acceptodds
Under review as a conference paper at ICLR 2027

CARVEPrep: Model-Aware Data Preparation for Tabular Foundation Model Post-Training

Abstract

Tabular foundation models such as TabPFN are post-trained on task tables, where each record both updates the parameters and serves as context for prediction. Preparing a dirty table thus amounts to allocating a data action to each record. Detecting an error does not show that fixing it helps the model, and action effects need not add, so these decisions should be judged by the trained model rather than by data correctness. We present CARVEPrep, which flags errors from dependency violations and label checks, trains a predictor on self-injected errors, without true values, to choose each record's action, and keeps a prepared table only when repeated short post-training shows that it beats the unchanged one. Feedback from the trained model then reprioritizes records for the next round. We prove that, under the same budget on changed records, removing repair or partial weights can strictly hurt. On six tables with injected errors, CARVEPrep averages 0.8119 accuracy, against 0.7343 for the unchanged data and 0.7853 for the strongest of fourteen baselines, which rewrites 70 times as many cells, and matches or beats every baseline on four of six tables. With data-specific evidence, the same actions carry over to sensor time series, where CARVEPrep ranks first on five of six fault scenarios, and to instruction tuning of Qwen2.5-1.5B, where held-out loss drops from 1.396 to 1.312. Tuning data and post-training hyperparameters together matches or beats tuning either alone on 11 of 12 tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
data-centric AI, data preparation, data quality, data cleaning, tabular foundation models, post-training, joint data and model optimization

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

This paper proposes CARVEPrep, a model-aware data preparation framework that jointly decides whether to retain, repair, delete, or downweight each record. Candidate preparations are evaluated through downstream model feedback, and the resulting losses are further used to update record priorities across rounds. The method is evaluated on tabular data, sensor time series, and instruction tuning, with extensive comparisons and ablations.

Critical strengths
  1. The paper provides a compelling formulation of data preparation as model-aware action allocation rather than simply maximizing data correctness.

  2. The experimental evaluation is extensive, covering many baselines, ablations, multiple data types, scale studies, and held-out acceptance tests.

Critical weaknesses
  1. The action-value regressors are trained using self-injected error types that closely match the synthetic corruption used in the main evaluation, so robustness to natural or mismatched errors remains unclear.

  2. CARVEPrep evaluates many candidate preparations through repeated downstream post-training, but the paper does not provide a compute-matched comparison with the baselines. It is therefore unclear how much of the gain comes from the proposed allocation strategy versus additional search budget.

Questions for authors
  1. How does CARVEPrep perform when the test-time corruption types differ from those used in the self-injected perturbation trials?

  2. Can the authors provide a comparison under a matched candidate-evaluation or post-training compute budget?

  3. For the instruction-tuning experiment, can the full learned action-value mechanism be evaluated instead of the evidence-based allocation rule?

5 comments
AuthorsOpenReview ↗

Response to Reviewer DEGn (1 of 2)

Thank you for the focused review and for finding the model-aware formulation compelling. Below are the new runs for mismatched errors, matched compute and the learned values on text.

To W1 and Q1, In our design the trials only determine which candidates are proposed, and a model-based test on the reference set determines whether a candidate is adopted. Based on your question, we added three comparisons using new feature errors, pair-flip label noise or both (Appendix Table 13, Table D1 below). CARVEPrep ranks first among the non-oracle methods run in all three settings. The mean accuracies are 0.846 for new feature errors with pair-flip labels, 0.875 for new feature errors alone and 0.849 for pair-flip labels alone, against the corresponding strongest baselines at 0.800, 0.867 and 0.811. These are changes of error family, evaluated on the submitted data seeds with three training runs per entry. Separately, changing the perturbation trials while keeping the submitted training-table errors leaves accuracy within 0.2 points of the default-trial rerun.

Table D1 (new runs, Section 4.3, Appendix Table 13). Six-table means of test accuracy over three training runs, rerun on the submitted data seeds. Strongest baseline is the best of the baselines run in that setting. Picket cannot run on Adult and Breast Cancer under added feature errors alone, since its prepared labels exceed the ten classes TabPFN-v2 supports. Flights has no numeric column for the three error types of that setting, so Flights keeps its clean table there.

SettingNoFixStrongest baselineCARVEPrepClean
Added feature errors and pair-flip labels0.7700.800 (RHO-LOSS)0.8460.875
Added feature errors alone0.8670.867 (BoostClean)0.8750.876
Pair-flip labels alone0.7900.811 (RHO-LOSS)0.8490.876
Trials without label redraws, submitted errors0.7340.785 (Jellyfish)0.8040.876
Trials without dependency violations and typos, submitted errors0.7340.785 (Jellyfish)0.8050.876

The first three settings differ from the trials in kind, since the added feature errors are heavy-tailed outliers, unit inconsistencies and cross-column dependency breaks that ignore any declared dependency (Appendix C.8). In the trials, dependency violations rewrite only values that a declared functional dependency determines. Pair-flip noise sends each class to one fixed class, whereas the trials redraw labels.

CARVEPrep adopts a preparation only after it passes the gain margin and paired bootstrap test on the reference set, whose records carry no injected errors. Records with rejected labels are deleted by the allocation rule without using the learned values, so their deletion does not depend on the trials.

On the natural dirty versions of Beers, Flights, Hospitals and Tax, CARVEPrep retains the original table and its mean accuracy of 0.854, while eleven of the fourteen baselines fall below it. This is within 0.35 points of clean training at 0.857. These runs evaluate natural feature errors with intact labels and no injected corruption (Appendix C.9).

AuthorsOpenReview ↗

Response to Reviewer DEGn (2 of 2)

To W2 and Q2, The gain holds without the additional search budget. We evaluated CARVEPrep with in-context adoption, removing short post-training from candidate comparison (Appendix Table 14). It reaches 0.779, above Jellyfish's 0.741. CARVEPrep takes 97 seconds per run, compared with 366 for Jellyfish on the same hardware (Table D2). These times include preparation and final training. Over five data seeds, the gains are 7.8 points over NoFix ([4.7, 11.6]) and 3.8 over Jellyfish ([1.6, 6.4]), at about a quarter of Jellyfish's cost. We ran a second control with the same adoption rule and the same number of short runs across candidate sets, as detailed below. Appendix C.8 of the submission specifies the fixed search and equal short-training counts per finalist.

Under the same adoption rule and short-training budget, combining CARVEPrep candidates with the fourteen baseline tables reaches 0.816, above either candidate set alone (Appendix Table 3). The means are 0.803 for CARVEPrep candidates and 0.814 for the baseline tables plus selection candidates. CARVEPrep's candidates give higher accuracy than the baseline tables on five of six tables on the submitted data seeds. Over five seeds they do so on four of six tables, while the baseline tables under the rule reach a mean of 0.784.

Table D2 (new runs, Section 4.3, Appendix Tables 9 and 14). Five-seed six-table means. Time in seconds per run covers preparation and one final training on shared hardware.

MethodSecondsShort post-training runsAccuracy
NoFix2400.701
RHO-LOSS2000.727
Ref. likelihood1900.727
Raha+Baran23900.708
Jellyfish36600.741
CARVEPrep, in-context adoption9700.779
CARVEPrep-ER (external repairs, in-context adoption, no record cap)111 plus the external methods (605)00.795
CARVEPrep82218 to 240.771

To Q3, We ran the full learned action-value mechanism on Alpaca over three seeds. It reduces held-out loss from 1.397 for NoFix to 1.336, and deletion by reference loss gives 1.339 in the same runs. The new experiment injects text errors in the trials, fits per-action value regressors on the text evidence and uses the same allocation rule as for tables (Appendix Tables 7 and 17, Appendix C.5). The submission used the evidence rule on text for a practical reason. In-context loss does not separate a clean pool from a corrupted one there, so every perturbation trial needs a short LoRA training run, and a run with learned values takes about 10,000 seconds against about 1,500 for the evidence rule (Appendix C.5).

We welcome further discussion if any concerns remain.

Dear Reviewer DEGn,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

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

This paper presents CARVEPrep, a model-aware data preparation framework for post-training tabular base models. Rather than just focusing on data correctness it treats data preparation as a record level assignment between retention, repair, deletion and downweighting. It combines structural error evidence with self-supervised regressors that estimate the utility of each action and iteratively evaluates against the downstream model to select the prepared dataset.

Experiments on six tabular datasets report a mean performance of 0.8119 as opposed to 0.7853 for the best baseline and 0.7343 for the unchanged data. Further experiments are conducted on sensor time series and LLM instruction tuning, along with a number of conceptual analyses of action interactions and acceptance guarantees.

Critical strengths
  • Well motivated problem formulation. Treating data preparation as a model dependent decision over a range of record level actions is very much a realistic scenario. The paper also acknowledges the interactions between actions which independent record ranking is unable to capture.
  • Substantial experimental evaluation. The comparison against a range of cleaning and selection baselines, along with candidate set studies, ablations and experiments outside of the normal tabular classification scenario, all provide pretty convincing evidence for the overall framework.
Critical weaknesses
  • Mismatch between the formal budget and the actual preparations evaluated. The proposed formulation limits the number of records that can be intervened upon, but the full candidate selection process is able to make use of preparations which go beyond that budget. For instance the selected Adult preparation intervenes on 523 of the 900 records even though the nominal cap is 0.3 record. This is all acknowledged but it does mean that the link between the constrained optimisation formulation and the headline results is somewhat obscured.
  • Insufficient isolation of the novel mechanism. The final system is effectively selecting from the learned allocations and the existing preparation strategies so it's very difficult to isolate how much of the improvement is actually due to the self-supervised action regressors as opposed to the more general candidate search and model-based selection. Related work in the area of model guided cleaning and joint data-model optimisation further underpins the need for a more rigorous experimental separation. The generalisation claims would also be strengthened by testing across a range of corruption seeds and downstream tabular base models.
Questions for authors
  1. Can the authors report results when all adopted preparations fall within the same intervention budget (including the composite baseline-derived candidates)? And how does that change the primary comparison?
  2. To what extent is the improvement due specifically to the learned record-wise action allocator as opposed to simply selecting amongst the various preparation strategies under matched evaluation and computation budgets?
  3. How sensitive are the results to the quality and size of the trusted reference set, to different corruption realisations and to the reuse of reference data during the course of iterative selection? Which of the results reported actually use fully independent acceptance samples?
5 comments
AuthorsOpenReview ↗

Response to Reviewer tLVA (1 of 2)

Thank you for the constructive review and for recognizing that record-level preparation reflects a realistic setting. The new runs you asked for are reported below.

To W1 and Q1, We reran the main comparison with every candidate constrained to the same record budget (Table 1 of the revision). CARVEPrep retains its six-table mean lead over all fourteen external baselines. In the submitted Adult preparation, the cap applies to the deletion stage. It deletes 270 records, exactly the cap, and verified repairs then act on 253 retained records, giving 523 intervened records (Appendix C.3 of the submission). The other five submitted preparations stay within their caps. Your comment shows that this scope should be stated with the formulation and that the main result should also be given under one budget. The revision does both, stating the scope of the budget in Section 2 and reporting the constrained run in Table 1.

With every candidate constrained to the record budget, CARVEPrep reaches a five-seed six-table mean of 0.769, compared with 0.771 for the standard setting (Table 1 of the revision). Both settings remain above all fourteen external baselines by six-table mean. The change is −0.2-0.2 points ([−1.4-1.4, 1.2]), with gains of 6.8 points over NoFix ([3.9, 10.3]) and 2.8 over Jellyfish ([0.4, 5.6]). Candidates above the cap are removed or scaled down. On the submitted data seeds the mean changes from 0.8028 to 0.8021. Five tables adopt the same candidate and only Adult changes, by −0.44-0.44 points. The revised row also reports 0.986 balanced accuracy over six TEP scenarios, 14.1 points above Ref. likelihood. Its TEP candidates include external feature repairs combined with CARVEPrep deletions under the same record cap. Appendices C.8 and C.9 give the configurations and runs.

To W2 and Q2, Our thinking is that actions interact, so record-wise estimates propose the actions and the trained model judges complete preparations. The improvement comes from both parts, and based on your suggestion we measured each of them with comparisons of candidate sets and an evidence-rule control (Appendix Table 3 and Appendix C.9). The controlled comparisons separate the gains from model-based adoption, record-wise allocation and candidate composition. Table T1 varies the candidate set under one adoption rule, with the same number of short runs per table. These experiments extend the nested comparison of Appendix C.1 of the submission.

(a) The adoption rule alone adds 2.7 points to the best single baseline on the submitted data seeds and raises it from 0.741 to 0.784 over five seeds.

(b) Both implementations of the action rule preserve the gain. Replacing the learned regressors with a closed-form evidence rule reaches 0.773 against 0.771 over five seeds. Under label errors, rejected-label evidence determines the deletions. Appendix C.9 reports this evidence-rule comparison.

(c) Record-wise allocation adds 0.4 points on the submitted data seeds over CARVEPrep's candidates without it, with the largest table-level gain, +5.4+5.4, on Hospitals.

Table T1 (new runs, Appendix Table 3). Six-table means on the submitted data seeds, three training runs.

Candidate setMeanTables adopting a record-wise allocation
Best single baseline (Jellyfish), no adoption0.7850
The fourteen baseline tables0.8120
The fourteen baseline tables and the selection baselines0.8140
CARVEPrep candidates without record-wise allocations0.7990
CARVEPrep candidates0.8034
CARVEPrep candidates and the fourteen baseline tables0.8162

With in-context adoption and the record budget lifted, CARVEPrep-ER reaches a five-seed mean of 0.795, exceeds the best of the fourteen baselines on every table and ranks first among non-oracle methods on five (Table 1 of the revision). Its gain over Jellyfish is 5.4 points ([3.3, 7.9]). The combinations address complementary errors. External methods repair feature values, while CARVEPrep removes records with rejected labels. Eleven of the 15 adopted combinations use CARVEPrep's deletions. Under the matched adoption comparison in Table T1, CARVEPrep's candidates give higher accuracy than the baseline tables on five of six tables on the submitted data seeds.

The closest related methods are part of this comparison. ActiveClean and BoostClean are among the fourteen baselines in Table T1, and they use model feedback to guide one kind of operation, as does meta-learned reweighting. CARVEPrep compares complete preparations that combine value changes, deletion and weights record by record (Appendix A).

AuthorsOpenReview ↗

Response to Reviewer tLVA (2 of 2)

To the rest of W2, The gain holds across corruption seeds and downstream models, and both extensions you suggest are new experiments here.

(a) Across five data seeds that redraw the injected errors, the gain over NoFix is positive on all six tables, and on five further data seeds as well.

(b) A new transfer study on the submitted data seeds tested each method's prepared table on six other models (TabICL, XGBoost, LightGBM, CatBoost, MLP and logistic regression) and on TabPFN-v2 in context under two settings. CARVEPrep's table exceeds NoFix's in all eight settings, by 3.3 to 14.6 points (Section 4.3, Appendix Table 15), and ranks first among 23 methods in five of them. We repeated the study on the five data seeds of Table 1 of the revision. On those seeds CARVEPrep's table again exceeds NoFix's in all eight settings, by 3.3 to 13.6 points, and ranks first in six of them, up from five (Appendix C.9). These studies change only the model trained on the prepared tables, so we also replaced TabPFN-v2 with TabICL inside the method, on the submitted data seeds. CARVEPrep then reaches 0.814 against 0.749 for NoFix and 0.764 for Jellyfish, first of the 23, with gains of 6.5 and 5.0 points whose 95% intervals exclude zero.

To Q3, We added studies varying the size and quality of the reference set (Appendix C.9), and controls that allow the same reference records into training (Table 11). The mean gain stays positive as reference size and quality change, and preparation still adds value at a cap of 200 records when reference records enter training.

On size, with the training pool fixed, CARVEPrep gains 3.3, 4.1, 7.2, 6.9 and 7.0 points over NoFix using 25, 50, 100, 200 and all reference records over five data seeds. Separate controls allow the reference records into final training. At a cap of 200, all 30 combinations of table and data seed are included. Appending the CARVEPrep table gives a six-table mean gain of 0.8 points over reference-only training (95% interval [0.4, 1.1]) and 0.6 points over appending the unchanged corrupted table ([0.2, 0.9]). All three options use the same reference-record training and stopping split (Appendix C.9, Table 11).

On quality, a separate five-seed study corrupts 5%, 10% and 20% of the reference records in features and labels. The gains over NoFix remain 9.6, 9.2 and 8.6 points, positive on all six tables. Their difference from the clean-reference gain is driven mainly by changed candidate choices on Flights and Hospitals (Appendix C.9).

On reuse, the default bootstrap rule reuses the reference set for search, comparison and early stopping. To isolate reuse, a five-seed run keeps the bootstrap rule and the candidates inside the budget but reserves 40% of the reference records for the adoption test alone. It reaches 0.767 against 0.769 with reuse, a change of 0.2 points (Appendix C.9).

On independent acceptance, CARVEPrep-Cert uses certified acceptance and reaches a six-table mean of 0.749 over five data seeds, 4.8 points above NoFix, and TEP balanced accuracy of 0.984 over three training runs (Table 1 of the revision). This configuration includes Jellyfish's preparation and reserves 40% of the reference set for independent acceptance (Appendix C.8). In the submission, independent acceptance is used by its CARVEPrep-Cert row of Table 1, the held-out results of Appendix C.3, the 10,000-record Adult result of Appendix C.4 and Figure 6(g,h). A separate five-seed run of the held-out rule with every candidate inside the budget and no external table reaches 0.733 against 0.701 for NoFix.

We welcome further discussion if any concerns remain.

Dear Reviewer tLVA,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

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

This work covers data preparation for post-training TabPFN-v2. It gives each record one of four actions under a record budget: retain, repair, delete, or downweight. CARVEPrep flags suspect cells from dependency violations, format and typo checks, and label rejection by a model fitted on a clean reference set. Per-action regressors, trained on errors the method injects into the working table itself, pick each record’s action. The resulting candidate preparations are pre-screened by in-context accuracy. Leading candidates are then compared with the unchanged table by repeated short post-training and adopted only if they pass a gain margin and a paired bootstrap. Losses from the trained model re-rank records for the next round. A held-out variant (CARVEPrep-Cert) replaces the bootstrap with a betting test on a separate acceptance sample. The theory covers a separable weighted least-squares model.

On six tables with errors injected at rate ρ = 0.4, CARVEPrep reaches a mean accuracy of 0.8119, against 0.7343 for the unchanged table (NoFix) and 0.7853 for Jellyfish, the strongest of fourteen baselines. CARVEPrep matches or exceeds every baseline on four of six tabular datasets.

Critical strengths

The paper judges data preparations by the accuracy of the post-trained model, not by how many cells end up correct, and reports both so the reader can see where they diverge. Without injected errors, NoFix is selected on all six tables. The held-out variant is reported even though it costs 3.17 points, and adding Jellyfish’s table as a candidate shows the adoption rule choosing an external preparation where it is better (Flights). The baseline set is broad: fourteen methods plus five uniform strategies.

The evidence analysis is a concrete, checkable result. On the 18 dependencies whose median group size exceeds one, structural evidence recovers 0.947 of the injected violations, against 0.184 for label-conditional scoring, and the gap grows with group size.

Critical weaknesses

The main comparison cannot separate CARVEPrep’s gain from noise, or from the clean reference set it uses to make choices.

  • Each table is evaluated on one split and one corruption instance, and each cell of Table 1 is the best of three training runs as judged by reference accuracy (Appendix C.3). There is no variation across splits or corruption seeds.
  • It holds up to 500 clean records, which is 500 clean records against 900 corrupted training records on Adult. It is used for prior fitting, in-context screening, adoption and early stopping. No baseline uses these records directly, for example by training or conditioning TabPFN on the reference set, or appending it to the corrupted table.

The paper’s central claim is that preparations must be judged by model feedback. So the main text needs to show the gain holds across splits and corruptions, and beats simply using the clean records.

The four-action framing is supported by theory and by test-set comparisons after the fact, not by the preparations the method actually adopts.

  • At ρ = 0.4, almost every training record carries an error and a third to over half of the cells are corrupted. Every adopted preparation in the main comparison deletes records. With feature errors alone, NoFix is close to clean training and three of six tables are left unchanged. The gains therefore come mostly from removing label-noisy records.
  • The claim that they help rests on a separable least-squares model and on test-best candidates (Figure 6(b)). As section 4.2 says, whether that capacity is realised “depends on the adoption step”.
  • Finding 1: the motivating example is a single-run counterexample chosen from 9,681 policies and scored on 180 test records. The ground-truth repair it is compared against fixes uniformly random erroneous cells, while the mixed policy deletes records by label rarity.
  • Finding 2: the “positive interaction” on Beers is measured where every method lies between 0.48 and 0.58, and where ties among “dozens of runs” are broken by a fixed order.
Questions for authors

How large is the gain across splits and corruption seeds? Can you report means and paired intervals over several splits and corruption instances, using the three-run mean rather than the run selected on the reference set? Which of the six table-level gains survive?

What do the 500 clean reference records achieve on their own? Please compare against TabPFN trained or conditioned on the reference set alone, on the reference set appended to the corrupted table, and against the best of all fourteen baseline preparations chosen by the same adoption rule. Evidence that CARVEPrep still leads would materially strengthen the claim.

Please clarify the TEP results and the motivating examples. In TEP, most methods score exactly 0.90 on most scenarios while clean training reaches 1.00. What is the class balance, and do the rankings hold under balanced accuracy? How was Mixed #8449 chosen from the 9,681 policies, and why do its accuracy and NoFix’s differ between Figure 1 and Appendix C.6?

6 comments
AuthorsOpenReview ↗

Response to Reviewer dHRf (1 of 3)

Thank you for the close reading. Your requests for more splits and for controls on the clean records strengthened the evaluation.

To W1a and Q1, We reran CARVEPrep and all fourteen baselines over five data seeds that redraw the splits and injected errors, with three training runs per seed (Table 1 of the revision). CARVEPrep's ER setting reaches a six-table mean of 0.795. It exceeds all fourteen baselines on every table. Across the framework's settings in Table 1, a CARVEPrep variant holds the highest non-oracle value on all six tables and all six TEP scenarios, including ties, and the lowest held-out loss on Alpaca (Appendix Table 17). Its gains over NoFix and Jellyfish are 9.4 points ([6.2, 13.4]) and 5.4 points ([3.3, 7.9]), and both 95% intervals exclude zero (Table G1, general comment). On Flights and Hospitals it reaches 0.636 and 0.953, against Jellyfish's 0.634 and 0.925.

The standard setting also retains positive gains over NoFix on all six tables, with five-seed gains of 7.0 points over NoFix and 3.0 over Jellyfish, compared with 6.9 and 1.9 in the submitted three-run means. It leads Jellyfish on Adult, Beers, Breast Cancer and Tax, and Jellyfish leads on Flights and Hospitals (Table W1). On five further data seeds, the standard setting's six-table mean reaches 0.790, above all fourteen baselines, and its gains over NoFix and Jellyfish rise to 9.0 and 3.8 points. The 95% intervals of gains over NoFix, Jellyfish, RHO-LOSS, Ref. likelihood and Raha+Baran all exclude zero (Appendix Table 18). In the submitted comparison, every method selected its final run among three training runs by reference accuracy, without using the test set. Appendix C.3 and Figure 6(a) also report its three-run means. Appendix Table 10 preserves the submitted comparison.

Table W1. Five-seed means in points, relative to NoFix.

TableClean trainingJellyfishCARVEPrepCARVEPrep-ER
Flights54.5+17.9+9.1+18.1
Hospitals22.6+15.9+14.1+18.7
Breast Cancer9.7−0.7-0.7+8.9+9.0
Adult7.9−0.1-0.1+4.8+5.0
Tax4.5−3.4-3.4+4.4+4.2
Beers3.5−5.5-5.5+0.8+1.6

To W1b and Q2, In our setting the reference set simulates the small amount of verified information a practitioner has, such as trusted values, labels and expert rules (Definition B.4). CARVEPrep uses it to guide and judge preparations, through prior fitting, candidate comparison and early stopping. It is not used as training data, and repair values come from the working table itself. Even so, what the same records achieve as training data is a fair question. We ran these controls at five reference sizes over five data seeds (Appendix Table 11). With the full reference set, training on the records alone reaches 0.866, close to clean training at 0.872. At 200 records, appending the CARVEPrep table gains 0.8 points over reference-only training ([0.4, 1.1]) and 0.6 points over appending the corrupted table ([0.2, 0.9]). At 100 records they are 1.3 points ([0.1, 2.2]) and 3.6 points ([2.6, 4.7]). A small reference set lacks information that the corrupted pool carries, and preparation makes that information usable. All three options use the same training and stopping split of the reference records (Table R1, Appendix C.9), and the prepared table gives the highest mean at each reference size. Under the standard protocol, which does not train on reference records, CARVEPrep keeps its full gain with 100 of them, 7.2 points over NoFix against 7.0 with all.

Table R1 (new runs). Means over tables of the per-table means, over the tables and seeds counted in the second column. A table and seed enter a row when the NN records contain every class, and the last row uses the full reference set.

NNTable-seed pairsReference records aloneCorrupted table + referenceCARVEPrep table + reference
25220.6870.7030.709
50250.7190.7110.734
100280.8090.7860.822
200300.8390.8410.847
All300.8660.8660.867

For your third comparison, we ran the same adoption rule on all fourteen baseline tables. On the submitted data seeds it reaches 0.812, and 0.816 with CARVEPrep's candidates added, both above the 0.785 of the best single baseline without the rule (Appendix Table 3). Over five seeds the rule reaches 0.784 on the fourteen baseline tables. CARVEPrep's own candidates are higher than the rule's choice among those tables on Adult, Beers, Breast Cancer and Tax. Flights and Hospitals have the largest gap to clean training (Table W1), so value repair has the most room there. Combining record-wise deletions with external repairs therefore reaches 0.795 with the record budget lifted, above 0.784.

AuthorsOpenReview ↗

Response to Reviewer dHRf (2 of 3)

To W2a and W2b, Our thinking is that the four actions define the search space and the adopted actions follow the errors present. Based on your point, we added matched controls on adopted preparations (Appendix C.9 and Table 17). They show gains from combining external repairs with deletion and from keeping repairs and weights in the adopted Alpaca preparation. Under the same in-context adoption rule with no record cap, external-repair combinations add 1.6 points ([0.8, 2.5]). Of 30 adopted preparations, 15 combine deletion and external repair. CARVEPrep-ER reaches 0.795 and leads non-oracle methods on five tables in the revised Table 1. The Alpaca control uses the evidence rule and keeps the adopted deletion set fixed. Removing repairs and weights raises held-out loss from 1.314 to 1.337 (Appendix Table 17).

The adopted actions depend on the error setting. At ρ=0.4\rho = 0.4, every adopted preparation in the submission deletes records. Across the new five-seed runs, 27 of 30 delete and 4 of them also repair, 2 others repair without deletion, and 1 retains the table. With feature errors alone, the table is near clean-training performance and three of six remain unchanged. Section 4.2 and Appendix Table 12 distinguish adopted actions from the test-best candidates in Figure 6(b).

The theory establishes that removing repair or fractional weights can restrict attainable performance under a common budget. The new external-repair comparison quantifies the gain from combining preparations, and Appendix C.9 compares learned action values with the evidence rule.

To W2c, The example was meant to show that the same data-level correctness can give very different model accuracy, and your question led us to measure this directly. We added three-seed retraining to measure the effect of repair placement (Appendix Table 16). Across placements, accuracy differs by 28, 24 and 26 points in the three runs even with the number of ground-truth repairs fixed. The submission presents Figure 1(b) as a selected counterexample (Appendix C.6). The comparison covers randomly chosen subsets of the erroneous cells of Flights, each receiving 359 ground-truth repairs. In the run behind Figure 1, their accuracies range from 43.9% to 75.0%. We have rewritten Finding 1 in the revision to compare placements of the same number of ground-truth repairs. On the choice among the 9,681 policies, only 192 are trained and the rest are interpolated, as Example C.1 of the revision states, so Mixed #8449 comes from 192 measured policies.

To W2d, We reanalyzed the submitted runs under alternative treatments of ties (Appendix C.6). The Beers interaction stays positive in every check. The submission breaks ties by a fixed order that uses no test data (Appendix C.3), and the "dozens of runs" it mentions occur on Adult and Breast Cancer. The interaction is +3.17 points under the fixed order, as submitted, and +3.08 when the tied runs are averaged. The tied runs on Beers have equal or nearly equal test accuracy. When the training rule is selected within each training seed, the interaction is +2.33, +2.17 and +1.00. We have restated Finding 2 in the revision as non-additivity. Beyond Beers, the submission already reports that tuning both is highest or tied highest on 11 of 12 tasks (Section 4.2).

AuthorsOpenReview ↗

Response to Reviewer dHRf (3 of 3)

To Q3, The score of 0.90 is the majority-class rate, because each TEP test set has 261 faulty and 29 normal records. A predictor that outputs fault for every record scores exactly 261290=0.9000\tfrac{261}{290} = 0.9000 in accuracy and 0.50.5 in balanced accuracy. To test whether the rankings hold under balanced accuracy, we added retraining experiments with this metric (Table 1 of the revision). These runs retrain the prepared tables of the 21 non-oracle methods with three seeds, 378 runs in total, of which 171 score exactly that accuracy and 168 of those predict fault everywhere.

CARVEPrep has the highest six-scenario mean balanced accuracy, and its lead over the strongest baseline grows from 3.4 points in accuracy to 14.1 points in balanced accuracy (Table R2). The other settings of the same framework also exceed the strongest baseline, with 0.986 for CARVEPrep within bb and 0.984 for both CARVEPrep-ER and CARVEPrep-Cert (Table 1 of the revision, configuration details in Appendix C.8). CARVEPrep-ER ranks first among non-oracle methods on F05 at 0.957. We also expanded the comparison to all 21 TEP fault scenarios. CARVEPrep averages 0.829 against 0.694 for Ref. likelihood, the strongest of the other methods run there (Appendix C.5).

Table R2 (new runs). Balanced accuracy 12(TPTP+FN+TNTN+FP)\tfrac{1}{2}\big(\tfrac{TP}{TP+FN} + \tfrac{TN}{TN+FP}\big) on TEP with faulty records as positives (three-run means). Differences in the text are computed before rounding.

F01F04F05F07F12F14Mean
NoFix0.8490.5990.5000.5000.6070.5000.592
RHO-LOSS0.9761.0000.9550.9940.5800.5000.834
Ref. likelihood0.9810.9940.9450.9710.6840.4970.845
Jellyfish0.8650.8150.5000.5000.5000.5000.613
CARVEPrep0.9811.0000.9431.0000.9971.0000.987

On how Mixed #8449 was chosen, among the 192 trained policies it has the unique highest checkpoint accuracy, so selecting on the 84 checkpoint records of the reference set alone yields this policy. On test accuracy, the same policy ties with #91 for the highest value. We display this policy because it reaches high accuracy with few repairs, which illustrates our motivation and problem definition. It makes 27 validated edits, deletes 44 records and downweights the remaining suspect records, without any true value, and reaches 56.67% in that run, against 48.89% for the best uniform strategy and 41.11% for NoFix (Example C.1). It is also reachable by search, since model feedback on the reference records alone selects it. The method of Section 3 is designed to find such preparations. The accuracies of this policy and of NoFix differ between Figure 1 and Examples C.1 and C.2 of the submission only through the encoder, which Appendix C.6 of the submission describes. In Figure 1 the encoder appends categories from the clean training table, which gives Mixed #8449 57.22% and NoFix 44.44%, against 56.67% and 41.11% in the examples.

We welcome further discussion if any concerns remain.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response

We thank the three reviewers for the careful reviews. They find the formulation of data preparation as model-aware action allocation compelling and well motivated (DEGn-S1, tLVA-S1) and the evaluation extensive and convincing (DEGn-S2, tLVA-S2, dHRf-S1). They also credit the evidence analysis as concrete and checkable (dHRf-S2), and the held-out tests and the adoption of external preparations where they are better (dHRf-S1, DEGn-S2).

The reviews raise three shared questions. The first concerns the stability of the gains and the evaluation metrics (dHRf-W1a, Q1, Q3, tLVA-Q3). The second concerns the contribution of reference records and action allocation (dHRf-W1b, W2, Q2, tLVA-W2, Q2, Q3). The third concerns the method's scope and its record and compute budgets (tLVA-W1, W2, Q1, DEGn-W1, W2, Q1 to Q3). We summarize the results below, followed by the replies to each reviewer.

(1) The gain holds as the splits and the injected errors change. (a) We reran CARVEPrep and all fourteen baselines on six tabular datasets over five data seeds that redraw the split and injected errors. With external repairs and no record cap, CARVEPrep-ER gains 9.4 points over NoFix (95% CI [6.2, 13.4]) and 5.4 over Jellyfish (95% CI [3.3, 7.9]) (Table G1). The standard setting gains 7.0 points over NoFix, positive on all six tabular datasets, and 3.0 over Jellyfish. Its lead over Jellyfish grows from 1.9 points in the submitted three-run means to 3.0, and to 3.8 on five further data seeds, where its gain over NoFix reaches 9.0.

(b) On TEP, CARVEPrep reaches 0.987 in balanced accuracy against 0.845 for the strongest baseline. Constraining every candidate to the record budget retains a mean of 0.986, 14.1 points above that baseline. Across all 21 fault scenarios, CARVEPrep reaches 0.829 against 0.694.

Table G1 (new runs). Five-seed test accuracy, three training runs per seed. CARVEPrep-ER uses external repairs, in-context adoption and no record cap. The last two columns give its paired gains in points. The Mean row gives 95% bootstrap intervals over data seeds.

TableNoFixJellyfishCARVEPrepCARVEPrep, in-context adoptionCARVEPrep-ERER gain over NoFixER gain over Jellyfish
Adult0.7600.7590.8080.8140.810+5.0+5.1
Beers0.5370.4820.5450.5510.553+1.6+7.1
Breast Cancer0.8690.8620.9580.9560.959+9.0+9.7
Flights0.4550.6340.5460.5980.636+18.1+0.2
Hospitals0.7660.9250.9070.8940.953+18.7+2.8
Tax0.8160.7820.8610.8590.858+4.2+7.6
Mean0.7010.7410.7710.7790.795+9.4 [6.2, 13.4]+5.4 [3.3, 7.9]

(2) CARVEPrep gains from comparing and combining complementary preparation methods. (a) Given the prepared tables of all fourteen baselines, CARVEPrep's adoption rule raises the best single baseline from 0.741 to 0.784 over five seeds.

(b) In the new experiments addressing your requests, the reported CARVEPrep variants collectively rank first among non-oracle methods on all six tabular datasets and six TEP scenarios (Table 1, including ties), and obtain the lowest held-out loss on Alpaca (Appendix Table 17, upper block). CARVEPrep-ER reaches 0.795 with external repairs and no record cap. On Flights and Hospitals it reaches 0.636 and 0.953, above Jellyfish's 0.634 and 0.925. These candidates combine record-wise deletions and external repairs under in-context adoption.

(c) Preparing the corrupted records adds value on top of the reference records. The reference set simulates the small amount of verified information available in practice and is not used as training data. When 100 such records are added to training, appending them to the CARVEPrep table reaches 0.822, against 0.809 for the records alone and 0.786 for appending them to the corrupted table (Table R1, reply to Reviewer dHRf).

(d) With independent certified acceptance, CARVEPrep still reaches a five-seed six-table mean of 0.749, 4.8 points above NoFix, and a TEP mean balanced accuracy of 0.984 (Table 1).

(3) The results extend beyond the error types and the model of the submission. (a) New runs on the submitted data seeds replace the injected training-table errors by other feature errors, by pair-flip label noise, or by both. CARVEPrep ranks first among all non-oracle methods run in each setting (Table D1, reply to Reviewer DEGn).

(b) With TabICL in place of TabPFN-v2 inside the method, on the same seeds, CARVEPrep reaches 0.814 against 0.749 for NoFix and 0.764 for Jellyfish.

(c) Comparing candidates in context requires no short post-training and reaches 0.779, 3.8 points above Jellyfish at about a quarter of its cost.

The revised PDF and supplementary material have been uploaded. Changes are in red, with comment labels in green for dHRf, orange for tLVA and blue for DEGn. Anonymous code: https://anonymous.4open.science/r/CARVEPrep-ICLR2027/. Section and table numbers refer to the revision.

AuthorsOpenReview ↗

Summary of changes in the rebuttal revision PDF

We thank all three reviewers again for their time and comments. We have uploaded the revised PDF and updated our replies to each of your questions with the new experiments and analyses. The substantive changes to the paper are in red, and an underlined label at the end of each changed paragraph names the review comment it answers. The labels are colored by reviewer, green for dHRf, orange for tLVA and blue for DEGn. The main changes are listed below.

  • Table 1, five-seed means and TEP balanced accuracy, now also reporting CARVEPrep under the record budget, with external repairs and with certified acceptance (dHRf-W1a, dHRf-Q3, tLVA-W1, tLVA-W2, tLVA-Q3)
  • Appendix Table 18, five further seeds with 95% intervals (dHRf-W1a, tLVA-W2)
  • New Section 4.3 on reference records in training, further error types, cost and transfer (dHRf-W1b, dHRf-Q2, DEGn-W1, DEGn-Q1, DEGn-W2, DEGn-Q2, tLVA-W2)
  • Appendix C.5, all 21 TEP fault scenarios (dHRf-Q3)
  • Appendix C.9 and its Tables 11 to 17, with the added analyses for all three reviews (dHRf-W1b, dHRf-W2a, dHRf-W2b, dHRf-W2c, dHRf-Q2, dHRf-Q3, tLVA-W2, tLVA-Q1, tLVA-Q2, tLVA-Q3, DEGn-Q1, DEGn-Q2, DEGn-Q3)
  • Appendix Table 10, the submitted Table 1 (dHRf-W1a)
  • Figure 5, rerun with three training runs per cell (dHRf-W1a)

The supplementary material contains the added code and the per-run records of these experiments. We hope the new experiments and analyses address your concerns, and we would be glad to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.