acceptodds
Under review as a conference paper at ICLR 2027

INCONVENIENT: A BENCHMARK TO ASSESS LLM GENERALIZATION CAPABILITIES FROM CONVENIENCE SAMPLES

Abstract

Conventional machine learning models for tabular prediction typically assume that training data are representative of the population on which they will be deployed. In practice, however, data are often collected based on convenience, availability, accessibility, or cost, resulting in systematic sampling biases that can distort feature–target relationships and undermine real-world generalization. Existing tabular distribution-shift benchmarks primarily construct out-of-distribution (OOD) settings by partitioning observations across naturally occurring domains. While such partitions may induce changes in P(Y | X), the resulting shifts are neither explicitly controlled nor isolated, and the evaluation distributions themselves are often not representative of the broader target population. We introduce INCONVENIENT, a benchmark designed to study robustness to convenience-sample bias. Rather than relying on naturally occurring domain partitions, INCONVENIENT directly perturbs feature–target relationships in the training distribution while preserving the original population distribution as the deployment environment. This setting more closely reflects real-world scenarios in which models are developed from unrepresentative samples and subsequently deployed on a broader population. The distinction is particularly relevant for large language models (LLMs), whose pretraining may encode population-level knowledge that extends beyond the relationships observed in a biased training dataset. Unlike existing distribution-shift benchmarks, INCONVENIENT provides explicit control over both the direction and severity of relationship shifts, enabling a systematic assessment of robustness to convenience sampling. The benchmark comprises 200 tabular classification and regression datasets spanning diverse application domains and generates around shifted training distributions which lead to substantial downstream performance degradation.

Primary area:
datasets and benchmarks
Keywords:
NLP in healthcare, AI in healthcare, machine learning, natural language processing, tabular prediction, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nj6ROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes INCONVENIENT, a benchmark for tabular prediction when the training data come from a biased sample but the test data come from the full population. Starting from 200 public datasets, the authors select training subsets in which the relationship between one feature and the target is made stronger, weaker or reversed, and they provide knobs to control how severe the shift is. Standard tabular models, including gradient boosted trees and TabPFN, lose a large part of their performance under these shifts. A small experiment also suggests that an LLM can partly tell which feature relationships changed the most.

Critical strengths

S1: The setting is worth studying. The deployment population stays fixed and only the training sample is unrepresentative, which differs from the domain split design of WILDS, TableShift and WhyShift. Changing P(Y|X) with a controlled direction and strength is a good idea for a benchmark.

S2: The benchmark is large, with 200 datasets and about 15,500 shifts, and it comes with a command line tool that exposes the severity knobs. The severity analysis is reported honestly, including the finding that the magnitude knob does not reliably control shift difficulty in Section 5.3

Critical weaknesses

W1: The main purpose of the benchmark is not evaluated. The title and abstract present INCONVENIENT as a benchmark for LLM generalization, and Section 5.1 argues that "correction must therefore come from outside the data". However, no LLM based predictor or prior elicitation method is evaluated, although TabLLM, LIFT, AutoElicit, LLM Processes and LoID are all cited in Section 2. The only LLM result is a Recall@5 probe with one model, compared only with random ranking. A simple heuristic that ranks features by the size of their shifted coefficients is needed before the gain can be attributed to domain knowledge. Many of the datasets, such as Titanic, Boston, California Housing, Pima diabetes and COMPAS, may also have been seen during LLM pretraining, which is a contamination problem for exactly the methods the benchmark targets. Without at least one method that closes part of the gap, it is not possible to tell whether the benchmark is solvable or whether it can separate methods.

W2: The construction and the filtering make the main results hard to interpret. (a) The influence score in Eq. (5) is built from each example’s residual, so the selection depends directly on the label, and Appendix A.1 itself calls it residual-conditioned selection. This is closer to targeted data poisoning than to the examples in the introduction, such as one hospital or one platform, where selection mostly depends on covariates or site. (b) αmin⁡\alpha_{\min} can be as small as 0.1, so a convenience set can be a small part of the pool, while the IID model is trained on the whole pool. The reported drop therefore mixes the effect of fewer samples with the effect of a distorted relationship, and a random subset of the same size is needed as a control. (c) Shifts with a performance drop below 0.1 are removed (Section 4.1, step 6), so an average drop of at least 0.1 is guaranteed by design. Table 1 in the appendix also shows different shift counts for different methods under the same configuration. For binary tasks with attenuation, bin 0 and αmin⁡\alpha_{\min} of 0.1, there are 80 shifts for CatBoost, 103 for LightGBM and 56 for logistic regression. This suggests the filter is applied per method, so each method is averaged over a different set of shifts, and statements such as Ridge and MLP "suffer among the largest degradations" are not like for like comparisons.

W3: There are several inconsistencies, and the presentation makes the benchmark hard to assess. (a) Figure 1 shows normalized classification drops of about 0.02 to 0.06. This cannot happen if every retained shift has an absolute drop of at least 0.1, unless Figure 1 uses unfiltered shifts, which is not stated. (b) The conclusion says the baselines "lose between 20% and 45% of their performance", while Section 5.1 reports 20% to 29% for classification and 53% to 82% for regression. (c) In Appendix A.2 the stated counts do not match the lists. Regression from TALENT says 7 but lists 6, binary from OpenML says 23 but lists 25, binary from Kaggle says 11 but lists 9, and multiclass from OpenML says 78 but lists 81. At least 13 datasets appear in both the regression and the multiclass lists, for example Abalone, Airfoil Self Noise, Concrete Compressive Strength, Kin8nm and Wine Quality, and several entries appear to be copies of the same source, such as the Pima diabetes, wine quality and medical insurance files. The real diversity is therefore below 200 datasets. (d) Most results are averaged plots and a configuration table in Appendix A.3 that runs for more than ten pages, with little analysis per dataset or per domain. In Figure 2 most of each panel is empty, and in Figure 1 the methods are separated only by color with the same marker. There is also no figure that shows how a shift is constructed.

Questions for authors

Q1 (see W1): Please evaluate at least two methods that use LLM knowledge, for example TabLLM or LIFT style prediction, or AutoElicit or LoID style priors used to regularize a linear model, together with a strong data only baseline, and report whether any of them reduces the drop. For Section 5.4, please add a baseline that ranks features by coefficient size, and a contamination check, such as asking the LLM to reproduce the reference coefficients of well known datasets from feature names only.

Q2 (see W2): What is the ratio between the convenience set size and the pool size across the retained shifts? Could the authors add a random subset of the same size as a control and report the drop relative to it? Is the 0.1 filter applied per method or per shift? Please report results without the filter and on the set of shifts shared by all methods.

Q3 (see W2): Can the selection procedure be checked against at least one real case where the sampling mechanism is known, for example a site or covariate based selection, or the head and neck cancer example cited in the paper? A covariate based selection mode could also be offered next to the residual based one.

Q4 (see W3 ): Please state which filter is used in Figure 1, reconcile the 20% to 45% in the conclusion with Section 5.1, and explain the dataset counts in Appendix A.2 and how duplicate sources were handled. I also suggest adding a figure that shows how a shift is constructed, moving some analysis per dataset or per domain into the main text, and using distinct markers or line styles with a colorblind friendly palette in Figure 1, so that the methods can be distinguished without relying on color.

5 comments

Thank you for your insightful review! Please find our responses below:

W1:

We originally excluded LLM-based methods from the benchmark because none of the ones we tested (TabLLM, LoID, AutoElicit) beat the best ML model (on avg 0.018 worse). Each method has a structural handicap. TabLLM sees only a few labeled examples of the train set in its context. AutoElicit and LoID only produce priors for linear models, so comparing them to tree-based models, which can fit more complex functions, is unfair. Their underperformance therefore might not mean that the LLM lacks the relevant knowledge.

Comparing AutoElicit to Bayesian regression with uninformative priors is fair since the methods only differ in the priors used. This isolates the LLM's contribution to downstream performance, and the results confirm it:

Binary tasks: AutoElicit increased AUC 0.682 to 0.811. The full-data fit AUC is 0.821. Regression tasks: AutoElicit raised mean R² from 0.482 to 0.502. The full-data fit R² is 0.526. Coefficients: the elicited priors cut the mean squared error of the coefficients, measured against the full-data coefficients, by 70%. We added the justification for this exclusion and the above analysis to the revised manuscript.


W2-(a):

  1. Selecting on covariates alone can’t create this shift. If selection depends only on X, P(Y|X) doesn’t change, and a well-specified model still learns the right relationship. Label dependence is therefore the mechanism, not an artifact.

  2. Real convenience samples distort relationships through the same mechanism: site effects driven by severity or referral patterns, outcome-dependent recruitment, and collider bias all select on the outcome given the features. The head and neck cancer recurrence cohort is a real example. In this single-institution cohort of 508 patients, a standard model learns that marijuana protects against recurrence, contrary to the clinical literature (Rezaeimanesh et al., 2026). Similarly, early COVID-19 studies of hospitalized patients suggested a spurious protective effect of smoking due to collider bias (Griffith et al., 2020).

  3. INCONVENIENT is a controlled stress test. As in ImageNet-C, synthetic shifts give you ground truth: which relationship changed, in which direction and by how much. Natural-shift benchmarks can’t provide that, and both the severity analysis and the probe depend on it.

  4. The worst-case tilt. We agree that greedy selection leans toward worst-case shifts, but the severity knobs moderate this. Unlike poisoning, no rows are changed or injected. Every example is a real population row, and only the sampling is biased.

W2-(b):

We have added the requested size-matched control to the Appendix. We drew three random subsets of the IID pool, each the same size as the convenience set, and trained all six baselines on them using the same tuning protocol as in the main experiments. The results show that the distortion contributes more to the performance drop:

Total dropFewer samplesDistortion
Classification (AUC)0.2240.097 (43%)0.127 (57%)
Regression (R²)0.3460.152 (44%)0.194 (56%)

We also note that small size is itself a defining property of real convenience samples. The total drop therefore reflects the penalty a practitioner would actually face. We now report this in the manuscript.

W2-(c):

>0.1 degradation is a feature of the benchmark rather than an empirical finding: INCONVENIENT contains only shifts with a large OOD–IID gap that standard models cannot close, which leaves room for methods that use information from outside the data. We agree that the wording can be misleading and have revised the introduction accordingly.

Per your suggestion, we also limited our baseline comparison analysis to shifts that are shared by all the baselines. Now we can correctly confirm that ridge and MLP suffer the largest degradations.


W3-(a):

Figure 1 includes all the possible shifts (without the filtering) to show how the knobs effect performance. We added this to the table description.

W3-(b,c):

We apologize for the inconsistencies. Thank you for pointing them out. Section 5.1 reports the correct numbers and the conclusion is now fixed. We also fixed the inconsistencies in the data counts (The dataset count remains 200).

Datasets that appear in more than one category have multiple types of targets. For example, the household income has both a regression type target for income and a classification type target for when whether a household is high income. It is true that this means the 200 datasets are not completely distinct. We will flag this in the manuscript.

W3-(d):

We made all the figures color-blind friendly. We will also add a figure that explains how the shifts are made.

Thank you for the detailed response. I accept the argument on label-dependent selection, and I have raised my soundness score accordingly

For W1, I would like the authors to clarify one point. In the new AutoElicit experiment, the flat-prior Bayesian baseline seems to perform better on the convenience samples than all baselines in Figure 3. I also suggest adding TabLLM, LoID and AutoElicit to the main results

Some points in my review are not addressed in the response,see w3c. The description of the filter is also inconsistent. The response to Reviewer QPho says the 0.1 filter is computed per shift from the best baseline. However, the number of shifts in Table 1 seems to differ across methods under the same configuration

The PDF on OpenReview has not been updated yet, so I cannot verify the new experiments and revisions in the response. For these reasons, I keep my overall rating.

Thank you for your prompt response!

The updated manuscript is uploaded and we applied all of the requested changes. We would truly appreciate it if you could review it.

  1. The experiment in conducted on a randomly selected subset of 1{,}570 datasets. We removed this detail for the sake of brevity, but it is mentioned in the updated version of the manuscript.

  2. We thank the reviewer for checking the dataset list carefully. Some datasets with similar names are distinct (for example, they differ in rows, region or time period), but others are copies of the same source obtained from different repositories. We have deduplicated the datasets within each task type by comparing their contents, corrected the counts in Appendix A.2, and rerun all results on the deduplicated set. [The main findings are unchanged: e.g., average degradation changes from X to Y.] Datasets that appear in both the regression and multiclass lists are retained, since each is used with a different target formulation. The benchmark now includes 165 datasets and roughly 13,000 shifts.

  3. We exclude baselines that degrade less than 0.1 compared to their own in-distribution performance from the table. Therefore, the shifts that are kept in the dataset have best_iid_performance - best_ood_performance > 0.1. Some methods fail to fit the in-distribution data as well as others. Therefore, their own iid - ood performance falls below 0.1. Table 1 excludes these methods. We acknowledge that this is an important distinction and have mentioned it in the description of Table 1 (Table 3 in the revised manuscript).

Reviewer_3kKnOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a methodology to adjust benchmark for realistic / real-world application. The authors mainly apply perturbation of features / target relationships. The paper has merits in the approach, but not sure the perturbation applied could be representative of the full data set, or some OOD version of it, as this is usually unknown. This is a good attempt to address a very complex problem, but unfortunately with limited impact given the unknown target variable range, scope.

Critical strengths
  • controlled shift is clear and precise
  • large scale experiments
Critical weaknesses
  • No LLM evaluation even though LLM is claimed in the title, paper
  • Not sure the shift is representative of anything natural or at least representative of real-world applications
  • Bias remains artificial
Questions for authors
  • No LLM evaluation even though LLM is claimed in the title, paper: why?
  • Not sure the shift is representative of anything natural or at least representative of real-world applications: why?
  • Bias remains artificial: any alternative?
3 comments

Thank you for your insightful review! Please find our responses below:

No LLM evaluation even though LLM is claimed in the title, paper

We originally excluded LLM-based methods from the benchmark because none of the ones we tested (TabLLM, LoID, AutoElicit) beat the best ML model (on avg 0.018 worse). Each method has a structural handicap. TabLLM sees only a few labeled examples of the train set in its context. AutoElicit and LoID only produce priors for linear models, so comparing them to tree-based models, which can fit more complex functions, is unfair. Their underperformance therefore might not mean that the LLM lacks the relevant knowledge.

Comparing AutoElicit to Bayesian regression with uninformative priors is fair since the methods only differ in the priors used. This isolates the LLM's contribution to downstream performance, and the results (on a randomly selected subset of 1{,}570 shifts) confirm it:

  • Binary tasks: AutoElicit increased AUC 0.682 to 0.811. The full-data fit AUC is 0.821.
  • Regression tasks: AutoElicit raised mean R² from 0.482 to 0.502. The full-data fit R² is 0.526.
  • Coefficients: the elicited priors cut the mean squared error of the coefficients, measured against the full-data coefficients, by 70%.

We added the justification for this exclusion and the above analysis to the revised manuscript.

Not sure the shift is representative of anything natural or at least representative of real-world applications

Thank you for raising this point. Real convenience samples distort relationships through the same mechanism: site effects driven by severity or referral patterns, outcome-dependent recruitment, and collider bias all select on the outcome given the features, and collider bias is the textbook way to attenuate or reverse an association.

The head and neck cancer recurrence cohort cited in the paper is a real example. In this single-institution cohort of 508 patients, marijuana users make up 26.5% of the non-recurrence group but only 12.4% of the recurrence group, so a standard model learns that marijuana protects against recurrence, contrary to the clinical literature (Rezaeimanesh et al., 2026). Similarly, early COVID-19 studies of hospitalized patients suggested a spurious protective effect of smoking due to collider bias (Griffith et al., 2020). This sign reversal reflects who ended up in the sample, and it is the type of shift INCONVENIENT generates.

INCONVENIENT is also intentionally a controlled stress test. As with ImageNet-C, synthetic shifts provide ground truth about which relationship changed, in which direction and by how much. Natural-shift benchmarks cannot provide this. We added this explanation about how INCONVENIENT can represent real-world convenience samples to the introduction.

Dear Reviewer,

Thank you again for your thoughtful review. We wanted to check whether our response has addressed your concerns. In brief, we added LLM-method experiment, and justified the real-world representativeness of our benchmark. The details are in our response above and the changes are reflected in the revised PDF.

We would be glad to answer any further questions or clarify anything that is still unclear while the discussion period is open.

Dear Reviewer,

As the discussion period is coming to a close, we wanted to kindly follow up on our response. We would greatly appreciate it if you could let us know whether it addresses your concerns, and we would be happy to clarify anything further before the deadline.

Reviewer_QPhoOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • This paper introduces INCONVENIENT, a benchmark for robustness to convenience-sample bias in tabular prediction.
  • It constructs biased training sets from 200 datasets using greedy subset selection, while maintaining the test set drawn from the original population as the deployment environment.
  • It evaluates seven existing tabular baselines on 15.5k shifted sets with perturbed feature-target relationships and reports substantial performance degradation.
Critical strengths
  • The approach of keeping the test distribution identical and shifting only the training set is interesting. It isolates the cause of degradation to the training sample rather than to differences in the test population.
  • The benchmark is large-scale, with 15.5k shifts generated from 200 datasets and seven baselines evaluated, including TabPFN.
Critical weaknesses
  • The title and the motivation focus on assessing LLM generatlization, yet no LLM is evaluated as a predictor on the benchmark. The only LLM experiment uses a single model (claude haiku 4.5) to detect coefficient anomalies. Thus, the question raised in this paper, assessing LLM generalization capabilities, remains open.
  • The convenience sets are much smaller than the training pool, so it is unclear whether the degradation comes from reduced sample size or from distorted feature-target relationships. Experiments on an unbiased same size subset, e.g., a random subset, are needed to clarify this.
  • The claim that “every baseline degrades” is overstated, because shifts with small performance differences are filtered out by construction (delta perf < 0.1).
Questions for authors
  • Popular datasets such as Titanic and California Housing are likely to have been seen during LLM pretraining. Since the benchmark targets LLM-based methods, how is data contamination addressed?
  • Could the authors report standard deviations or confidence intervals for the results in Table 1?
  • Why are recent LLM-based methods not evaluated, considering the title and the motivation?
  • What is the criterion for the delta perf < 0.1 filtering? Is it computed per method or for a specific reference model?
4 comments

Thank you for your insightful review! Please find our responses below:

The title and the motivation focus on assessing LLM generalization, yet no LLM is evaluated as a predictor on the benchmark.

We originally excluded LLM-based methods from the benchmark because none of the ones we tested (TabLLM, LoID, AutoElicit) beat the best ML model (on avg 0.018 worse). Each method has a structural handicap. TabLLM sees only a few labeled examples of the train set in its context. AutoElicit and LoID only produce priors for linear models, so comparing them to tree-based models, which can fit more complex functions, is unfair. Their underperformance therefore might not mean that the LLM lacks the relevant knowledge.

Comparing AutoElicit to Bayesian regression with uninformative priors is fair since the methods only differ in the priors used. This isolates the LLM's contribution to downstream performance, and the results (on a randomly selected subset of 1{,}570 shifts) confirm it:

  • Binary tasks: AutoElicit increased AUC 0.682 to 0.811. The full-data fit AUC is 0.821.
  • Regression tasks: AutoElicit raised mean R² from 0.482 to 0.502. The full-data fit R² is 0.526.
  • Coefficients: the elicited priors cut the mean squared error of the coefficients, measured against the full-data coefficients, by 70%.

We added the justification for this exclusion and the above analysis to the revised manuscript.

The convenience sets are much smaller than the training pool, so it is unclear whether the degradation comes from reduced sample size or from distorted feature-target relationships.

Thank you for your suggestion. We added the requested size-matched control to the Appendix. We drew three random subsets of the IID pool, each the same size as the convenience set, and trained all six baselines on them using the same tuning protocol as in the main experiments. The results show that the distortion contributes more to the performance drop:

Total dropFewer samplesDistortion
Classification (AUC)0.2240.097 (43%)0.127 (57%)
Regression (R²)0.3460.152 (44%)0.194 (56%)

We also note that small size is itself a defining property of real convenience samples. The total drop therefore reflects the penalty a practitioner would actually face. We now report this in the manuscript.

The claim that “every baseline degrades” is overstated, because shifts with small performance differences are filtered out by construction (delta perf < 0.1).

"Every baseline degrades" is a feature of the benchmark rather than an empirical finding: INCONVENIENT intentionally keeps only shifts with a large OOD–IID gap that standard models cannot close, which leaves room for methods that use information from outside the data. We agree that the wording can be misleading and have revised the introduction to explain this more clearly.

Popular datasets such as Titanic and California Housing are likely to have been seen during LLM pretraining.

Our AutoElicit prompt gives the LLM only the target definition and the feature names. We compared the elicited prior means with the reference coefficients fitted on each dataset's full training pool. If the models had memorized these datasets, priors on well-known datasets would be near-exact. They are not:

On popular datasets, only 10% (binary) to 31% (regression) of features fall within 5% of their reference value. More importantly, the improvement in prior–reference correlation over a no-knowledge baseline is similar for well-known and less well-known datasets, so the LLM's advantage does not come from dataset popularity.

Could the authors report standard deviations or confidence intervals for the results in Table 1?

Since the average is across all the datasets, and the datasets have varying performances, the confidence interval becomes quite wide and unrepresentative of baseline performance.

What is the criterion for the delta perf < 0.1 filtering?

delta perf is the performance difference between the best IID and the best OOD baseline (i.e. performance of the best IID trained baseline - performance of the best OOD trained baseline). We compute this for every shift and exclude the ones with delta perf < 0.1. This makes sure that even the best OOD baseline could not close the gap. We have made this comparison clear in the text in section 3.1.

Dear Reviewer,

Thank you again for your thoughtful review. We wanted to check whether our response has addressed your concerns. In brief, we added LLM-method experiment, a size-matched control to separate the impact of αmin\alpha_{min} and coefficient distortion, and cleared up the requested definitions. The details are in our response above and the changes are reflected in the revised PDF.

We would be glad to answer any further questions or clarify anything that is still unclear while the discussion period is open.

Dear Reviewer,

As the discussion period is coming to a close, we wanted to kindly follow up on our response. We would greatly appreciate it if you could let us know whether it addresses your concerns, and we would be happy to clarify anything further before the deadline.

Reviewer_QPhoOpenReview ↗

Response to the authors

Thank you for your response. I read it carefully and found that the experiments clearly addressed most of my questions. I am raising my rating to Weak Accept.

Comments on the paper

1 comment
Area_Chair_PHsoOpenReview ↗

Please respond to rebuttal

Dear reviewers,

as the discussion phase with the authors is coming to an end in a few hours, please answer to the rebuttal and summarize the reasons for your score decision.

Thanks a lot, the AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.