acceptodds
Under review as a conference paper at ICLR 2027

Hard, Yet Reducible: Controlled Forward Transfer for Synthetic Degradation Curation

Abstract

Selecting synthetic degradations for dense prediction requires an estimate of their training utility, the generalization gain they bring under a finite training budget. Clean and degraded twins share content and labels, suggesting a score based on how much short training reduces the excess error caused by degradation. However, this gap can also shrink when clean performance deteriorates. Measuring the improvement on degraded images alone avoids that confound, but it still credits progress that the same amount of clean training would have produced. We propose the controlled Reducible Degradation Gap (cRDG) for regions defined by degradation type and severity. From a common checkpoint, cRDG runs two budget-matched probes that differ only in one augmentation slot, which holds either a synthetic degradation or a clean augmentation. The score is the gain on held-out degraded images relative to the clean-control probe. Clean harm is a separate feasibility constraint. cRDG reveals a correctable severity band in which training on the degradation yields high controlled gain under the available budget, and the band moves with the predictor, the starting checkpoint, and the training budget. Curation of Reducible Bands (\method) uses cRDG to select synthetic data without changing the predictor. On semantic segmentation and salient object detection, improves representative predictors under matched synthetic-data budgets and training schedules, extends to existing data-generation pipelines, and preserves clean performance. Code and supporting materials will be publicly released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Synthetic Data Curation, Data Valuation, Controlled Reducibility, Forward Transfer, Robust Segmentation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_dMGmOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes controlled Reducible Degradation Gap (cRDG), a metric for estimating the training utility of synthetic degradations through short, matched training probes. Based on this metric, the authors introduce CURB, a framework for selecting useful degradation regions under a fixed training budget. Experiments on semantic segmentation and salient object detection demonstrate improvements over several data selection baselines.

Critical strengths
  1. Well-motivated problem and sound methodology. The paper addresses an important challenge in synthetic data selection: difficult degradations are not necessarily useful for training. The matched-control formulation offers a principled way to estimate degradation-specific utility while accounting for generic training progress and clean-performance degradation.

  2. Improvements across multiple settings. The method improves results on both semantic segmentation and salient object detection, across different architectures, datasets, and data-generation pipelines. The gains are consistent in direction, which suggests practical value beyond a single setup, although their magnitude should be confirmed with repeated runs (see Weaknesses).

Critical weaknesses
  1. Novelty relative to proxy-based search. Estimating the value of an augmentation type and magnitude through short proxy training and held-out evaluation is closely related to augmentation policy search (e.g., AutoAugment, Fast AutoAugment) and proxy-based data-mixture selection (e.g., DoReMi, RegMix). These works are not discussed. The added value of the matched control over such approaches should be clarified, ideally with a compute-matched comparison.

  2. Insufficient experimental validation. Most results come from single runs without variance estimates, while the key margins are small, so the reliability of the improvements is hard to assess. In addition, over 90% of the synthetic budget is drawn from a single region. The paper should report which region is selected, how it ranks under real retraining utility, and how the result compares with an oracle choice, to show that the gains come from correct ranking rather than a fortunate selection.

Questions for authors
  1. Since most of the budget is drawn from a single region, which region is selected in each setting, how does it rank under real retraining utility, and how does CURB compare with an oracle choice? Does spreading the budget across the top few regions change the results?

  2. Since probe cost grows linearly with the number of regions, how does the method scale to substantially larger candidate spaces, such as finer severity bins or more degradation families?

6 comments
AuthorsOpenReview ↗

Response to Reviewer dMGm [1/4]: W1

W1. Relation to augmentation search and proxy-based mixture selection. These methods use model-based feedback to select augmentation policies or data mixtures. AutoAugment evaluates policies using child-model validation accuracy, while Fast AutoAugment uses fixed-model density matching. RegMix predicts validation loss from mixture weights using small-model runs. DoReMi optimizes domain weights using a proxy model's excess loss relative to a separately trained reference.

cRDG instead estimates the additional held-out gain from degradation relative to matched clean training. Two KK-step probes start from the same checkpoint and use the same scenes, labels, and number of updates. Their augmentation slots contain degraded twins or clean augmentations, and the resulting losses are compared on degraded twins held out from probe updates. The subtraction matters because clean training alone reduces held-out loss by a region-dependent amount, which a score without the control would credit to degradation training. Unlike DoReMi's reference-model loss gap, our control compares two matched training interventions. Clean harm is screened separately. We clarify these distinctions in Related Work.

We test the subtraction directly and adapt all four methods to the same gated 32-region pool, synthetic budget, predictor, and final-training recipe. Selection cost includes method-specific training, calibration, evaluation, and search, excluding shared candidate-pool generation. CuRB costs 0.66 final-run equivalents, including 0.59 for probe updates. Results are mean ± sample SD over three paired final-training seeds, with θ0\theta_0 and each selected set fixed (Tab. S8).

SelectionSelection costIn-family mIoU
DgD_g (no control subtraction)0.6650.86±0.15
AutoAugment-style search0.6650.59±0.21
Fast AutoAugment-style search0.0347.39±0.17
DoReMi-style reweighting0.6650.47±0.23
RegMix-style, 66 total runs0.6650.72±0.19
RegMix-style, 128 total runs1.2650.89±0.20
RegMix-style, 256 total runs2.5051.03±0.22
cRDG (CuRB)0.6651.90±0.23

Control ablation. DgD_g removes the clean-control subtraction from the ranking score. Both methods use the same 66 probes and clean-harm constraint, giving identical probe-training compute. The subtraction raises Spearman correlation with full-retraining utility from 0.70 to 0.79 on ACDC and from 0.62 to 0.72 on Dark Zurich (Tab. 6). In-family mIoU improves by 1.04 points on average, with paired gains of 0.95, 1.24, and 0.92 (Tab. 7a).

Adapted baselines. At matched selection cost, CuRB exceeds AutoAugment-style, DoReMi-style, and RegMix-style selection by 1.31, 1.43, and 1.18 mIoU. It remains 0.87 ahead of RegMix-style with 256 runs, which uses 3.8 times the selection cost. CuRB exceeds every compared method in all three paired final-training seeds. Implementation protocols and additional diagnostics are reported in Part 2/4.

AuthorsOpenReview ↗

Response to Reviewer dMGm [2/4]: W1, continued

W1. Protocols of the adapted baselines. Each policy defines a mixture over the same 32 regions, subject to the gated pool's sample capacities. All methods use the same final-training objective and budget (Tab. S8, App. G).

AutoAugment-style search. A policy-gradient controller proposes region mixtures. Each candidate is evaluated by a KK-step child run from θ0\theta_0, using mIoU on a fixed held-out twin set as the reward. The five highest-reward candidates passing the clean-loss screen are combined into the final policy, following AutoAugment's policy aggregation. The 66 total runs include one clean-control run.

Fast AutoAugment-style search. Following Fast AutoAugment's fixed-model search, Bayesian optimization minimizes the loss of θ0\theta_0 on held-out twins sampled from candidate mixtures. The top-ranked mixtures are combined without child-model training. Selection costs 0.03 final-run equivalents. At matched total compute, extending final training to 1.63T1.63T steps gives 47.98 versus 47.56 mIoU at TT, using the same final-training seed.

DoReMi-style reweighting. The 32 regions serve as domains. A reference model is trained from θ0\theta_0 on a uniform region mixture and then frozen. A proxy initialized from θ0\theta_0 uses Group DRO on excess loss relative to the reference. Trajectory-averaged domain weights define the final mixture. Reference and proxy updates, including reference forward passes, are included in the 0.66 selection cost.

RegMix-style regression. Each proxy trains for the probe horizon on a Dirichlet-sampled region mixture, covering sparse and near-uniform allocations. All runs are evaluated on the same fixed held-out twin set and clean held-out scenes. We fit ridge and LightGBM predictors for these losses and select the regression variant by out-of-fold rank correlation. One million candidate mixtures are scored, and the 100 with the lowest predicted twin loss among those passing the clean-loss screen and sample-capacity limits are averaged. The budgets of 66, 128, and 256 total runs each include one clean-control run, and 66 matches the number of CuRB probes.

Candidate clean-loss filtering. For AutoAugment-style search, candidates are screened using their measured clean held-out loss; for RegMix-style regression, using predicted clean loss. Both apply the source-calibrated tolerance ϵ\epsilon relative to the clean-control run. Screening precedes policy aggregation. CuRB instead applies its measured clean-harm constraint to individual regions.

RegMix-style diagnostics. Out-of-fold Spearman correlation on held-out mixtures increases from 0.38 to 0.52 and 0.61 at the three budgets, alongside increasing final mIoU. Additional regression and mixture-selection ablations are reported in Tab. S8.

AuthorsOpenReview ↗

Response to Reviewer dMGm [3/4]: W2, Q1

W2 and Q1. Repeated runs, the selected region, and the oracle. We repeat final training over three paired seeds, keeping θ0\theta_0 and each selected set fixed. Results are mean ± sample SD and include the submitted run (Tab. 7a).

MethodIn-family mIoUDark ZurichCityscapes val
CuRB51.90±0.2328.60±0.3570.09±0.14
DgD_g50.86±0.1527.77±0.1768.64±0.22
Fixed-Mid50.63±0.1427.53±0.2168.99±0.13
All-gen49.19±0.1826.62±0.1968.10±0.12

CuRB exceeds DgD_g, Fixed-Mid, and All-gen by 1.04, 1.27, and 2.71 in-family mIoU on average. Its paired gains over DgD_g are 0.95, 1.24, and 0.92. Across the four hosts in Tab. 2, each retrained with and without CuRB, the mean paired gains over three seeds range from 1.01 to 1.60 mIoU on the four-condition ACDC average and from 1.03 to 1.68 on Dark Zurich, with positive gains in every run. On SOD (NUN, CSOD10K), three seeds give Fβmax⁡F_\beta^{\max} of .878±.004 for CuRB against .866±.002 for uncurated twins (App. E).

Selected region and allocation. In the main setting, CuRB selects fog s5s_5 for 94% of the synthetic budget and low light s4s_4 for the remaining 6%. Tab. S6 reports the regions and shares for every setting. The top region alone, with clean data filling the remaining slots, reaches 51.78±0.18, compared with 51.90±0.23 for the full selection and the single-run source-only reference of 46.46 (Tab. 3). This region therefore accounts for most of the observed gain. The full selection averages 0.12 mIoU higher, while also filling the remaining slots with synthetic rather than clean data. Even splits over the top two and top four regions give 51.76±0.15 and 51.31±0.16 (Tab. 7c), which are 0.14 and 0.59 below the full selection. Neither tested equal split improves on rank-order filling.

Oracle comparison. We retrain each of the 32 regions once under the same final-training protocol, using candidates from that region and clean data to fill the remaining budget. In these single runs, fog s5s_5 achieves 51.84 mIoU and ranks 2/32 overall and 2/25 among feasible regions, with an observed regret of 0.16 mIoU relative to the best feasible single region, fog s4s_4 (52.00). For NUN on CSOD10K, the selected region, low light s4s_4, fills the whole budget and ranks 3/32 overall in the same single-run retraining, with an observed regret of 0.003 in Fβmax⁡F_\beta^{\max} relative to the best feasible single region. Both oracle comparisons use target-domain evaluation retrospectively and do not affect selection.

Selection stability. In the main setting, across three probe seeds and three cross-fitting partitions, the mean pairwise Spearman correlations between region rankings are 0.95 and 0.94, respectively. The top selected region remains unchanged, and each selected set shares at least 94% of its sample IDs with the default selection. With the final-training seed fixed, the changed selections achieve 51.81 to 52.04 in-family mIoU, compared with 51.97 for the default selection (App. C).

Reviewer_Z4BKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes CURB, a framework for curating synthetic degradations under a fixed training budget for dense prediction. Its central metric, controlled Reducible Degradation Gap (cRDG), compares short degradation-training probes with budget-matched clean-control probes on held-out degraded twins, while imposing a separate clean-harm constraint. The resulting scores rank degradation–severity regions to construct a fixed synthetic training set. Experiments on semantic segmentation and salient object detection demonstrate improvements over the evaluated selection baselines and benefits across several existing data-generation pipelines.

Critical strengths
  1. Clear motivation for controlled utility estimation. The paper distinguishes visual quality, instantaneous difficulty, and training utility. Equation (1), Proposition 1, and Figure 1 explain why paired gap reduction can reward clean forgetting and why degraded-loss reduction alone may still credit progress obtainable from clean training. This problem decomposition is a clear strength.

  2. Practical integration with existing pipelines. CURB operates as a curation layer without changing the predictor architecture and is evaluated with ISSA, H-Weather, Gen4Seg, and RobustNet+ISSA. Sharing clean-control probes across regions also limits the incremental cost of introducing the control branch.

  3. Useful ablations and utility analysis. Table 4 examines cross-fitting, held-out evaluation, control subtraction, and clean-harm gating. Table 6 reports global and within-family correlations, together with results on Dark Zurich, and appropriately acknowledges that between-family differences can inflate global correlation.

  4. Informative analysis of the correctable band. Figure 4 investigates how useful severity ranges depend on the probe budget, starting checkpoint, and predictor. Per-bin retraining at three final-training budgets provides useful evidence connecting band migration to downstream utility.

Critical weaknesses

W1. Limited methodological novelty relative to closely related work

The counterfactual comparison underlying cRDG has a close precedent in BSGAL [1], whose Definition 4.4 and Eq. (5) estimate generated-data contribution by comparing evaluation losses after updates on real and generated-data-augmented batches. DisCL [2] also uses model feedback to select synthetic guidance levels, although its data spectrum and feedback mechanism differ from CURB’s. CURB’s combination of degradation twins, region-level aggregation, matched probes, cross-fitting, and clean-harm gating is sensible, but the methodological advance beyond adapting these existing principles to synthetic degradations needs clearer articulation. Proposition 1 usefully explains a scoring failure mode, but it is an algebraic diagnostic rather than a theoretical guarantee for the estimator or selection policy. Overall, the contribution reads more as a carefully designed curation protocol than as a substantially new selection principle.

The scope of “forward transfer” also warrants clearer qualification: source images held out from the current probe may have contributed to training θ0\theta_0, so they are not necessarily scenes unseen throughout training.

W2. The region-filling policy is insufficiently validated

Section 3.4 exhausts regions in descending cRDG order, and Appendix A reports that the top-ranked region contributes over 90% of the selected data on Cityscapes and the entire synthetic budget on the SOD benchmarks. This raises an unresolved question about diminishing returns within a region: a high probe score does not establish that a region’s marginal benefit remains higher than that of other regions as more samples are selected. Comparisons with allocation across several high-scoring regions, together with performance curves as a function of samples selected per region, would help justify the current policy. On Cityscapes, a top-region-only baseline would directly test whether selecting beyond the dominant region adds value.

The checkpoint-dependent band shifts in Figure 4 also motivate exploring a training-time dynamic curriculum that periodically recomputes cRDG and updates severity selection or sampling proportions. Studying ranking stability and comparing staged re-scoring with static curation under matched total compute could clarify whether following these shifts improves performance. I view this as a promising extension, rather than a flaw in the current static formulation.

W3. The experimental comparison does not preserve key mechanisms of the closest related methods

Appendix G converts DisCL [2] into a one-time region ranking, removing the stage-dependent guidance selection of its adaptive variant. The gradient-contribution baseline associated with BSGAL [1] is described as candidate-level gradient alignment averaged within regions, but it is unclear whether it retains BSGAL’s real-batch control subtraction or online acceptance/rejection mechanism. Table 3 establishes an advantage within the evaluated static region-ranking protocol, but does not establish superiority over the original training-time selection strategies. Comparisons preserving the relevant mechanisms under matched total compute would strengthen the paper’s positioning.

The paper also lacks a model-specific breakdown of the selected degradation types, severity bins, and proportions in S⋆S^\star, making it difficult to connect the benchmark gains to the actual curated data. Reporting these compositions and testing targeted removal or replacement of the dominant region would strengthen gain attribution. A few repeated runs for key comparisons, particularly the 0.95 mIoU gain over DgD_g in Table 4, would also improve confidence in the reported margins.

W4. Clarity and organization require substantial improvement

Clarity and organization are currently the main barrier to understanding. The abstract and Introduction introduce many technical concepts—regions, matched probes, augmentation slots, clean-harm constraints, and correctable bands—before establishing a clear picture of the procedure. Although regions are described by degradation type and severity, their meaning as candidate groups rather than spatial image regions is only made explicit in Section 3.1, creating avoidable ambiguity in a dense-prediction paper.

The narrative also mixes motivation, operational details, and diagnostic findings without a clear hierarchy. In particular, the prominence of “correctable bands” and the name CURB suggest that bands drive selection, whereas the algorithm directly ranks regions and uses bands only for diagnosis. The method section then repeats the gap-reduction motivation while requiring readers to piece together probing, scoring, selection, and final retraining. These issues obscure a relatively straightforward central idea. Substantial restructuring is needed so that readers understand the overall procedure before encountering its notation, implementation details, and diagnostic analysis.

References

[1] Muzhi Zhu, Chengxiang Fan, Hao Chen, Yang Liu, Weian Mao, Xiaogang Xu, and Chunhua Shen. Generative Active Learning for Long-tailed Instance Segmentation. International Conference on Machine Learning (ICML), 2024.

[2] Yijun Liang, Shweta Bhardwaj, and Tianyi Zhou. Diffusion Curriculum: Synthetic-to-Real Data Curriculum via Image-Guided Diffusion. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 1697–1707, 2025.

Questions for authors
  1. Could the authors report the selected degradation types, severity bins, and proportions in (S*) for each evaluated model? How much of the final improvement can be attributed to the dominant selected region?

  2. How does the region-exhaustion policy compare with distributing the budget across several high-scoring regions, including a top-region-only baseline where meaningful? Do performance curves show diminishing returns as more samples are selected from the same region?

  3. Would CURB retain its advantage if the comparison preserved DisCL’s adaptive curriculum and BSGAL’s controlled contribution estimation and online acceptance mechanism under matched total compute?

  4. Figure 4 shows that the band centroid varies across training checkpoints and budgets. Could CURB be extended into a training-time dynamic curriculum, periodically recomputing cRDG and updating the selected severity regions or their sampling proportions? Under matched total compute, would this improve performance over one-time static curation? I view this as an exploratory extension motivated by the findings, rather than a prerequisite for the current contribution.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Z4BK [1/3]: W1, W2, Q1, Q2

W1. Contribution relative to BSGAL and DisCL. BSGAL is a direct precedent for controlled contribution estimation. Its Definition 4.4 and Eq. (5) compare evaluation losses after updates on a real batch and on the same batch with generated instances. CuRB instead estimates the additional held-out gain from content-preserving synthetic degradation relative to equally long clean training, using matched KK-step probes and a separate clean-harm constraint.

AspectBSGALcRDG in CuRB
InterventionGenerated instances pasted into a real batchA clean-augmentation slot replaced by a degraded twin of the same scene and label
Horizon and unitOne-step contribution per batch, estimated with a gradient cacheKK steps from a common checkpoint, cross-fitted and held out, per region
Separate clean-harm constraintNoYes
DecisionOnline accept or rejectOne-time fixed-budget selection

DisCL's adaptive curriculum selects guidance levels at each stage using validation progress after real-data training, whereas CuRB selects a fixed dataset for a given checkpoint and budget.

Two ablations test the clean-control correction and held-out evaluation. With the same probes and clean-harm constraint, omitting the correction (DgD_g) lowers correlation with full-retraining utility from 0.79 to 0.70 (Tab. 6) and in-family mIoU by 1.04 on average over three paired final-training seeds (0.95, 1.24, 0.92). Scoring on the probe-update split instead of held-out scenes lowers correlation from 0.79 to 0.44 (Tab. 4) and in-family mIoU by 2.58 on average over the same seeds (Tab. 7a). Online BSGAL and DisCL adaptations are compared in W3 (part 2/3).

Proposition 1 is now described as an algebraic diagnostic. The introduction clarifies that held-out scenes are excluded from probe updates but may have contributed to θ0\theta_0. As a retrospective check, we rescore the same probe checkpoints on twins of the 500 Cityscapes val scenes, which were not used to train θ0\theta_0 or the probes. The resulting ranking has a Spearman correlation of 0.92 with the default ranking across all 32 regions, and the top selected region is unchanged (App. C). These diagnostic scores were not used for data selection or hyperparameter tuning.

W2, Q1, and Q2. Region selection and budget allocation. In the main setting, CuRB selects fog s5s_5 for 94% of the synthetic budget and low light s4s_4 for the remaining 6%. After validity filtering, fog s5s_5 supplies Bdiag=5,593B_{\mathrm{diag}}=5{,}593 of the Bsel=5,950B_{\mathrm{sel}}=5{,}950 required samples. On SOD, the top region fills the entire budget for every predictor (Tab. S6). Under single-region retraining with clean filling, fog s5s_5 ranks 2/32 overall and 2/25 among feasible regions, with an observed regret of 0.16 mIoU relative to the best feasible single region.

We test allocation at two budgets (Tab. 7c). Ranks follow cRDG among feasible regions. Top-2, Top-4, and Ranks 2 and 3 use equal regional shares. Common-budget runs use BdiagB_{\mathrm{diag}} distinct synthetic samples, with repeated samples filling 5,950 synthetic slots, and report single-run results. At the original budget, Top-2, Top-4, and CuRB use 5,950 distinct synthetic samples. Top-1 uses 5,593 and fills the remaining 357 slots with clean data. Original-budget results are mean ± sample SD over three paired final-training seeds.

BudgetAllocationIn-family mIoUCityscapes val
CommonTop-151.8670.10
CommonTop-251.7970.02
CommonTop-451.3769.91
CommonRanks 2 and 351.2169.88
OriginalTop-1, remaining slots clean51.78±0.1870.08±0.10
OriginalTop-251.76±0.1570.05±0.09
OriginalTop-451.31±0.1669.94±0.13
OriginalCuRB (rank exhaustion)51.90±0.2370.09±0.14

At the common budget, Top-1 and Top-2 differ by 0.07 mIoU, while Top-4 and Ranks 2 and 3 trail Top-1 by 0.49 and 0.65. Ranks 2 and 3 is the allocation with the dominant region removed. Under the same protocol, nested subsets containing 25%, 50%, 75%, and 100% of the top region's distinct samples give 50.58, 51.27, 51.72, and 51.86. The observed increments decrease across these levels: 0.69, 0.45, and 0.14 mIoU points.

At the original budget, CuRB exceeds Top-1, Top-2, and Top-4 by 0.12, 0.14, and 0.59 mIoU on average. The Top-1 comparison also changes synthetic exposure because its remaining slots contain clean data. For Q1, Top-1 reaches 51.78±0.18, compared with the source-only reference of 46.46 (Tab. 3) and 51.90±0.23 for CuRB. Among the tested allocations, equal splits yield lower observed in-family mIoU than Top-1 at the common budget and CuRB at the original budget.

AuthorsOpenReview ↗

Response to Reviewer Z4BK [2/3]: W3, Q3, W4

W3 and Q3. Online selection baselines. We adapt DisCL's progress-based curriculum and BSGAL's cached-gradient acceptance using their released implementations. All methods use the same 32-region gated pool, clean source images, predictor, mixed-data objective, and base optimization settings. The adaptations retain their online selection rules rather than converting them to static region rankings.

DisCL-style selection. Each stage measures ground-truth class confidence on fixed feedback twins before and after clean-source updates, then selects the region with the largest progress for the next mixed-data stage. Clean-source updates count toward the training budget. Feedback twins are excluded from gradient updates, while their source scenes remain in training.

BSGAL-style acceptance. Candidate batches are proposed uniformly across regions. A momentum-cached reference gradient and the gradient difference between twin-filled and clean-filled batches determine the contribution score. The dynamic threshold controls acceptance, and rejected synthetic slots revert to clean data. We test clean-only and clean-and-twin reference batches.

We report comparisons at TT final-training steps and at the same selection-plus-training compute as CuRB. Costs are normalized by one final-training run. CuRB costs 1.66, comprising 1.00 for final training, 0.59 for probes, and 0.07 for calibration, scoring, and selection. Each adaptation includes its own overhead, including the feedback twins of DisCL-style, while the shared candidate pool is excluded. Synthetic exposure is the cumulative number of synthetic sample presentations during updates to the final predictor (probe updates excluded), normalized by default CuRB at TT steps. Random acceptance keeps the BSGAL-style proposals and acceptance rate. CuRB at BSGAL-style exposure keeps the default selected set and reverts slots to clean views. Values are mean ± sample SD over three paired seeds (Tab. S5).

MethodStepsCostSynthetic exposureIn-family mIoU
DisCL-styleTT1.180.9050.71±0.24
DisCL-style1.44$T$1.661.3050.93±0.29
BSGAL-style, clean referenceTT2.380.5250.62±0.27
BSGAL-style, clean reference0.70$T$1.660.3450.18±0.29
BSGAL-style, clean and twin reference0.70$T$1.660.3650.41±0.23
Random acceptanceTT1.000.5249.08±0.26
CuRB at BSGAL-style exposureTT1.660.5251.56±0.20
CuRBTT1.661.0051.90±0.23

At TT steps, adaptive DisCL-style selection reaches 50.71 versus 50.04 for the static DisCL-style ranking in Tab. 3. At matched compute, CuRB exceeds the strongest online adaptation, DisCL-style, by 0.97 mIoU on average (0.75, 1.44, and 0.71 across paired seeds), with DisCL-style using 1.30 times the cumulative synthetic exposure. CuRB exceeds every tested online adaptation in all three paired seeds.

At matched steps and synthetic exposure, CuRB exceeds BSGAL-style acceptance by 0.94 mIoU at lower total compute (1.66 versus 2.38). BSGAL-style exceeds random acceptance by 1.54 mIoU, supporting its contribution-based selection signal. With 250 source scenes and their twins withheld from subsequent parameter updates for CuRB and DisCL-style, and CuRB re-selected, matched-compute results are 51.74±0.25 versus 50.78±0.26 (App. G).

Selection composition and the dominant-region ablation are reported in W2, while repeated runs for the DgD_g comparison are reported in W1 (part 1/3).

W4. Clarity and organization. We have reorganized the method section to introduce the procedure before the formal definitions. The abstract and introduction define a region at first use as a group of candidates sharing a degradation type and severity bin. A four-step overview summarizes the pipeline: generate degraded twins of clean source images, run matched probes, select a fixed-budget synthetic set by ranking clean-harm-feasible regions with cRDG, and train the final predictor on clean data and selected degraded twins. The gap-reduction motivation is now stated more compactly. The score band is presented as a diagnostic summary and is not used for sample selection. Sec. 4.5 and Tab. 7 report seed variability, selected regions, and budget-allocation results.

AuthorsOpenReview ↗

Response to Reviewer Z4BK [3/3]: Q4

Q4. Training-time dynamic curation. We evaluate a two-stage extension. At T/2T/2, Dynamic CuRB re-estimates cRDG from the current checkpoint with a probe horizon scaled to the remaining training budget, then updates the selected regions for the second half. To reduce refresh cost, we restrict re-estimation to 12 regions fixed from the initial ranking: the highest-ranked severity bin of each degradation family and its two neighbors. This local search is motivated by the small centroid shifts in Fig. 3c of the revised PDF.

Staged static CuRB follows the same two-stage schedule but retains the initial ranking. Both use TT final-training steps, matched cumulative synthetic exposure, and 5,950 distinct synthetic samples across the two stages. We also evaluate longer static training at matched total compute, rerunning one-time curation for 1.09T1.09T steps with probe horizon 1.09T/801.09T/80. Cost includes selection and final training, as in W3 (part 2/3). Results are mean ± sample SD over three paired seeds (Tab. S7).

MethodStepsCost (× final run)In-family mIoUCityscapes val
Static CuRBTT1.6651.90±0.2370.09±0.14
Staged static CuRBTT1.6651.86±0.1770.06±0.12
Dynamic CuRBTT1.8052.03±0.2269.97±0.14
Static CuRB, matched compute1.09T1.09T1.8051.99±0.2070.13±0.11

The refreshed selection changes the top region from fog s5s_5 to fog s6s_6 in all three seeds. As a check, full 32-region re-estimation in each seed gives a mean Spearman correlation of 0.85 with the initial ranking and produces the same selected samples as the 12-region search, but at a refresh cost of 0.35 rather than 0.14 final-training runs.

At equal final-training steps, Dynamic CuRB exceeds staged static selection by 0.17 mIoU on average, with positive paired differences of 0.26, 0.05, and 0.20. At matched total compute, Dynamic CuRB reaches 52.03 versus 51.99 in-family mIoU for longer static training, while Cityscapes mIoU is 69.97 versus 70.13. With a full 32-region refresh, the matched-compute static run at 1.22T1.22T reaches 52.12 (App. H). Thus, dynamic updating provides a modest observed gain at equal training steps but no consistent advantage over longer static training at matched compute. We report it as an exploratory extension in Sec. 4.5 and retain one-time curation as the default.

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

The paper addresses the selection of synthetic degradation data in dense prediction. It proposes to estimate the true training utility of degradation regions under a finite training budget using matched clean-control probes, and builds a plug-and-play CuRB curation framework.

Critical strengths
  • The problem is clearly defined and well motivated, shifting synthetic degradation selection from realism/difficulty to training utility under a finite budget.
  • The matched clean-control design of cRDG is reasonable and can simultaneously expose that paired gap reduction is confounded by clean forgetting and that using only the degraded loss drop is not self sufficient and may overestimate the utility.
  • CuRB is plug-and-play, does not change the predictor architecture, and is validated on SOD, Cityscapes→ACDC/Dark Zurich, and multiple generation hosts.
  • Using clean and degraded views of the same source scenes also helps interpret the score.
  • The authors comparec RDG against actual region-level retraining utility, reporting good correlations. This is meaningful validation of the proxy.
  • The principal selection baselines are evaluated under matched budgets, and the clean-harm constraint is treated separately rather than hidden in a combined score.
Critical weaknesses
  • The fixed K=1/80 lacks sufficient validation across models and tasks. While it may be defensible as an efficient empirical proxy, its reliability must be validated for each new setting rather than assumed. In particular, with nonlinear training dynamics, scaling the probe length proportionally may not preserve ranking when learning-rate schedules, optimization phases, or data interactions change.
  • The clean-harm threshold depends on task or predictor settings, and its impact on the feasible region set needs further discussion. The theoretical analysis in the main text needs to be strengthened.
  • Most newly trained models are single runs, and seed variance is not reported.
  • Some plug-in comparisons do not provide a clean causal estimate : Several original host-pipeline numbers are taken from published reports, while the +CURB variants are run by the authors. Differences of roughly 1.2–1.5 mIoU may contain an implementation or recipe effect.. The plug-in claims would be more convincing if each host was rerun without and with CURB in the same environment.
  • The relationship to proxy-based data selection needs to be developed. DoReMi, RegMix, Skill-It, and related approaches also use inexpensive training signals to value data groups. They do not address the same content-matched degradation problem, and the clean-control design is a useful specialization. Still, the relevant conceptual difference should be explained more clearly.
  • The results may depend on the definition of the control and the probe. The same clean-control trajectory is used for all regions in a split direction. This is computationally efficient, but it leaves open whether the region ordering changes with the strength or type of the clean augmentation.
  • The probes also use a reduced learning rate and omit the DINOv2-weighted consistency component present in final training. The observed correlation with retraining utility is encouraging, but it would be helpful to test how sensitive the ranking is to these choices.
  • A potential concern is that cRDG evaluates each degradation on held-out twins from the same synthetic family and severity used in the treatment probe. Although the held-out split prevents memorization of the specific source scenes, it still measures in-family transfer under the generator’s own degradation distribution. Thus, a positive cRDG score shows that training on, for example, synthetic fog at a given severity improves performance on new source scenes rendered with that same synthetic fog relative to matched clean augmentation; it does not by itself establish that the degradation matches the unknown deployment distribution or will generalize equally well to real-world adverse conditions. The paper addresses this only through separate external evaluations on real target domains, so the reliability of cRDG as a selection criterion depends on how well its synthetic within-family ranking predicts those downstream gains.
Questions for authors
  • cRDG uses a fixed probe horizon K=1/80 of the final schedule. How sensitive are the rankings and final gains to this ratio?
  • cRDG uses a weak photometric augmentation as the clean control. What evidence shows that the control probe is a fair counterfactual for “equally long clean training” across different predictors and training budgets? Along the same lines, sensitive is the cRDG ranking to the control augmentation and to using the full final-training objective during probing?
  • Which region is selected in each principal experiment, and what is its rank and performance relative to the oracle-best region from full retraining?
  • What is the variation across training and probe seeds?
  • Can the host-pipeline comparisons be repeated with matched implementations?
3 comments
AuthorsOpenReview ↗

Response to Reviewer visA [1/3]: W1, W2, W3, Q1, Q4

W1 and Q1. Sensitivity to the probe horizon. Proportional scaling is tested in Fig. 3b of the revised PDF. From 0.5x to 2x the final budget, per-bin retraining shifts the empirical centroid by +0.20 for fog and +0.19 for low light, and the scaled-KK cRDG prediction stays within 0.03 of it at each tested budget. To complement this, we fix the final training schedule TT and vary the probe horizon KK, keeping other probe settings and the default clean-harm tolerance ϵ\epsilon fixed. Both treatment and control run for KK steps. We evaluate each ranking against the same full-retraining utility UrealU_{\mathrm{real}} and compare the selected sets (Tab. S3). Overlap is ∣S∩Sdef∣/∣Sdef∣|S\cap S_{\mathrm{def}}|/|S_{\mathrm{def}}| over unique synthetic sample IDs.

KKρglob\rho_{\mathrm{glob}} with UrealU_{\mathrm{real}}Top selected regionOverlapIn-family mIoU
T/160T/1600.75Unchanged0.9451.88
T/80T/80 (default)0.79Unchanged1.0051.97
T/40T/400.80Unchanged1.0051.97†

† Identical selected set and final-training protocol, so the default checkpoint is reused.

The top selected region remains unchanged across all three horizons. On SOD (NUN, CSOD10K), the three horizons also select identical sets. These results support K=T/80K=T/80 as a stable operating choice in the evaluated settings.

W2. Clean-harm threshold and theoretical analysis. We use the same source-only calibration rule across tasks and predictors, with its numerical value adapting to the loss scale and image-level variability. For the per-image clean-loss difference di=ℓi(θa)−ℓi(θb)d_i=\ell_i(\theta_a)-\ell_i(\theta_b) between two degradation-free matched probes differing in their slot-augmentation draws, we set ϵ=s(di)/n\epsilon=s(d_i)/\sqrt{n}, one standard error (SE), where ss is the sample standard deviation over nn held-out source images. The resulting values are 0.011 for ISSA SegFormer (25/32 feasible regions), 0.014 for RobustNet DeepLabV3+, and 0.006 for NUN on CSOD10K, without target-domain tuning.

Tab. S2 reports the sensitivity to ϵ\epsilon:

ToleranceFeasible fractionIn-family mIoUCityscapes mIoU
0 SE0.4151.4370.21
0.5 SE0.6351.8670.24
1 SE (default)0.7851.9770.13
2 SE0.9151.7969.82

From 0.5 to 2 SE, more regions become feasible while in-family mIoU varies by at most 0.18. Cityscapes mIoU decreases from 70.24 to 69.82, showing the clean-retention trade-off.

Theoretical analysis. We clarify the two confounding effects through the exact decompositions RDGg=Dg+Fg\mathrm{RDG}_g=D_g+F_g (Eq. 1) and Dg=Gg+CgD_g=G_g+C_g (Sec. 3.3), where Fg=Lc(θgdeg)−Lc(θ0)F_g=L_c(\theta_g^{\mathrm{deg}})-L_c(\theta_0) is the clean-loss change and Cg=Ld(θ0;g)−Ld(θctrl;g)C_g=L_d(\theta_0;g)-L_d(\theta^{\mathrm{ctrl}};g) is the degraded-loss reduction from clean-control training. These identities hold for each held-out split direction and their average. Thus, clean forgetting can inflate RDGg\mathrm{RDG}_g, while DgD_g still includes clean-control progress. cRDG subtracts this progress through Gg=Dg−CgG_g=D_g-C_g. Separately, CuRB requires Lc(θgdeg)≤Lc(θctrl)+ϵL_c(\theta_g^{\mathrm{deg}})\leq L_c(\theta^{\mathrm{ctrl}})+\epsilon in each held-out direction, constraining probe-time clean harm relative to the matched control. Final clean performance is evaluated in Tabs. 4 and 5.

W3 and Q4. Training and probe variability. We repeat final training with three paired seeds, keeping θ0\theta_0 and each method's selected set fixed. Tab. 7a reports mean ± sample standard deviation (SD), including the submitted run.

MethodIn-family mIoUDark ZurichCityscapes val
CuRB51.90±0.2328.60±0.3570.09±0.14
DgD_g (same HH constraint)50.86±0.1527.77±0.1768.64±0.22
Fixed-Mid50.63±0.1427.53±0.2168.99±0.13
All-gen49.19±0.1826.62±0.1968.10±0.12

CuRB exceeds DgD_g by 0.95, 1.24, and 0.92 mIoU across the three paired seeds, respectively.

For probe variability, we vary the probe seed over three runs at a fixed A/B partition and the partition over three runs at a fixed probe seed. The mean pairwise Spearman correlations across the 32 region rankings are 0.95 and 0.94, respectively, with the same top selected region in every run. Every resulting selected set retains at least 94% of the default set's synthetic sample IDs. Changed sets are trained with the same final-training seed and recipe as the submitted run, yielding 51.81–52.04 in-family mIoU versus 51.97 for the default set. Identical sets reuse the default checkpoint (App. C).

On SOD (NUN, CSOD10K), three final-training seeds give Fβmax⁡F_\beta^{\max} of .878±.004 for CuRB versus .866±.002 for uncurated twins (App. E).

AuthorsOpenReview ↗

Response to Reviewer visA [2/3]: W4, W5, W6, Q2, Q5

W4 and Q5. Host comparisons with matched implementations. We train each host in Tab. 2 with and without CuRB over three paired seeds, matching the implementation, initialization, source data, synthetic-data budget, training recipe, and evaluation protocol. The paired final-training runs differ in synthetic-data selection alone. Values are ACDC / Dark Zurich mIoU, reported as mean ± sample SD over three runs, with ACDC averaged over its four conditions.

Host (predictor)Reproduced+CuRBPaired gain
ISSA (SegFormer)52.18±0.23 / 26.97±0.2453.74±0.23 / 28.61±0.37+1.56 / +1.64
H-Weather (SegFormer)49.40±0.16 / 23.30±0.2750.74±0.24 / 24.75±0.34+1.34 / +1.45
Gen4Seg (SegFormer)50.60±0.21 / 25.38±0.3052.20±0.21 / 27.06±0.31+1.60 / +1.68
RobustNet+ISSA (DeepLabV3+)47.81±0.16 / 23.35±0.2148.82±0.23 / 24.38±0.38+1.01 / +1.03

CuRB improves every host, with mean paired gains of 1.01 to 1.60 ACDC mIoU and 1.03 to 1.68 Dark Zurich mIoU. Gains are positive in all three paired runs for every host on both datasets. Tab. 2 now reports these matched comparisons instead of using published host results as baselines.

W5. Relation to DoReMi, RegMix, and Skill-It. These methods also use training signals to guide data allocation. DoReMi reweights domains using proxy-model excess loss relative to a reference model, RegMix predicts mixture performance through regression on small-model runs, and Skill-It updates skill sampling using learned skill relationships and evaluation losses. CuRB instead estimates the additional training gain from synthetic degradation relative to equally long clean-augmentation training on the same scenes and labels. cRDG compares two matched KK-step probes from the same checkpoint that differ only in the augmentation slot: degraded twins versus clean augmentations of the same scenes and labels. Their loss difference is evaluated on held-out degraded twins, while clean harm is constrained separately during selection. The DgD_g ablation omits the matched clean-control progress correction while retaining the same probes and HH constraint. Its rank correlation with retraining utility is 0.70 versus 0.79 for cRDG (Tab. 6), and it trails CuRB in every paired seed (W3 in part 1/3). Adaptations of DoReMi and RegMix to our pool at matched selection cost reach 50.47±0.23 and 50.72±0.19 in-family mIoU against 51.90±0.23 for CuRB (W1 of Reviewer dMGm, Tab. S8). We clarify these distinctions in Related Work.

W6 and Q2. Choice of the clean control. Although the clean-control trajectory is shared across regions, its loss is evaluated on the degraded twins of each region, so changing the control can change their ranking. We compare three geometry-preserving controls on ISSA SegFormer, recomputing region scores and clean harm with the default ϵ\epsilon fixed (Tab. S3).

Control in the slotρglob\rho_{\mathrm{glob}} with UrealU_{\mathrm{real}}Top selected regionOverlap with defaultIn-family mIoU (single run)
No photometric augmentation0.73Unchanged0.9451.70
Weak photometric (default)0.79Unchanged1.0051.97
Strong photometric0.76Unchanged1.0051.97†

Overlap is the fraction of default selected sample IDs retained. Changed sets use the same final-training seed and recipe as the default. † The selected set is identical to the default, so the default checkpoint is reused.

The top selected region remains unchanged, and each selected set retains at least 94% of the default set. For the clean-training counterfactual, treatment and control share the starting checkpoint, source scenes, labels, clean replay, and number of updates, differing in the augmentation slot. The same matching is used for each evaluated predictor and training budget. Under this comparison, the clean-control correction raises the correlation with full-retraining utility from 0.70 (DgD_g) to 0.79 (cRDG, Tab. 6).

AuthorsOpenReview ↗

Response to Reviewer visA [3/3]: W7, W8, Q2, Q3

W7 and Q2. Probe learning rate and objective. On ISSA SegFormer, we compare probe learning rates of 0.25x and 1.0x the final-training base rate, using either the supervised terms or the full objective with DINOv2-weighted consistency (Tab. S3). Within each setting, treatment and control use the same learning-rate schedule and loss coefficients. Under the full objective, the consistency term pairs the clean view with the degraded twin in the treatment and with the augmented view in the control. Other probe settings, the held-out supervised scoring loss, and the default ϵ\epsilon are fixed.

Probe settingρglob\rho_{\mathrm{glob}} with UrealU_{\mathrm{real}}Top selected regionOverlap with defaultIn-family mIoU
0.25x LR, supervised (default)0.79Unchanged1.0051.97
1.0x LR, supervised0.76Unchanged1.0051.97†
0.25x LR, full objective0.80Unchanged0.9452.00
1.0x LR, full objective0.78Unchanged1.0051.97†

Overlap is the fraction of default selected sample IDs retained. Changed sets use the default final-training seed and recipe. † Identical selected set and final-training protocol, so the default checkpoint is reused.

Across the four settings, ρglob\rho_{\mathrm{glob}} ranges from 0.76 to 0.80, the top selected region is unchanged, and each set retains at least 94% of the default. The full objective requires 1.3x the default probe-update compute without changing the top selected region, supporting the supervised probe as the default.

W8. Scope of forward transfer. cRDG measures the gain over matched clean-control training on held-out twins from the same generator, degradation family, and severity bin. Its value for real-domain selection is tested through full retraining of all 32 regions. On ACDC, the rank correlation with in-family retraining utility is 0.79 globally and 0.74 averaged over within-family correlations, compared with 0.70 and 0.56 for DgD_g. On Dark Zurich, the corresponding correlation is 0.72 versus 0.62 (Tab. 6). At the final-model level, CuRB improves all four hosts on ACDC and Dark Zurich in paired runs (W4 in part 2/3). In the matched curation experiment, snow is excluded from the candidate pool and final training, yet CuRB reaches 51.38 mIoU on ACDC snow versus 50.79 for the best baseline in Tab. 3 and 50.15 for All-gen. The revised introduction clarifies the distinction between synthetic forward transfer and real-domain evaluation.

Q3. Selected regions and oracle comparison. In the main setting, CuRB selects fog s5s_5 for 94% of the synthetic budget and low light s4s_4 for the remaining 6%. We retrain each of the 32 regions once under the same final-training protocol, using candidates from that region and filling the remaining slots with clean data. Ranked by ACDC in-family utility, fog s5s_5 (51.84 mIoU) ranks 2/32 overall and 2/25 among feasible regions, with an observed regret of 0.16 mIoU points relative to the best feasible single-region run, fog s4s_4 (52.00).

We also retrain all 32 regions once for NUN on CSOD10K. The selected region, low light s4s_4, fills the entire synthetic budget and ranks 3/32 overall, with an observed regret of 0.003 in Fβmax⁡F_\beta^{\max} relative to the best feasible single-region run. In both settings, feasibility is determined by the source-side clean-harm constraint. Target-domain evaluation is used for retrospective oracle analysis, not selection. Tab. S6 lists the selected regions and their shares across all evaluated settings. In the main setting, the top selected region remains unchanged across probe seeds and partitions (W3 in part 1/3).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers (visA, Z4BK, and dMGm) for their thoughtful and constructive feedback, which has helped us strengthen our work. We are encouraged that the reviewers recognize the clear motivation of valuing synthetic degradations by training utility rather than realism or difficulty (visA, Z4BK, dMGm), the matched clean-control formulation that separates clean forgetting and generic training progress from degradation-specific gain (visA, Z4BK, dMGm), the validation of cRDG against region-level retraining utility with useful ablations (visA, Z4BK), the informative analysis of the correctable band (Z4BK), and the plug-and-play design with consistent gains across tasks, architectures, and data-generation pipelines (visA, Z4BK, dMGm).

In response to the reviewers' valuable comments, we have carefully addressed each concern in our individual responses and summarize the main revisions below:

  • We isolated the effect of the matched clean-control subtraction and clarified its relation to BSGAL, DisCL, and proxy-based selection. With the same 66 probes and clean-harm constraint, removing the subtraction (DgD_g) lowers the rank correlation with full-retraining utility from 0.79 to 0.70 on ACDC and from 0.72 to 0.62 on Dark Zurich. Over three paired final-training seeds, CuRB reaches 51.90±0.23 in-family mIoU versus 50.86±0.15 for DgD_g, a 1.04 gain that is positive in every seed (Tabs. 6 and 7a). These results support the benefit of the clean-control correction beyond the held-out loss drop.
  • We added online DisCL- and BSGAL-style adaptations that retain their training-time selection rules, together with AutoAugment-, Fast AutoAugment-, DoReMi-, and RegMix-style proxy search on the same candidate pool and final-training protocol. At matched selection-plus-training compute, the strongest online adaptation, DisCL-style, reaches 50.93±0.29 versus 51.90±0.23 for CuRB. At equal training steps and synthetic exposure, CuRB reaches 51.56±0.20 versus 50.62±0.27 for BSGAL-style at lower total compute. At CuRB's selection cost, AutoAugment-, DoReMi-, and RegMix-style search reach 50.47 to 50.72, and RegMix-style reaches 51.03±0.22 with 3.8 times that cost (Tabs. S5 and S8). These results support the advantage of CuRB under matched compute and synthetic exposure.
  • We reported repeated runs, matched host retraining, and the actual selection. All four hosts are retrained with and without CuRB over three paired seeds, with mean gains of 1.01 to 1.60 mIoU on ACDC and 1.03 to 1.68 on Dark Zurich that are positive in every run (Tab. 2). In the main setting, CuRB selects fog s5s_5 (94%) and low light s4s_4 (6%). Retrospective single-region retraining ranks fog s5s_5 second among 25 feasible regions with an observed regret of 0.16 mIoU, and the top region with clean filling is within 0.12 mIoU of the full selection (Tabs. 7 and S6). These results support the reliability of the selection and of the reported margins.
  • We tested probe sensitivity, larger candidate spaces, and dynamic curation. The top selected region is unchanged across probe seeds, cross-fitting partitions, probe horizons, clean controls, learning rates, and objectives (App. C, Tab. S3). With 64 regions, halving the probe horizon gives a selection cost of 0.69 final-run equivalents, close to 0.66 for 32 regions, while keeping 93% of the full-horizon selection, and a two-stage search that scores 40 regions recovers the exhaustive selection (Tab. S9). A mid-training cRDG refresh improves in-family mIoU by 0.17 over staged static selection at equal training steps but shows no consistent advantage at matched compute, so one-time curation remains the default (Tab. S7). These analyses clarify the stability and cost of the probe protocol.
  • We reorganized the presentation. A region is defined at first use as a degradation type and severity bin, the four-step procedure precedes the formal definitions, the score band is presented as a diagnostic rather than a selection rule, and Proposition 1 is described as an algebraic diagnostic. We also clarified the scope of forward transfer. Held-out scenes may have contributed to the initial checkpoint, and rescoring on twins of 500 Cityscapes val scenes unseen by θ0\theta_0 and the probes gives a rank correlation of 0.92 with the default ranking (App. C).

We thank the reviewers again for their valuable suggestions. We have revised the manuscript accordingly, updating the main text, experimental results, and appendix to incorporate these improvements, with the main changes marked in blue.

We appreciate the reviewers' time and consideration in reviewing our responses. We hope that the additional experiments and clarifications have addressed the concerns raised, and we welcome any further questions or discussion.

Best Regards,

All #4105 Authors

Area_Chair_5WvZOpenReview ↗

AC Reminder: Please Follow Up on the Authors’ Responses

Dear Reviewers,

Thank you for your time and thoughtful reviews. The authors have posted responses to your questions and concerns.

If you have not yet done so, please read their responses and any relevant revisions, and leave a follow-up comment visible to the authors indicating which concerns have been addressed and which remain unresolved. Please update your review and rating if warranted. If your assessment remains unchanged, a brief explanation would still be helpful.

Thank you for your continued engagement in the discussion.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.