acceptodds
Under review as a conference paper at ICLR 2027

TICDA: Tabular In-Context Data attribution

Abstract

Tabular foundation models (TFMs) achieve strong predictive performance by conditioning on labeled demonstrations provided in context, without any parameter update. Yet how individual demonstrations shape a given prediction remains poorly understood. This gap matters in practice: the context is often assembled from whatever labeled data is available, potentially leading to the inclusion of mislabeled, redundant, or low-quality examples that degrade performance. Standard data attribution methods do not transfer to TFM setting: resampling-based approaches such as DemoShapley require a combinatorial number of forward passes, and gradient-based estimators such as influence functions require to compute training point's effect on the model parameters, which in-context learning never updates. We introduce TICDA, a method that measures the influence of every demonstration in the context directly from linear surrogates trained on TFM latent embeddings, in a single forward pass and at negligible cost. We show that TICDA offers the best compromise against competitors across four tasks: detecting labeling errors, curating context to preserve predictive accuracy while lowering inference cost, producing attribution scores that transfer across TFMs, and supporting an acquisition strategy for efficient active learning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
data attribution, influence function, tfm, tabular fondation model, explainability, interpretability

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_f69qOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates how to identify low-quality in-context demonstrations for tabular foundation models. It proposes TICDA, a simple method that extracts internal model representations, fits a ridge regression surrogate, and computes demonstration influence scores using closed-form expressions. The authors evaluate TICDA’s computational efficiency on 38 classification datasets. They also explore applications to context curation, cross-model transfer, and active learning.

Critical strengths

[S1] The paper addresses an important and challenging problem: identifying low-quality in-context demonstrations at low computational cost.

[S2] The method is simple and computationally inexpensive, requiring only a ridge regression fit on extracted TFM representations.

Critical weaknesses

[W1] As shown in Table 1, TICDA’s average detection performance across 38 datasets on TabDPT is significantly lower than that of DemoShapley (0.79 vs. 0.91).

[W2] The experiments cannot be reproduced from the information provided. The paper does not specify the names of the 38 datasets, their sizes, or how the in-context demonstrations are constructed.

[W3] TICDA uses an additional labeled validation set that is neither counted toward the labeling budget nor used by the baselines. Thus, the active learning comparisons are unfair.

Questions for authors

[Q1] TICDA relies on an attribution surrogate and the fixed-representation simplification in Equation (6). The authors should provide theoretical explatation or empirical experiments clarifying how these two approximations affect agreement with the TFM’s LOO influence in Equation (1), particularly its magnitude and sign.

[Q2] Please provide the per-dataset results underlying Table 1 to clarify whether the findings are consistent across datasets.

4 comments
AuthorsOpenReview ↗

Answer to reviewer f69q (1/2)

We thank reviewer f69q for their insighful comments. Please find our answer below:

[...] TICDA’s average detection performance across 38 datasets on TabDPT is significantly lower than that of DemoShapley (0.79 vs. 0.91)

Let us kindly clarifiy this result. The 0.79 vs 0.91 gap concerns faithfulness, not detection. On labeling error detection, which is the practical task, TICDA and DemoShapley are tied on TabDPT (0.88 vs 0.88). On faithfulness for TabDPT, DemoShapley does score higher than TICDA, but, it is roughly 1000 times more expensive to compute than TICDA, as shown in Figure 2.

The experiments cannot be reproduced from the information provided. The paper does not specify the names of the 38 datasets [...]

All 38 datasets are public. They are the 38 classification datasets of the TabArena benchmark [1], and each one can be downloaded from OpenML with its ID. We follow the TabArena preprocessing and evaluation protocol and limit each context to at most 4,096 demonstrations.

TICDA uses an additional labeled validation set that is neither counted toward the labeling budget nor used by the baselines. Thus, the active learning comparisons are unfair.

We thank the reviewer for this important and insightful comment. We reran the experiment on 36 TabArena datasets. We also fixed the protocol so that our approach and its competitors have exactly the same labeling budget: we removed the access to validaiton set labels. TICDA now scores each candidate from the self influence of its addition, based on the TFM predicted label. The new results are below:

ComparisonAULC difference
TICDA vs entropy+0.18
TICDA vs random+0.97

With the same labeling budget for every strategy, TICDA achieves a higher area under the accuracy learning curve (AULC) than both entropy and random sampling. This shows that data attribution can be used to build a competitive acquisition strategy for TFMs. We added these experiments in the updated version of our paper.

AuthorsOpenReview ↗

Answer to reviewer f69q (2/2)

[...] The authors should provide theoretical explatation or empirical experiments clarifying how these two approximations affect agreement with the TFM's LOO influence [...]

In short, the two approximations largely preserve the ranking of the demonstrations, which is what Tables 1–2 measure, but they do not preserve the magnitude of the TFM's LOO effect, and the self-score TiT_i carries no sign information. TICDA's scores are therefore ranking scores, not calibrated estimates of the LOO effect. For the applications of the paper, the rank agreement is what matters, and not the magnitude: label-error detection ranks the demonstrations by TiT_i, and context curation removes the lowest-ranked demonstrations under a fixed budget (Section 3.1). We will state this limitation explicitly in Section 2.3 and in the conclusion, and the new Appendix F of the revised paper reports the experiment below.

Theory: Appendix B.2 shows that, inside the surrogate, the self-score TiT_i is nonnegative and never exceeds the exact deletion effect, and that for another query the exact loss change can be positive when the first-order influence is negative, but never the reverse. These properties give no guarantee about the effect of the TFM, because the representations change when a demonstration leaves the context and the surrogate loss differs from the loss of the TFM (Appendix B.3).

Experiment: We therefore measured the magnitude and the sign against the native effect, that is, the TFM's own change of log loss when a demonstration is deleted, which is the reference of Tables 1–2. To compare magnitudes in one unit, we evaluate the ridge surrogate with a log loss. Table 8 of the revised paper (Appendix F.3) reports two estimates of this surrogate, one with fixed representations and one after the TFM has encoded the reduced context again. For each, it gives the error relative to the size of the native effect, so that predicting no change gives an error of 1, and the sign agreement with the native effect. As a reference for the sign, it also gives the agreement of a score that is positive for every deletion, the always-positive baseline.

Sign: Most native self-deletions raise the loss, so a score that is always positive already agrees with the TFM in 93–100% of the deletions. TiT_i is such a score: it is nonnegative by construction, and it cannot reproduce the negative native effects, which make up the remaining 7% of the TabICL deletions and 1% of the TabDPT deletions at the submitted sizes. Neither of the two estimates exceeds this baseline in any row of Table 8.

Magnitude: The surrogate recovers little of the magnitude of the native effect: with fixed representations, its errors of 0.74–0.98 are only moderately below the error of 1 that predicting no change gives. Recomputing the representations lowers the error by 0.04–0.19 in every row but one, and the errors of 0.61–0.93 that remain are due to the surrogate, including the uncalibrated probability mapping.

Ranking: By contrast, the Spearman correlation of TiT_i with the native effect is 0.64–0.79 for TabICL and 0.76–0.83 for TabDPT across the four context sizes (Table 7, last column).

[...] Please provide the per-dataset results underlying Table 1 to clarify whether the findings are consistent across datasets.

In the revised version, we added Appendix E.1 with a table that gives the per-dataset results underlying Table 1, for both TabDPT and TabICL. For each dataset, the table lists its OpenML ID, the labeling error detection AUC of all methods, and the faithfulness of each method.

The per-dataset results are consistent with the averages in Table 1:

  • TabDPT, detection. TICDA and DemoShapley are within 0.05 AUC of each other on 31 of the 38 datasets.
  • TabDPT, faithfulness. DemoShapley has higher faithfulness than TICDA on 36 datasets. Unlike TICDA, it re-evaluates the TFM on eight permuted contexts, which makes it two to three orders of magnitude slower (Figure 2). TICDA still keeps a faithfulness above 0.7 on 32 datasets, and is above IG and DETAIL on all 38.
  • TabICL, detection. TICDA has a higher AUC than DemoShapley, IG and DETAIL on all 38 datasets, and is above LOO on 31 of them.
  • TabICL, faithfulness. TICDA has the highest faithfulness on average (0.67 vs 0.65 for DemoShapley), and is higher than DemoShapley on 21 of the 38 datasets. This way, the findings are consistent across datasets.

References

[1] Nick Erickson, Lennart Purucker, Andrej Tschalzev, David Holzmüller, Prateek Mutalik Desai, David Salinas, and Frank Hutter. TabArena: A Living Benchmark for Machine Learning on Tabular Data. NeurIPS, 2025.

Thank you for your response. It addresses my concerns. I will raise my score to 3.

Dear reviewer, Thank you for your comment and updating your score. Kind regards,

Reviewer_8srAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TICDA attributes the effect of in-context demonstrations in tabular foundation models using ridge surrogates fitted to frozen row representations. It derives efficient self- and query-influence scores and evaluates attribution faithfulness, label-error detection, context curation, transfer between models, and active learning. The central practical benefit is avoiding repeated evaluation of the full foundation model for individual deletions.

Critical strengths
  1. The ridge-surrogate approach addresses the difficulty of attributing demonstrations in models that adapt through context, and offers a useful runtime–quality tradeoff relative to exhaustive leave-one-out evaluation.
  2. The evaluation covers many tabular datasets and downstream uses. The appendix distinguishes first-order influence from exact fixed-design deletion and examines representation changes, making the approximation assumptions explicit.
Critical weaknesses
  1. The active-learning comparison does not match total labeling cost. On p. 8, lines 400–402, V contains 20% of the final budget, with at least 32 labeled examples, but these labels are excluded from the budget and unavailable to baselines. Table 4 therefore does not establish greater label efficiency at equal annotation cost. Count V in the budget and equalize resources, or narrow the claim to an externally supplied labeled validation set. This issue does not by itself invalidate attribution or curation.
  2. The faithfulness target is ambiguous. Equation 1 concerns query loss after deleting a demonstration, but the strongest results use self-influence on a demonstration representation. Appendix B.1 notes that a row can have different representations in demonstration and query roles. Specify the LOO target, loss, roles, and query aggregation in Tables 1–2 to establish that the scores are evaluated against the same quantity. This is a clarification request, not an established implementation error. (pp. 2, 4–6, 13–14)
Questions for authors
  1. Does the active-learning advantage remain when all validation labels count toward the budget and baselines receive comparable resources?
  2. What exact deletion effect is the LOO reference for self- and query-influence in Tables 1–2?
  3. Against that same target, how much error comes from influence linearization, frozen representations, and surrogate mismatch, particularly in small contexts?
3 comments
AuthorsOpenReview ↗

Reponse to reviewer 8srA (1/2)

We thank reviewer 8srA for their insightful comment. Below is our answer.

The active-learning comparison does not match total labeling cost [...]. Does the active-learning advantage remain when all validation labels count toward the budget and baselines receive comparable resources?

We thank the reviewer for this important and insightful comment. We reran the experiment on 36 TabArena datasets. We also fixed the protocol so that our approach and its competitors have exactly the same labeling budget: we removed the access to validaiton set labels. TICDA now scores each candidate from the self influence of its addition, based on the TFM predicted label. The new results are below:

ComparisonAULC difference
TICDA vs entropy+0.18
TICDA vs random+0.97

With the same labeling budget for every strategy, TICDA achieves a higher area under the accuracy learning curve (AULC) than both entropy and random sampling. This shows that data attribution can be used to build a competitive acquisition strategy for TFMs. We added these experiments in the updated version of our paper.

The faithfulness target is ambiguous. Equation 1 concerns query loss after deleting a demonstration, but the strongest results use self-influence on a demonstration [...]

Thank you for your comment. Please find our clarifications below:

Section 3.1 points to Equation 1, which is written for a generic query. All scores in Tables 1–2 are evaluated against the same quantity, which we specify below. We will add this specification to Section 3.1 and to the captions of both tables.

Target: Every faithfulness value in Tables 1–2, for every method and every row, is computed against one reference vector per dataset: the TFM's native self-LOO effect, which is Equation 1 with the removed row itself as the query. We remove a labeled demonstration, present the same row to the TFM as an unlabeled query, and measure the change of its log loss against its observed label. This observed label remains the target even when we have corrupted it.

Loss and roles of the scores: TICDA's self-score uses the representation of its related row in its role of labeled demonstration and the squared-error loss of the ridge surrogate. As Appendix B.1 notes, this representation can differ from the one that the TFM computes for the same row as an unlabeled query. Neither the loss nor the representation role of the self score TiT_i therefore coincides with those of the reference: TiT_i is an approximate score, and Tables 1–2 measure how closely its ranking agrees with native self-LOO.

Query aggregation: The query-influence rows of Table 2 rank demonstrations by −ui-u_i, where uiu_i is the mean influence of Section 2.3 under the same squared-error loss, averaged over clean labeled examples of the held-out split. We flip the sign so that a large score marks a demonstration that hurts these examples, as a large TiT_i marks a suspect label. In Table 2 the held-out split serves only as this query set. The reference for these rows is again native self-LOO. Their faithfulness measures agreement with self-deletion rankings. We will clarify that in the text.

Dataset aggregation: Faithfulness is the Spearman correlation over the demonstrations of one dataset. We average it with equal weight over the 38 datasets, and the ± values are standard deviations across datasets.

Does the active-learning advantage remain when all validation labels [...]

See our response above.

What exact deletion effect is the LOO reference for self- and query-influence in Tables 1–2?

The reference is the change of the TFM's log loss on row ii when the labeled demonstration ii leaves the context and the same row is queried without its label. Let D\mathcal D be the labeled context and D∖i\mathcal D_{\setminus i} the context without demonstration (xi,yi)(x_i,y_i), as in Equation 1. Let pf(c∣x;D)p_f(c\mid x;\mathcal D) be the probability that the TFM ff assigns to class cc for an unlabeled query xx given the context D\mathcal D (that is, pfp_f is the probability output of the TFM), and let cic_i be the class of the observed label yiy_i, including any corruption that we injected. The reference is the log loss after the deletion minus the log loss before it:

$ \log p_f(c_i\mid x_i;\mathcal D) -\log p_f(c_i\mid x_i;\mathcal D_{\setminus i}), $

where both terms use the same class cic_i. This is Equation 1 with q=iq=i, the log loss as LL, and the query presented without its label. A positive value means that removing the demonstration lowers the TFM's confidence in the observed label, and a negative value means that the removal raises it.

Each of the two predictions runs the whole pipeline on its context with the pretrained weights fixed. We compute this effect for every demonstration of a context. The resulting vector is the reference for the self-influence and the query-influence rows alike, and for every method of Table 1.

AuthorsOpenReview ↗

Reponse to reviewer 8srA (2/2)

Against that same target, how much error comes from influence linearization [...]

We ran a new experiment that answers this question against the native self-LOO target of Tables 1–2, and we added it to the revised paper as a new Appendix F. We summarize the conclusions here:

  • Linearization costs almost nothing: it changes the rank agreement with the TFM by at most 0.01 at every context size.
  • Frozen representations (the fixed representations of the paper) cost rank agreement in the smallest contexts only (0.15 for TabICL at 64 demonstrations). In larger contexts they change it by less than 0.03.
  • Surrogate mismatch is the main source of error at every size. Once the other two approximations are removed, the surrogate still ranks the demonstrations with a correlation of 0.74–0.83 with the TFM in all settings but one, and the gap to 1 is due to the surrogate.

These conclusions support the motivation of TICDA. TFMs usually operate at the scale of thousands of demonstrations or more, in which a single demonstration is a small fraction of the context (Section 2.2). This is the regime in which the two approximations that TICDA adds to the surrogate cost the least: at the submitted sizes, neither linearization nor frozen representations costs rank agreement. Both approximations are therefore sound for ranking demonstrations in the contexts that TICDA targets, and the small contexts in which frozen representations cost more lie outside this regime.

Dear Reviewer,

We posted a detailed response to your review, including additional experiments. Since the discussion period ends today, we would be grateful if you could take a look and let us know whether it addresses your concerns. If it does, we would be grateful if you could consider updating your score to reflect this.

If anything remains unclear, we can reply quickly today.

Best regards,

Reviewer_3DpCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Tabular Foundation Models (TFMs) achieve state-of-the-art performance on tabular prediction tasks by conditioning on labeled demonstrations in their context, without any parameter updates. However, little is understood about how TFMs use their context, specifically how each individual demonstration influences a given prediction. This paper introduces TICDA (Tabular In-Context Data Attribution), a gradient-based method that measures the influence of every in-context demonstration directly from linear surrogates trained on TFM latent embeddings, in a single forward pass efficiently. The paper shows that TICDA offers the best compromise across four tasks: detecting labeling errors, curating context to preserve predictive accuracy while lowering inference cost, producing attribution scores that transfer across TFMs, and supporting an acquisition strategy for efficient active learning.

Critical strengths
  • The paper introduces an approach that estimates the influence of in-context demonstrations in TFMs and does not require parameter updates

  • TICDA runs in under 3 seconds (orders of magnitude faster than other methods)

  • The empirical evaluation is extensive (across 38 datasets, two TFM backbones, and 4 tasks)

Critical weaknesses
  • The experiments focus on classification while extension to regression and multimodal settings are left to future work.

  • TICDA relies on the quality of TFM latent representations. If the representations capture the task structure poorly, attribution quality may decrease

  • The method is efficient, but it requires extracting all representations upfront, which may be memory-intensive for very large contexts.

  • The paper does not evaluate performance on regression tasks, which are common in tabular data.

  • The active learning experiment uses a limited number of datasets and a specific protocol and broader validation is needed.

Questions for authors

How does TICDA perform on regression tasks (e.g., predicting continuous values) compared to classification?

  • Can the method be extended to multimodal TFMs (e.g., combining tabular and text data)?

  • How sensitive are the attribution scores to the choice of the hidden layer where representations are extracted?

  • Why does the self-influence score detect labeling errors? Is it because it measures whether the demonstration's label is supported by the rest of the context or something else?

  • Can TICDA be integrated into existing TFM training pipelines to improve context construction?

  • How does the method perform on out-of-distribution contexts?

3 comments
AuthorsOpenReview ↗

Response to reviewer 3DpC (1/2)

We thank reviewer 3DpC for their insightful comments. Below are our answers.

The experiments focus on classification [...]

We focused on classification because it is the most widely studied setting in the data attribution literature [1, 2]. Our influence function formulation is not specific to classification and extends directly to regression. We are currently running regression experiments and will report the results as soon as they are available.

TICDA relies on the quality of TFM latent representations [...]

Our method is indeed sensitive to the choice of the hidden layer from which the representation is extracted. To highlight this, we ran additional experiments over all layers of both TFM backbones. Attribution scores are stable up to approximately layer 20 for TabDPT and up to the pre-ICL representation for TabICL, and then degrade in deeper layers. The layers used in the paper lie inside this stable region. We added the corresponding figure to the Appendix of the new version.

The method is efficient, but it requires extracting all representations upfront [...]

Thank you for this comment. The stored matrices can become large when the context is long, but this is not a blocker: dimension reduction can be applied (e.g., random projection to 20% of the dimensions) while still yielding good results, as already shown in Table 2.

The active learning experiment uses a limited number [...]

Thank you for this insightful and important comment. We reran the experiment on 36 TabArena datasets. We also fixed the protocol so that our approach and its competitors have exactly the same labeling budget: we removed the labeled validation set. TICDA now scores each candidate from the influence of its addition on the TFM's predicted probabilities, so it uses no labels beyond those it acquires.

ComparisonAULC difference
TICDA vs entropy+0.18
TICDA vs random+0.97

With the same labeling budget for every strategy, TICDA achieves a higher area under the accuracy learning curve (AULC) than both entropy and random sampling. This shows that data attribution can be used to build a competitive acquisition strategy for TFMs. We added these experiments in the updated version of our paper.

Can the method be extended to multimodal TFMs [...] ?

Yes, since our approach only need hidden representations from the model. The exact same methodology can then be applied.

How sensitive are the attribution scores to the choice [...] ?

See our response above.

Why does the self-influence score detect labeling errors?

Yes. A demonstration gets a high self-influence when its label disagrees with what the surrogate predicts from the rest of the context.

Can TICDA be integrated into existing TFM training pipelines to improve context construction?

Could you clarify whether your question concerns pretraining (e.g., selecting or cleaning the datasets used to pretrain a TFM), or the construction of the context at inference time?

References

[1] Pang Wei Koh and Percy Liang. Understanding Black-box Predictions via Influence Functions. ICML, 2017.

[2] Sung Min Park, Kristian Georgiev, Andrej Ilyas, Guillaume Leclerc, and Aleksander Madry. TRAK: Attributing Model Behavior at Scale. ICML, 2023.

AuthorsOpenReview ↗

Response to reviewer 3DpC (2/2)

Dear reviewer,

Please find below our additional answers to the points you raised in your review:

The experiments focus on classification while extension to regression and multimodal settings are left to future work.

We ran two new experiments, one on regression and one on multimodal data to show how TICDA applied to other settings.

Regression extension

Our influence function formulation is not specific to classification and extends directly to regression. We trained a linear regressor instead of a linear classifier as the surrogate on the TFM latent embeddings. For evaluation, we used the 13 regression datasets of TabArena. For each dataset, we corrupted 20% of the context targets by adding Gaussian noise, y_i ← y_i + s · σ_y · z_i with z_i ~ N(0, 1), where σ_y is the standard deviation of the clean context targets and s is the noise intensity. We report s = 2 and s = 3 below. Results are mean ± std over datasets.

s = 2

BackboneMetricTICDALOODataSHAPIGDETAILDETAIL-128
TabICLAUC0.834 ± 0.060.770 ± 0.070.820 ± 0.050.764 ± 0.050.455 ± 0.080.492 ± 0.07
Faithfulness0.551 ± 0.11n/a0.812 ± 0.080.634 ± 0.10-0.052 ± 0.09-0.011 ± 0.08
TabDPTAUC0.838 ± 0.050.878 ± 0.060.878 ± 0.060.712 ± 0.070.520 ± 0.030.531 ± 0.03
Faithfulness0.661 ± 0.10n/a0.910 ± 0.050.476 ± 0.170.013 ± 0.070.033 ± 0.07

s = 3

BackboneMetricTICDALOODataSHAPIGDETAILDETAIL-128
TabICLAUC0.872 ± 0.040.787 ± 0.070.849 ± 0.050.784 ± 0.060.411 ± 0.090.464 ± 0.09
Faithfulness0.528 ± 0.10n/a0.798 ± 0.080.612 ± 0.10-0.080 ± 0.08-0.032 ± 0.08
TabDPTAUC0.878 ± 0.040.909 ± 0.050.911 ± 0.040.720 ± 0.070.522 ± 0.050.536 ± 0.04
Faithfulness0.662 ± 0.09n/a0.908 ± 0.050.450 ± 0.160.021 ± 0.080.043 ± 0.07

These results show that TICDA transfer well to regression, with detection improving as the noise grows, still at roughly 1000x less compute than LOO and DataSHAP (as shown in Figure 2).

Multimodal (text + tabular)

We use 6 text + tabular classification datasets from MulTaBench (hosted on OpenML, with MulTaBench's column selection). Each text column is embedded with MulTaBench's frozen e5-small-v2 sentence embeddings (384 dimensions), standardized and reduced to 30 dimensions by a PCA fitted on the context rows. The TFMs receive the non-text columns plus these 30 features per text column, so TICDA is applied unchanged. 20% of the context labels are flipped to another class for labelling error detection.

BackboneMetricTICDALOODataSHAPIGDETAILDETAIL-128
TabICLAUC0.876 ± 0.100.867 ± 0.090.785 ± 0.140.789 ± 0.110.398 ± 0.140.501 ± 0.09
Faithfulness0.751 ± 0.08n/a0.728 ± 0.200.475 ± 0.09-0.130 ± 0.170.033 ± 0.14
TabDPTAUC0.859 ± 0.110.887 ± 0.100.851 ± 0.150.666 ± 0.170.345 ± 0.140.437 ± 0.13
Faithfulness0.710 ± 0.14n/a0.877 ± 0.170.417 ± 0.30-0.292 ± 0.20-0.106 ± 0.22

These results also show that TICDA applies unchanged to text + tabular data once text is embedded into the tabular input, leading to the best compromise in terms of AUC, faithfulness and computational cost.

We will add these experiments to the revised paper.

Dear Reviewer,

We posted a detailed response to your review, including additional experiments. Since the discussion period ends today, we would be grateful if you could take a look and let us know whether it addresses your concerns. If it does, we would be grateful if you could consider updating your score to reflect this.

If anything remains unclear, we can reply quickly today.

Best regards,

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.