arguments are generally logically structured.

Critical weaknesses

1.The means and reported error terms in all four rows of the Naive column in Table 2 exactly match those in Table 1 of the original VIVID paper. However, VIVID explicitly distinguishes its ablation models from its final model: they are trained for 2182^{18} and 2202^{20} iterations, respectively, with differences in batch size, EMA, and sampling settings. Appendix A.3 of this submission only states that the official frozen 64×6464\times64 model is used, without specifying which checkpoint corresponds to each row of Table 2 or whether the Naive results are taken from VIVID or independently re-evaluated. 2.Equation (14) compares the teacher-feature matching errors obtained by using PgP_g and PiP_i separately, whereas the actual operation is a product-of-experts fusion proportional to PiPgαρP_iP_g^{\alpha\rho}.

Questions for authors
  1. Are the comparisons in Tables 2 and 4 conducted under fully matched conditions?
  2. How is “Net benefit” in Figure 6 computed?
3 comments

Weakness 1

Thank you for pointing out this ambiguity. The four entries in the left column of the original Table 2 were directly adopted from the published geometry-encoding ablation results in VIVID (Base-Ablation configuration). Our intention was to use these results as references to motivate our central investigation: whether RCGA can more effectively exploit geometric conditioning.

Your comment prompted us to more clearly distinguish conventional geometry-conditioned generation baselines from uncalibrated geometric injection within our own framework. Our geometric correspondence encoding was designed not to introduce a stronger standalone geometry encoder, but to construct an explicit probabilistic representation for subsequent reliability estimation and adaptive integration. Accordingly, we have renamed the original "Naive injection" column as "Geometry-encoding baselines" and clarified the source and training configuration of these results. We have also clarified the relevant analysis in the main text and specified in the appendix that our experiments use the official frozen VIVID Base-Final model.

Weakness 2

Thank you for this important observation. This distinction is intentional in our design. During reliability supervision, we compare the individual feature-matching errors of the backbone and geometric experts to construct a gate-independent comparative target. This provides a stable assessment of whether explicit geometry offers better correspondence evidence than the backbone alone, without making the supervision target dependent on the predicted reliability or fusion strength. The actual PoE integration is subsequently performed using the learned reliability during generation.

We thank the reviewer for helping us recognize that the original manuscript did not sufficiently validate this design, we have added a PoE-aware consistency analysis on the held-out RealEstate10K and ScanNet test sets:

  1. Target–benefit consistency: We measure the Spearman correlation between the comparative target and the actual PoE correspondence benefit. Positive correlations 0.36 and 0.31 are observed on both datasets, indicating that the independently constructed supervision target is positively associated with the actual utility of geometric evidence after PoE integration.

  2. Reliability-stratified analysis: We divide queries into five groups according to predicted reliability. The proportion of queries with positive PoE benefits increases from 44.8% to 77.2% on RealEstate10K and from 42.7% to 73.1% on ScanNet, from the lowest- to highest-reliability group. The mean PoE benefit also exhibits an increasing trend across reliability groups.

DatasetGroupReliability (ρ̂)PoE Benefit (ΔPoE)Positive (%)
RealEstate10KQ10.10-0.00344.8
Q20.290.00152.1
Q30.480.00660.4
Q40.680.01268.9
Q50.880.01977.2
ScanNetQ10.10-0.00442.7
Q20.280.00150.6
Q30.470.00557.8
Q40.670.01065.3
Q50.870.01673.1

These results support Eq. (14) as a meaningful comparative surrogate for reliability supervision under the actual PoE integration, rather than an exact PoE-optimal target. We have added the corresponding analysis in Appendix A.5.

Question 1

Thank you for raising this question. The comparisons in Tables 2 and 4 serve different purposes and have different experimental sources. The left side of Table 2 presents externally reported geometry-encoding baselines from VIVID. The right side of Table 2 reports our RCGA results using the official frozen VIVID Base-Final model. Table 4 further conducts component ablations using the same frozen backbone and evaluation protocol, with the depth-projection prior fixed. We have clarified these distinctions in the revised table captions, main text, and Appendix A.3.

Question 2

Thank you for pointing out that the definition of Net Benefit was not sufficiently explicit. To provide an intuitive visualization of the contributions from clean and corrupted queries in Fig. 6, we normalize both correspondence benefits by the total number of valid queries. Net Benefit is therefore computed as the sum of these two contributions. It represents the overall mean correspondence benefit across all valid queries.

Finally, we are truly grateful for the thoughtful perspective you brought to improving the clarity and completeness of this work. Your comments helped us identify aspects that we had taken for granted but had not sufficiently explained, and strengthened the closed-loop validation of the comparative supervision strategy.

Dear Reviewer,

We would like to kindly follow up on our rebuttal and additional experiments. We would greatly appreciate it if you could let us know whether our responses have addressed your concerns or if any questions remain.

Thank you for your time and consideration!

Thank you for the detailed response and the additional experiments. Your rebuttal addressed most of my concerns. I will keep my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.