acceptodds
Under review as a conference paper at ICLR 2027

Cross-modal drifting

Abstract

Learning a conditional distribution from paired modalities can be challenging when each source observation has only one observed target. For conditional drifting, using that single paired target as the empirical target specifies a point mass rather than the population-conditional. We introduce Cross-Modal Drifting (CMD), which constructs a conditional target proxy from the paired targets of nearby source observations, weighted by their source-space proximity. A kernel-gradient field matches the generator's feature distribution to this proxy, further smoothed using local target-space covariance. The resulting generator produces each sample in a single network evaluation, with no neighborhood retrieval required at inference. Across cross-modal prediction tasks, CMD achieves the lowest mean Energy Score among the evaluated methods, including multi-step flow-matching baselines whose strongest evaluated setting costs fifty times more at inference. A component ablation supports the benefits of the proposed conditional proxy and, in most settings, of local covariance smoothing.

Primary area:
generative models
Keywords:
conditional generative modeling, cross-modal generation, one-step generation, drifting models, kernel methods

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

The paper studies Cross-Modal Drifting (CMD) which is proposed for conditional generation from paired cross-modal data where each source observation usually has only one observed target. As a result, the empirical conditional is a point mass. The paper constructs a richer local proxy and uses fixed source and target representations to retrieve nearby source observations and then assigns Gaussian distance weights. Those weights are transferred to the paired targets. A conditional generator is then trained with a kernel-gradient drifting field against this proxy. At inference, CMD generates a sample with one network evaluation and needs neither neighborhood retrieval nor field computation. The paper proves identifiability relative to the supplied proxy. Experiments cover a synthetic conditional circle, a cat-to-dog sketch task, BMMC CITE-seq RNA-to-protein generation, and HEST-IDC histology-to-gene-expression generation. CMD achieves the lowest mean Energy Score on both biological benchmarks and remains competitive as synthetic ambient dimension grows. CMD has higher training cost and its distributional gains do not consistently translate into better Pearson correlation.

Critical strengths
  1. The paper identifies a real obstacle in conditional generative modeling. This is especially important in biology, where a CITE-seq cell yields one RNA and one protein profile, and a spatial transcriptomics spot pairs one histology patch with one expression readout.

  2. CMD’s proxy is conceptually clean. It avoids storing or modeling a full conditional density and avoids assuming a parametric form. The use of Kish effective sample size to make MM comparable across queries is a suitable design choice.

  3. CMD builds on Kernel-Gradient Drifting, where the drift field is the difference between the scores of kernel-smoothed target and generator distributions. This gives the method an attractive-repulsive field.

  4. The optional CDC smoothing replaces point masses at paired targets with local Gaussian components estimated from target-space neighborhoods. This is a sensible way to inject local target geometry without changing the source-derived mixture weights.

  5. CMD requires only one neural function evaluation at inference. Once trained, the generator maps noise and a source representation to a target sample without neighborhood retrieval, drift computation, or iterative sampling. This is a real advantage over flow and consistency method.

  6. The component study separates reference drifting, KGD, CDC, the source-neighborhood proxy, and their combinations. This makes it possible to see which component drives gains.

Critical weaknesses
  1. CMD assumes that nearby source representations have similar conditional target distributions. This is the central statistical assumption, but the paper does not provide a diagnostic for when it holds. The sketch task is designed so that nearby points on different strokes can have different conditionals, yet the paper does not isolate how much of CMD's error comes from neighborhood mixing versus generator fitting. On real data, source representations are fixed PCA or pretrained embeddings, and there is no check that their local neighborhoods preserve target-relevant structure.

  2. CMD uses fixed source and target representation maps such as PCA for BMMC. The generator never adapts these representations. This simplifies training and makes the proxy stable, but it also means the method inherits any information loss or bias in the representations. If the source representation discards a factor that predicts target variability, no amount of proxy weighting can recover it.

  3. CMD posts the largest training total on HEST-IDC and, after CDC-FM, on BMMC. The proxy positives make its update the most expensive in the drift family. The paper notes that KGD-proxy is cheaper when CDC's gain is small, but the main method still carries a substantial training burden. For biological applications with large datasets or many folds, this cost may be prohibitive. The paper does not establish a clear cost-quality or show that CMD's ES gains justify the added training time in downstream terms. A cheaper approximation may be preferable in many settings.

  4. The evaluation covers two synthetic tasks and two biological benchmarks. There is no evaluation on image-text, audio-text, multimodal medical imaging, or other paired modalities. The biological tasks are both expression-prediction problems, and HEST uses one training seed per fold. The paper does not test robustness to missing modalities, noisy pairs, batch effects, or dataset shift. It also does not compare with strong conditional generative baselines beyond the flow/consistency family, such as conditional normalizing flows or conditional GANs.

  5. CMD involves several choices and the paper fixes many of these from prior work or inner validation, but does not provide a systematic sensitivity analysis.

  6. The code does not seem available which makes judgement about reproducing the results challenging.

Questions for authors
  1. hen does the source-neighborhood assumption actually hold, and can it be diagnosed before training? CMD's proxy is only as good as the claim that nearby source representations have similar conditional target distributions.

  2. Do CMD's distributional gains translate into downstream biological utility? ES improves, but Pearson correlation does not consistently improve and sometimes favors KGD or MeanFlow. Does better conditional calibration lead to better imputation, differential expression, cell-state recovery, or spatial prediction? The paper should connect ES gains to concrete downstream tasks, not only to a proper scoring rule.

  3. hyperparameters from prior work or inner validation. How much do ES and Pearson change across reasonable ranges of all hyperparameters? Is the method stable across seeds and folds or do the gains depend on favorable calibration? A sensitivity study would clarify robustness.

6 comments

We thank the reviewer for their careful reading and constructive feedback. We appreciate the opportunity to clarify CMD’s assumptions, discuss its downstream utility, and address its sensitivity to methodological choices. We address the three questions separately.

When does the source-neighborhood assumption hold, and can it be diagnosed before training?

We agree that this is the central statistical assumption (Section 3.1). It cannot be verified directly from single-observation data, because testing whether two nearby sources share a conditional distribution would require repeated targets per source. We therefore provide a direct measurement on the sketch task, where the conditional is known, and predictive checks that are available on real data before any generator is trained. These are the checks the current procedure already provides; the conclusion names dedicated diagnostics of proxy quality as a direction for further work.

Sketch: measuring neighborhood mixing. Appendix B.8 notes that nearby sources on different strokes can have different conditionals, and that the reported results alone do not separate neighborhood mixing from generator fitting. To complement that discussion, we computed for every held-out source the fraction of its proxy weight assigned to training sources on a different stroke, at the selected breadth M=16M = 16 (D=2D = 2, three replicates, 768 held-out sources). The mean cross-stroke weight is 2.8% ± 0.2%, and about 97% of the 16 nearest training sources lie on the query's stroke. We then split the held-out sources into three groups by this fraction, using cut-offs of 10% and 30% chosen before inspecting the results, and report the mean conditional W1W_1 within each group. W1W_1 is computed per source between a method's generated samples and samples drawn from the true conditional:

Cross-stroke proxy weightSourcesOracle samplerDirect proxyKGD + proxy + CDC (E)FM, 50 NFE
At most 10%705 (91.8%)0.0240.0560.0550.071
10% to 30%33 (4.3%)0.0200.2440.4080.485
More than 30%30 (3.9%)0.0210.5120.5980.605

The first method column scores an independent sample from the true conditional in the same way as a method; it is the lowest W1W_1 attainable at this sample size. Every method remains more than twice this floor, including on the sources with nearly pure neighborhoods.

On the 92% of sources with nearly pure neighborhoods, the trained generator matches the proxy it was trained against (0.055 vs 0.056). Where neighborhoods mix strokes, the proxy's W1W_1 is roughly 4 and 9 times higher, and the generator's is higher still. FM, which uses no neighborhoods, degrades on the same sources, so they are hard for every method, and we do not attribute the generator's error there to mixing alone.

Real data: checks available before generator training. The proxy is itself a predictive distribution, so its inner-validation ES can be computed for every candidate breadth without training a generator. This is the criterion that selects MM (Section 3.2, Appendix A.1). The full curves are:

MM48163264128
BMMC6.8006.1565.8665.7555.7355.765
HEST-IDC (mean of four folds)3.3493.0552.9452.9302.9653.028

Both curves have an interior minimum. Very narrow neighborhoods are 13–19% worse than the selected breadth. The informative part is what happens beyond the selected breadth. If distance in the fixed representation were unrelated to the target, far neighbors would be as useful as near ones, and adding them would not worsen the score. Instead the score worsens as the neighborhood grows past the selected breadth, by 0.5% at M=128M = 128 on BMMC and, in every HEST-IDC fold, by 0.9–1.6% at M=64M = 64 and 2.7–4.4% at M=128M = 128. Nearer sources therefore predict the target better than farther ones.

As a second check, the selected proxy can be compared with trained baselines before committing to generator training (Section 4.5, Table 3). These are held-out test scores for the same proxy at the selected MM. The scores in the table above are inner-validation scores, computed on queries held out at random from the training data, so the two are not comparable in magnitude. Sampled directly, the proxy attains held-out ES 6.712 on BMMC, lower than KGD trained on single paired targets (6.855) and all six competing objectives (6.938 or higher). On HEST-IDC it attains 6.555, lower than KGD (6.620) and five of the six competing objectives; CDC-FM at 50 NFE scores 6.528.

These are diagnostics of predictive usefulness, not guarantees that every neighborhood preserves the conditional law. As the sketch task shows, a favorable aggregate can coexist with a minority of mixed neighborhoods. Regarding the fixed representations, we agree that the proxy cannot recover information that the representation discards. Two remarks put this in context. All methods in our comparison are conditioned on the same fixed representations (Appendix B.1), so none of them has access to the discarded information. And the checks above test whether neighbors in the given representation predict the target; they cannot tell whether a different representation would do better. Section 3.1 states that the quality of the proxy depends on how well the source representation preserves target-relevant structure, and the conclusion names this as a limitation, with learning the representations jointly with the generator as a direction for further work.

Reviewer_xHWXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Cross-Modal Drifting (CMD) which is a method for conditional generation from paired multimodal observations when each conditioning input is associated with only one observed target. The central idea is to replace the Dirac empirical conditional distribution with a neighborhood-based proxy where nearby observations in the source representation space contribute their paired targets, weighted by source similarity. The neighborhood bandwidth is calibrated through an effective sample size criterion, and the resulting target distribution can be further smoothed using local target-space covariances inspired by carré du champ constructions. A conditional generator is trained to match this proxy using the kernel-gradient drifting (KGD) field and a stopped-gradient regression objective, then allowing conditional generation without neighbourhood retrieval at inference time. Beyond synthetic data, the method is evaluated on RNA-to-protein generation on BMMC CITE-seq and histology-to-gene-expression prediction on HEST-IDC where it achieves the lowest reported mean Energy Score. Some component ablations support the importance of the neighborhood proxy, with more modest and setting-dependent gains from covariance smoothing. These improvements come at increased training cost and do not consistently extend to predictive-mean accuracy.

Critical strengths
  1. The paper addresses a concrete limitation of conditional drifting when each conditioning observation has only one paired target. Constructing a weighted empirical target distribution from nearby source observations is an intuitive and relatively easy to implement. The effective-neighborhood-size calibration provides a useful way to adapt the weighting scale across queries. Importantly, the resulting generator amortizes this construction and requires only one forward pass without retrieval at deployment.
  2. The evaluation includes both synthetic examples with known conditional laws and biological prediction tasks. The BMMC ablations provide consistent evidence that adding the neighborhood proxy improves upon KGD.
Critical weaknesses

My main concern is that the experiments do not adequately establish which part is responsible for its improvements. CMD combines three largely established ingredients: (i) a neighborhood-weighted empirical conditional distribution, closely related to classical kernel conditional distribution estimation; (ii) the existing kernel-gradient drifting (KGD) objective; and (iii) local covariance smoothing adapted from CDC-FM. Related conditional drifting methods, including Keyed Drifting Policies [2] and Implicit Drifting Policy [1], have also addressed sparse conditional observations by borrowing information from neighboring conditioning inputs. The precise methodological distinction from these approaches needs to be clarified more clearly. Moreover, existing approaches can learn conditional stochastic generators directly from individually paired observations. For example, Energy Score minimization (as in Engression [3] ) provides a strictly proper population objective whose stochastic gradients can be estimated using one observed target per condition and multiple generated samples, or also less recent works [4]. Additionally, the neighborhood distribution is already a valid conditional predictive distribution, whose breadth is explicitly selected using validation Energy Score before generator training and can thus be sampled directly without fitting a neural generator. It seems to work well in the high-dimensional circle experiments, so should be considered a more serious baseline throughout the experiments.

References: [1] Yang, Zemin, et al. "Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry." arXiv preprint arXiv:2606.01098 (2026). [2] Puthumanaillam, Gokul, and Melkior Ornik. "Amortizing Trajectory Diffusion with Keyed Drift Fields." arXiv preprint arXiv:2603.14056 (2026). [3] Shen, Xinwei, and Nicolai Meinshausen. "Engression: extrapolation through the lens of distributional regression." Journal of the Royal Statistical Society Series B: Statistical Methodology 87.3 (2025): 653-677. [4] Ren, Yong, et al. "Conditional generative moment-matching networks." Advances in Neural Information Processing Systems 29 (2016).

Questions for authors
  1. What is the incremental benefit of drifting-based amortization over the conditional proxy and alternative training objectives? The ablation shows that the proposed proxy improves KGD, but it does not establish whether KGD is the most effective way to learn from that proxy. Could the authors report the held-out Energy Score and sampling cost of the discrete and CDC-smoothed proxies directly on BMMC and HEST.
  2. How does the proposed neighborhood construction and subsequent drifting objective differ from recent conditioning-aware drifting approaches such as Keyed Drifting Policies and Implicit Drifting Policy?
  3. What aspects of predictive distributions account for the biological Energy Score gains? Could the authors provide complementary diagnostics, such as predictive-mean RMSE, marginal CRPS, interval coverage, particularly for HEST?
6 comments

We thank the reviewer for the thorough and insightful review. The question of attribution led directly to the direct-proxy baselines and the controlled comparison of the two drifting fields, now in Section 4.5, which we believe make the paper considerably stronger. The suggested diagnostics and the pointers to related conditioning-aware drifting methods and to distributional-regression objectives were equally valuable. We address each question below.

Incremental benefit over direct proxy sampling and alternative objectives

We thank the reviewer for this suggestion, which led to one of the most informative additions to the paper. We agree that the neighborhood proxy is already a valid conditional predictor that can be sampled without training a generator, and we have added it as a baseline on both biological benchmarks (Section 4.5, Table 3). It performs better than we had anticipated: sampled directly, it attains a lower mean ES than KGD trained on single paired targets and than all six competing objectives on BMMC, and than all of these except CDC-FM at 50 NFE on HEST-IDC. The fitted generators nevertheless improve on it. On both datasets, every generator trained against a proxy attains a lower mean ES than direct sampling from that same proxy, by 0.4–3.8%. The details follow: the matched comparison, the choice of drifting field, sampling cost, and alternative objectives such as Engression.

All predictors use 64 samples per held-out query, with the same representations, reference bank, selected breadth, and reporting maps. Where CDC is used, its covariance construction is also shared.

PredictorBMMC ES ↓HEST-IDC ES ↓
Direct proxy6.712 ± 0.0016.555 ± 1.268
Reference Drifting + proxy6.661 ± 0.0046.401 ± 1.258
KGD + proxy (D)6.611 ± 0.0096.308 ± 1.178
Direct proxy + CDC6.617 ± 0.0006.363 ± 1.216
Reference Drifting + proxy + CDC6.589 ± 0.0076.214 ± 1.134
KGD + proxy + CDC (E)6.547 ± 0.0096.261 ± 1.158

BMMC values summarize three sampling seeds for direct predictors and three training seeds for learned generators. HEST values summarize four outer folds. These standard deviations reflect different sources of variability.

Relative to direct sampling, KGD with the discrete proxy lowers mean ES by 1.5% on BMMC and 3.8% on HEST. With CDC smoothing, configuration E lowers mean ES by 1.1% and 1.6%, respectively. The proxy therefore provides a strong baseline, but its held-out score is not a performance ceiling for the fitted generator.

The new controls also answer the question about the drifting objective. KGD achieves lower mean ES than Reference Drifting when both use the discrete proxy on either dataset. With CDC smoothing, KGD remains better on BMMC, whereas Reference Drifting is better on HEST. We therefore do not claim that KGD is universally the most effective field. The revised manuscript distinguishes CMD, the adaptation through a conditional proxy, from its tested field and smoothing configurations. Our experiments establish the benefit of the proxy within drifting and compare two drifting fields under matched target constructions. They do not establish that drifting is the best objective for learning from that proxy; we return to this point at the end of this response.

We also measured prediction cost on matched inner-validation queries using one A100-80GB GPU. The following values are amortized milliseconds per query for 64 samples, using identical batching within each dataset. They include retrieval, weighting, sampling, and decoding where applicable, while excluding one-time preparation and metric computation.

PredictorBMMC ms/queryHEST-IDC ms/query
Direct proxy0.2250.078
Direct proxy + CDC0.1100.099
KGD + proxy + CDC (E)0.0892.312

The timing advantage is dataset-dependent. Direct sampling is substantially faster on HEST and requires no generator training. A trained generator offers improved predictive scores in these comparisons and inference without retaining or querying the reference bank, but we do not claim that it is always faster than direct proxy sampling.

Regarding objectives that learn directly from individually paired observations, we agree, and we find the suggestion to compare with Engression particularly valuable. Engression minimizes the Energy Score directly, needs only one observed target per condition, and yields a generator that samples in one evaluation [1]. It is therefore a natural comparator for CMD and, because it optimizes the criterion we report, potentially a strong one. Conditional generative moment-matching networks [2] are relevant for the same reason. These methods also show that our motivation should be stated narrowly: with one paired target, it is the positive distribution supplied to drifting that reduces to a point mass, not conditional learning in general.

Unfortunately, we were not able to implement and tune Engression under our protocol on the four benchmarks within the response period, and we did not want to report a rushed baseline. We prioritized the analyses that isolate the components of CMD, which the reviewer also requested: the direct-proxy baselines and the matched comparison of the two drifting fields. Our baseline panel had followed the original drifting study ([3], Table 5), which compares against one-evaluation objectives such as iCT, Shortcut, MeanFlow, and iMeanFlow, and we recognize that this left distributional-regression objectives uncovered.

We therefore make no claim about how CMD compares with Engression. We see two experiments as the natural next step: Engression trained on the single paired targets, as a baseline, and an Energy Score objective trained against our proxy, which would test directly whether drifting is the most effective way to learn from it.

References:

[1] Shen, Xinwei, and Nicolai Meinshausen. "Engression: extrapolation through the lens of distributional regression." Journal of the Royal Statistical Society Series B: Statistical Methodology 87.3 (2025): 653-677.

[2] Ren, Yong, et al. "Conditional generative moment-matching networks." Advances in Neural Information Processing Systems 29 (2016).

[3] Deng, Mingyang, et al. "Generative modeling via drifting." arXiv preprint arXiv:2602.04770 (2026).

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

This paper proposes CMD (Cross-Modal Drifting) to address a key challenge in instance-level cross-modal generation, where each source observation is typically paired with a single observed target, making it difficult to adequately capture the variability of the conditional target distribution. Specifically, CMD assigns weights to reference samples based on their neighborhood distances in the source representation space and applies these weights to the corresponding paired targets to construct a richer conditional target proxy distribution. Building on this proxy, the method trains a generator using a kernel-gradient drifting objective, with an optional local covariance smoothing module (CDC), to enable single-step conditional generation. Experiments cover synthetic tasks with known conditional distributions, as well as two real-world cross-modal generation tasks, BMMC RNA-to-protein generation and HEST-IDC histology-to-gene-expression generation. Moreover, ablation studies and computational cost analyses examine the contributions of individual components and assess the method in terms of generation quality and inference efficiency.

Critical strengths
  1. The paper addresses a practical challenge in instance-level cross-modal generation, where each source observation is typically paired with a single target, making it difficult for the empirical conditional distribution to capture the variability of the underlying target distribution. This provides a clear motivation for developing more effective approaches to conditional distribution learning.
  2. The proposed method is clearly structured. Specifically, CMD exploits neighborhood relationships in the source representation space to construct a weighted proxy for the conditional target distribution, which is then used to train the generator through a kernel-gradient drifting objective. This enables single-step conditional generation while maintaining a focus on both distributional quality and inference efficiency.
  3. The experiments cover synthetic tasks with known conditional laws, as well as two real-world cross-modal generation benchmarks, BMMC RNA-to-protein generation and HEST-IDC histology-to-gene-expression generation. Moreover, the paper evaluates the proposed CMD from both generation quality and inference efficiency perspectives by combining Energy Score (ES), conditional Wasserstein-1 distance on synthetic tasks, and sampling cost analysis.
Critical weaknesses
  1. Although the experiments provide indirect evidence for the effectiveness of the source-neighborhood proxy, its construction relies on the assumption that nearby source representations have similar conditional target distributions. When neighboring samples exhibit different conditional distributions, the weighted proxy may mix incompatible target modes and introduce misleading conditional supervision.
  2. According to Table 1, adding CDC leads to modest reductions in ES on BMMC and W1 on the Circle task, suggesting that local covariance smoothing can provide additional benefits. However, the gains are modest, and it remains unclear how much they depend on the geometry of the target space.
  3. The captions of Figures 2 and 3 appear to conflate M with L. As defined in Section 3.2, M controls the effective neighborhood size of the proxy, whereas L specifies the candidate retrieval budget from the training reference bank. However, the captions describe M=16 and M=64 as the numbers of retrieved training sources, respectively, which is inconsistent with the notation defined in the main text.
Questions for authors
  1. The bootstrap fraction for Shortcut is set to 0.5, whereas the original paper uses 0.25. What motivated this change, and is there empirical evidence supporting the chosen setting?
  2. In Figures 2 and 3, does M refer to the effective neighborhood size of the proxy or the number of retrieved training sources? And how does this relate to the candidate retrieval budget L defined in Section 3.2?
  3. Table 1 shows only marginal gains from adding CDC on BMMC and the Circle task. What properties of the target distribution make local covariance smoothing beneficial?
5 comments

We thank the reviewer for their careful reading and constructive feedback. We appreciate the suggestions concerning the baseline settings, the contribution of covariance smoothing, and the clarity of the figure captions. We address each point separately, explaining the corresponding revisions and presenting the supporting experimental evidence.

The fraction of 0.5 was carried over from a preliminary validation study on BMMC. In that study, a fraction of 1/2 gave lower one-step validation ES than 1/4, but the comparison used short training budgets and 16–32 discretization levels rather than the published 128 levels of our final protocol, and it was not repeated on the other benchmarks. We do not consider this sufficient to justify departing from the published setting.

For consistency with the published method and with our other baselines, which use their published settings, we have retrained Shortcut at the fraction stated in the paper (1/4) on all four benchmarks (34 runs; same seeds, folds, budgets and checkpoint-selection rule, changing only the fraction) and replaced every reported Shortcut performance result. The values change as follows (mean ± SD, one NFE):

Shortcut, submission (1/2)Shortcut, revision (1/4)KGD + proxy + CDC (E)
BMMC, ES7.020 ± 0.0147.004 ± 0.1246.547 ± 0.009
HEST-IDC, ES6.354 ± 1.3047.066 ± 1.3326.261 ± 1.158
Circle, W1 at D = 64 / 128 / 256 / 5121.302 / 3.487 / 11.498 / 19.6651.155 / 3.185 / 11.508 / 19.7550.230 / 0.311 / 0.397 / 0.346
Sketch, W1 at D = 2 / 8 / 64 / 2560.124 / 0.155 / 0.500 / 0.8270.173 / 0.178 / 0.636 / 0.8670.091 / 0.095 / 0.088 / 0.084

The Shortcut reruns preserve the comparison with KGD + proxy + CDC (E), which has lower mean ES or conditional W1 on every benchmark under either bootstrap setting. Among the six competing objectives, CDC-FM at 50 NFE is now the closest comparator to E on HEST-IDC. Our new proxy-based controls, discussed in Section 4.5, additionally show that Reference Drifting with the CDC-smoothed proxy achieves a lower HEST-IDC ES than E.

Table 5 and the accompanying text now state the published fraction.

We agree that the original captions were inconsistent with the notation in Section 3.2. MM denotes the effective neighborhood size of the weighted proxy, while LL specifies the candidate retrieval budget.

The illustrations display the nearest 16 sources in Figure 2 and the nearest 64 in Figure 3, together with their paired targets. These display counts were chosen to equal the selected breadths, M=16M = 16 and M=64M = 64. Both figures use the nominal budget L=512L = 512 and exclude the training query itself. The full candidate sets therefore contain Lq=512L_q = 512 sources for Figure 2 (selected from 5,000 training pairs) and Lq=511L_q = 511 for Figure 3 (all remaining sources among 512 training pairs).

We have revised both captions to identify the displayed points as subsets and to state MM and LqL_q separately. The correction concerns the figure descriptions. Training follows Section 3.2 and samples positives from the full weighted candidate set.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of changes

Main changes

  1. Direct proxy and field comparison (xHWX, KELn, MSej). Section 4.5 and Table 3 compare direct sampling from the proxy with generators trained on it, for both fields, with and without CDC. The proxy improves both fields, and every fitted generator improves on sampling its proxy directly. The best configuration depends on the dataset: KGD + proxy + CDC on BMMC (ES 6.547) and Reference Drifting + proxy + CDC on HEST-IDC (6.214, against 6.261 for KGD + proxy + CDC). The abstract and the list of contributions are updated to reflect this: the contribution remains the use of a source-neighborhood proxy as the conditional target for drifting, now stated for both drifting fields instead of KGD alone.

  2. Shortcut re-run (KELn). We re-ran Shortcut on all benchmarks with the bootstrap fraction of the original paper (1/4 instead of 1/2) and replaced every Shortcut result. Its mean ES changes from 7.020 to 7.004 on BMMC and from 6.354 to 7.066 on HEST-IDC. The CMD variants have a lower mean ES (biological benchmarks) or W1 (synthetic tasks) than Shortcut under either setting, so the re-run changes no conclusion. Its largest effect is on HEST-IDC, where Shortcut was the closest competing method in the submission; that is now CDC-FM at 50 NFE.

Corrections

  1. CDC on HEST-IDC. A few exactly duplicated targets (two to four per fold) could be treated as distinct nearest neighbors, giving those points a near-zero local bandwidth. After correcting this and re-running the four folds, KGD + proxy + CDC changes from 6.237 ± 1.142 to 6.261 ± 1.158 (Pearson 0.514 to 0.510). Its position relative to every other method is the same with either value. BMMC is unaffected.

  2. Training cost. For runs that stopped early, the per-update cost had been computed with the update ceiling instead of the updates performed. Table 6 is recomputed (for example, 8.6 to 13.6 ms for KGD + proxy + CDC on BMMC), its HEST-IDC column now comes from a fixed-window benchmark (Appendix B.4), and Section 4.4 is rewritten accordingly. BMMC training totals are unchanged.

The conclusions on predictive performance are unchanged, and all results not listed here are unchanged.

Reported in the responses

HEST-IDC predictive-mean RMSE, marginal CRPS, and interval coverage (xHWX); checks of the proxy available before training, a neighborhood-mixing analysis on the sketch task, and a local sensitivity study on BMMC (MSej); anonymized code, uploaded as supplementary material.

Changes by table

TableChange
1BMMC KGD + proxy: rounding corrected (6.612 to 6.611)
2Shortcut replaced; HEST-IDC KGD + proxy + CDC corrected
3Expanded: direct proxies, Reference Drifting, and KGD, with and without CDC
5Shortcut bootstrap fraction 0.5 to 0.25
6Per-update costs recomputed; proxy controls added
8–9Shortcut replaced; HEST-IDC KGD + proxy + CDC corrected; controls added
11–15Shortcut replaced

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.