acceptodds
Under review as a conference paper at ICLR 2027

Context-aware Explainable Selective Regression

Abstract

Regression models can have low average error yet make large errors on individual samples, which can lead to suboptimal decisions. Selective regression mitigates this issue by rejecting unreliable predictions, but existing methods rely primarily on numerical data. Text associated with each sample can carry signals of unreliability that numerical data do not capture, and can thus provide evidence for rejection. We introduce *context-aware explainable selective regression*, a post-hoc framework that equips a fixed regressor with a context-aware rejector. Given a numerical input, the associated text, and the prediction from the regressor, the rejector outputs a score for threshold-based rejection. For each rejected sample, it also extracts evidence from the text and generates rationales explaining why the evidence indicates low reliability. We implement the rejector with a fine-tuned large language model. Across 16 real-world datasets, it outperforms existing methods on average, with clear gains from text on tabular data and domain-dependent effects on time-series data. While the evidence is only partially faithful to the scoring behavior of the rejector, the evidence and rationales appear plausible in case studies, which suggests their potential to support human review.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
selective prediction, selective regerssion, multimodality, large language model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_DM34OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a post-hoc selective regression method that uses textual context to identify unreliable predictions from a fixed regressor. An LLM takes numerical features, associated text, and the regressor's prediction as inputs and estimates a prediction-error score. A calibrated threshold is then used to accept or reject each prediction. For rejected predictions, a separate prompt extracts supporting textual evidence and generates explanations by underlining relevant text.

Critical strengths

I feel the problem is practically relevant, as assessing the reliability of an existing predictor without modifying it is useful in real-world applications. The main novelty lies in using textual information to estimate prediction errors and provide natural-language explanations for rejected predictions in selective regression.

Critical weaknesses
  1. I find the use of textual information for selective regression interesting, but I think the paper needs stronger evidence that it offers clear advantages over existing multimodal loss prediction methods. In particular, the experiments do not directly show whether text can help identify unreliable predictions even when it does not improve outcome prediction.

  2. The explanations are generated after scoring, without a clear connection between the selected evidence and the scoring process. Tables 5 and 6 show negative comprehensiveness on several datasets, meaning that removing the selected evidence can actually increase the predicted unreliability score. This raises concerns about whether the explanations truly support the rejection decisions. The qualitative examples are readable and reasonable, but they do not show whether the explanations reflect the model's actual decisions or help human reviewers. I feel the explainability contribution is not sufficiently supported by the current results.

Questions for authors
  1. What advantages does the proposed framework offer over a strong multimodal error predictor under the same prediction targets and tuning budgets? Can the authors directly test whether text helps identify unreliable predictions beyond its contribution to outcome prediction?

  2. How does the selected evidence compare with random sentence selection or deletion-based attribution methods? Can the authors explain the negative comprehensiveness results and provide stronger evidence that the explanations reflect the rejection scores?

2 comments
AuthorsOpenReview ↗

Response to Reviewer DM34 (1/2)

We thank the reviewer for the careful reading and for the concrete experimental requests. We address the questions and concerns below.

W1/Q1-1. Does text help beyond outcome prediction?

Yes, text that does not improve the regressor can therefore still indicate when its prediction is unreliable. For each tabular dataset, we compared the relative MSE reduction of the regressor (TabSTAR with vs. without text) with the relative NAURC reduction of the rejector (Qwen3.5-9B (SFT+GRPO) with vs. without text), both averaged over the four splits. We show the results the following table. On Airbnb, text increases the regressor's MSE by about 5% yet reduces the rejector's NAURC by nearly 40%. On Houses and Wine, text barely changes the regressor's MSE but reduces the rejector's NAURC by about 60%. This supports applying a context-aware rejector post hoc to a multimodal regressor. The time-series datasets are excluded from this analysis because text changes the regressor's MSE by less than 1% on all of them.

MSE reduction [%]NAURC reduction [%]
Airbnb-4.937.9
Beer4.519.2
Houses-0.265.6
Laptops12.564.6
Mercari20.837.7
Permits0.485.3
Wine2.857.7

W1/Q1-2. Why an LLM rather than a multimodal error predictor?

To our knowledge, no existing selective regression method uses text. SR-TabSTAR, SR-MMTSFlib, and SR-TaTS are not prior methods: we constructed them for our problem setting by replacing the loss predictor of SR-Kernel/SR-kNN with TabSTAR, MMTSFlib, and TaTS. We have revised Section 5.1 to state this explicitly.

The advantage of our LLM-based rejector over a multimodal error predictor is the explanation. In Table 1, the proposed rejector has the best average rank and performs comparably to SR-TabSTAR, the strongest multimodal error predictor we constructed. What SR-TabSTAR and the other baselines cannot do is explain a rejection. None of them produces evidence or a rationale. The problem we pose requires both the rejection and its explanation, and the LLM-based rejector is the component that makes it solvable. Comparable selective regression performance together with natural-language explanations is the advantage we claim.

For a comparison under the same budget, we also added a non-generative rejector built on an LLM of comparable size: a frozen Qwen3-Embedding-8B embedding of the text, concatenated with the numerical input and the prediction, is fed to a GBDT trained on the same target, the realized loss. We show the result in Table 1 in the revised manuscript. The proposed LLM-based rejector outperforms it on 6 of the 7 datasets and on average. The LLM's text representation alone therefore does not account for the performance. Rather, the generative formulation and the fine-tuning contribute. This rejector provides no explanation either.

AuthorsOpenReview ↗

Response to Reviewer DM34 (2/2)

W2/Q2-1. Why is extraction separate from scoring?

SFT supervises only the loss estimate, because no supervision is available for explanations. The extracted evidence is therefore a retrospective self-explanation and is not faithful by construction, which is why we evaluate faithfulness empirically.

W2/Q2-2. Do the explanations reflect the rejection scores?

At the sample level, these negative averages are not produced by the majority of rejected samples: on every dataset except Health and Social Good, fewer than half of the rejected samples have negative comprehensiveness (new Tables 15 and 16 in Appendix C.6). Our hypothesis for the negative cases is that removing the extracted sentences also removes context the rejector relies on. In time series, for example, a major event is often described repeatedly across consecutive timestamps. Extracting those sentences then removes a block of context around the event, and the rejector may estimate a larger loss, that is, lower reliability. We state this as a hypothesis, and we leave characterizing the datasets on which faithfulness fails to future work.

To test whether the extracted evidence reflects the sentences the score depends on, we identified, by leave-one-sentence-out occlusion, the sentences whose individual removal lowers the score (score-supporting sentences) and measured how many of them the extracted evidence recovers (capped recall@k), compared with the expected value for k sentences selected uniformly at random (new Table 17 in Appendix C.6).

Capped recall@kk = 3, extractedk = 3, randomk = 7, extractedk = 7, randomk = 10, extractedk = 10, random
Airbnb0.1540.1970.1740.2500.3450.397
Beer0.7390.6410.7880.9270.8080.993
Houses0.4300.2830.4790.4010.6380.726
Laptops0.0520.1410.0640.2080.0690.378
Mercari0.7410.6300.8800.9050.9350.988
Permits0.3110.4970.4980.8060.5990.999
Wine0.2490.3540.3000.5260.7190.912

On Beer, Houses, and Mercari, the extracted evidence recovers score-supporting sentences better than random selection at k = 3 (and at k = 7 on Houses), whereas on the other four datasets it does not. We therefore conclude that the evidence is faithful to the scoring on some datasets and not on others. This is consistent with prior findings that LLM self-explanations can be unfaithful and that faithfulness varies across datasets (Turpin et al., 2023; Madsen et al., 2024).

W2/Q2-3. Do the explanations help human reviewers?

We have not conducted a human study. The case studies show that the evidence and rationales are plausible, not that they help reviewers. Section 5.2.2 states this and lists a human evaluation as future work.

Reviewer_VUpLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a context-aware and explainable selective regression by building a post-hoc rejector using an LLM. The proposed approach comprises three steps: first, the rejector estimates the regression loss for each sample and uses it as a score. Second, the rejector compares this score against a prespecified threshold to accept or reject the prediction and third, it extracts evidence for a rejected prediction and explains why it suggests low prediction reliability.

Critical strengths
  1. The paper addresses selective regression in situations where bringing additional information (like unstructured data) can be practically meaningful. The house and energy example makes the motivation clear.

  2. The proposed method not only rejects a sample but also extracts evidence from the context and generates a rationale.

  3. The post-hoc nature of the proposed method makes it usable for any trained regressor.

  4. The experiments demonstrate the practical usefulness of the proposed method.

Critical weaknesses
  1. The methodological novelty seems to be incremental. The authors should demonstrate that using an LLM as the rejector has theoretical advantages compared to other methods.

  2. As the authors note, LLM-generated self-explanations can be unfaithful, and their faithfulness depends on the task.

  3. The accept/reject decision depends on a prespecified threshold on predicted loss. Therefore, the results depend strongly on how that threshold is calibrated. The paper should report the behavior of the proposed approach across different thresholds.

Questions for authors

Please address the weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer VUpL

We thank the reviewer for the review, and in particular for recognizing the practical setting, the explanation generation, and the post-hoc design. We address the concerns below.

W1. Methodological novelty and the advantage of an LLM as the rejector

We agree that the rejector's architecture is incremental. The paper's contribution is not the architecture. It consists of the problem setting the reviewer describes in S1 and S2, in which text is used both to identify unreliable predictions of a fixed regressor and to explain each rejection in natural language, together with a first method for this setting and empirical findings on when text helps. What the LLM provides beyond accepting text as input is the explanation. Existing selective regression methods such as SR-Kernel and SR-kNN use no text and produce no explanation, so they are not solutions to this problem. Neither are the multimodal rejectors we constructed for comparison, such as SR-TabSTAR, which performs comparably to the proposed rejector in selective regression but cannot explain its rejections.

The submitted version did not make this claim explicit. It introduced context-aware explainable selective regression as a "post-hoc framework", which reads as a method. We have revised the wording accordingly. The Abstract and Section 1 now formulate it as a new problem setting, and Section 1 presents the LLM-based rejector as a first method for this setting. Section 3 is retitled "Problem Setting", and Sections 4 and 6 use the same wording. Only the presentation has changed, and the method itself is the same.

We do not claim a theoretical advantage of the LLM as a loss estimator over other estimators, and the revision does not add such a claim. The practical advantage of using an LLM is that it can explain each rejection in natural language, which existing selective regression methods cannot do.

W2. Self-explanations can be unfaithful and task-dependent

We agree, and the paper says so (Section 4 and Section 5.2.2). The revised version adds a more detailed analysis of this point (Appendix C.6, summarized in Section 5.2.2), which gives the two results below.

First, at the sample level, the negative average comprehensiveness on some datasets is not produced by the majority of samples: on every dataset except Health and Social Good, fewer than half of the rejected samples have negative comprehensiveness (new Tables 15 and 16).

Second, we tested whether the extracted evidence reflects the sentences the score depends on. By leave-one-sentence-out occlusion we identified the sentences whose removal lowers the score, and measured how many of them the evidence recovers (capped recall@k) against the expected value for k randomly selected sentences (new Table 17).

Capped recall@kk = 3, extractedk = 3, randomk = 7, extractedk = 7, randomk = 10, extractedk = 10, random
Airbnb0.1540.1970.1740.2500.3450.397
Beer0.7390.6410.7880.9270.8080.993
Houses0.4300.2830.4790.4010.6380.726
Laptops0.0520.1410.0640.2080.0690.378
Mercari0.7410.6300.8800.9050.9350.988
Permits0.3110.4970.4980.8060.5990.999
Wine0.2490.3540.3000.5260.7190.912

On Beer, Houses, and Mercari, the extracted evidence recovers score-supporting sentences better than random selection at k = 3 (and at k = 7 on Houses), whereas on the other four datasets it does not. We therefore conclude that the evidence is faithful to the scoring on some datasets and not on others. This is consistent with prior findings that LLM self-explanations can be unfaithful and that faithfulness varies across datasets (Turpin et al., 2023; Madsen et al., 2024).

W3. Behavior across different thresholds

We added normalized risk–coverage curves, in which the selective risk is divided by the full-coverage risk, for all datasets and methods (new Figs. 7 and 8 in Appendix C.3). On most datasets, the curve of the proposed rejector lies below those of the other methods over most of the coverage range. Moreover, it stays at or below 1 over the entire range on all 7 tabular datasets and on 4 of the 9 time-series datasets, which indicates that the rejector stably rejects unreliable samples at every coverage level.

Reviewer_KtyrOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies context-aware explainable selective regression. The goal is to identify unreliable predictions of a given regression model, and to explain why. For a sample (X,S,Y)(X,S,Y) and a given regressor ff, the authors build an LLM-based rejector gg such that g(X,S,f(X,S))g(X,S,f(X,S)) indicates whether to accept or reject the prediction f(X,S)f(X,S). The rejector uses an LLM MθM_\theta to estimate the prediction loss and builds a threshold upon it. This paper also implements fine-tuning of the LLM MθM_\theta using SFT and GRPO in their proposed framework. For rejected predictions, the authors use a separate prompt to explain the reason for rejection by using evidence extracted from the given text context SS. In experiments on 7 tabular regression datasets and 9 time-series datasets, this paper demonstrates promising results on tabular datasets, whereas for time-series data, the gain from using text exists only in some domains.

Critical strengths
  1. This paper is globally well written, and I can follow most parts to understand the main contribution and framework. The tables and graphs clearly illustrate the results.
  2. The framework of using an LLM as a rejector for regressors and providing explanations in natural language is an interesting combination of existing methods.
Critical weaknesses
  1. The advantage over the baselines is very limited. For tabular datasets, the proposed method is better than SR-TabSTAR, the only multimodal baseline that takes text SS as input, on only 4 over 7 datasets. Even so, the improvement is slight (average 0.294 versus 0.313 of SR-TabSTAR, and there is no confidence intervals reported. For time-series datasets, this method performs better than the baselines on average, but the gain is not consistent across domains, and using text does not improve. In addition, I checked the paper "TabSTAR: A Tabular Foundation Model for Tabular Data with Text Fields". In their Table 7, TabSTAR has only 47.26M parameters, which is 200 times fewer than Qwen 9B. The running /inference time is not reported. Therefore, I'm not convinced at all whether this limited and inconsistent improvement is worth introducing such a heavy 9B model.
  2. The advantage of explanation is not clear. In Tables 5 and 6, I do not see consistent evidence that the extracted evidence is important for the rejector’s decision. Although the authors claim that on some datasets, e.g., Houses, the results support faithfulness, I'm not sure whether such support is a random observation or has practical meaning. Even if the authors report standard deviations in Figure 9, it is not clear whether these are over samples or experiments replications. Therefore, I'm not convinced by the dataset-specific conclusions about faithfulness.
  3. The novelty is incremental. It seems to me that the loss estimator in an existing selective regression framework is simply replaced by an LLM, and some existing fine-tuning techniques are applied.
Questions for authors
  1. Could the authors explain in table 1, for each number, how many experiments is done? If different splits are applied, could the author report the confidence interval for the average NAURC? Were SFT and GRPO independently trained for each split? Could the authors also report training time and the number of parameters? Do the multimodal baselines also take Y^\hat{Y} as input?
  2. Why the authors use this conformal calibration step? In table 2 and 4, I find the comparison somewhat misleading. The real coverage (of test set) can be different, e.g.in the traffic data in table 4 at α=0.5\alpha=0.5m the proposed method has actual coverage 0.07, whereas TEM achieves 0.39. Thus, I do not think the lower risk is an appropriate conclusion.
  3. Could the authors consider another baseline? In my opinion, one direct baseline is to use a pretrained embedding model combined with xgboost/lightgbm? For example, Qwen3-embedding-8B could be used to extract a vector, and then build a simple tree model on it. This could help clarify the benefit of using a generative LLM with different fine-tuning techniques.
3 comments
AuthorsOpenReview ↗

Response to Reviewer Ktyr (1/3)

We thank the reviewer for the detailed questions. Some of them call for discussion, and others are requests that we have met by revising the manuscript. We take the points for discussion first and then list the revisions that answer the remaining requests.

W1/W3-1. The contribution of this paper

W1 and W3 both depend on what the paper contributes, so we address this first. The contribution we intended has three parts. The first is a new problem setting in which text is used both to identify unreliable predictions of a fixed regressor and to explain rejections in natural language. The second is an LLM-based rejector as a first method for this setting. The third is a set of empirical findings on when text helps. The submitted version did not make this claim explicit. It introduced context-aware explainable selective regression as a "post-hoc framework", which reads as a method. We have revised the wording accordingly. The Abstract and Section 1 now formulate it as a new problem setting, and Section 1 presents the LLM-based rejector as a first method for this setting. Section 3 is retitled "Problem Setting", and Sections 4 and 6 use the same wording. Only the presentation has changed, and the method itself is the same.

In this setting, the baselines in Tables 1 and 3, including SR-TabSTAR, do not solve the problem. None of them produces an explanation, and the unimodal ones use no text. The NAURC tables test one of the two evaluation axes, selective regression performance, while the other is the faithfulness and plausibility of the explanations (Section 5.2.2).

We agree that the NAURC advantage over SR-TabSTAR is small. New Table 8 (Appendix C.1) gives confidence intervals over the four splits for every entry of Table 1. With Y^\hat{Y} added to SR-TabSTAR (see the list below), the proposed rejector has the best average rank and the second-best average NAURC. We therefore changed Section 5.2.1 from "lowest average NAURC" to "comparable to SR-TabSTAR". Against the existing unimodal rejectors (SR-Kernel and SR-kNN, with or without Y^\hat{Y}), the proposed rejector is better in both average NAURC and average rank.

On novelty (W3), we agree that the rejector is a simple instantiation, and the paper does not claim architectural novelty. Its contribution is the problem setting, a first method for it, and the empirical findings. Existing methods, including the strong multimodal baselines we constructed, do not address the setting because they cannot explain their rejections.

W1-2. Is a 9B model worth its cost?

New Appendix D reports the computational cost on Airbnb (Table 18) and the parameter counts (Table 19). Table 18 is reproduced below.

AirbnbSR-KernelSR-kNNSR-TabSTARQwen3-Embedding-8B + GBDTQwen3.5-9B (SFT+GRPO)
Training, total [s]0.9510.06124.7270.015,573.4
SFT stage [s]––––1,609.0
GRPO stage [s]––––13,966.4
Inference, loss estimation [s/sample]4.2e-31.4e-44.5e-30.1380.484
Inference, explanation generation [s/sample]––––8.790

Qwen3.5-9B has 9,424M parameters, of which 14M (0.15%) are trained by LoRA (Table 19). Training and inference are thus about two orders of magnitude more expensive than SR-TabSTAR, and we do not argue that the NAURC gain justifies this cost. What the 9B model provides is the explanation, which the problem setting requires and which TabSTAR cannot generate.

W1-3. Text does not help consistently on the time-series datasets

We agree that the gain varies across domains and that text does not improve the average NAURC on the time-series datasets, and Section 5.2.1 states both. The paragraph "Dataset-dependence of the text benefit" gives our hypothesis: tabular text describes instance-specific attributes, whereas time-series text consists of news of which only a part describes events that make the target hard to predict, and several Time-MMD domains contain only a few hundred observations. We present when text helps as one of the paper's empirical findings.

AuthorsOpenReview ↗

Response to Reviewer Ktyr (2/3)

W2-1. Is the Houses result meaningful, and are the dataset-specific conclusions supported?

The Houses result holds at the sample level. As shown in new Table 15, more than 60% of the rejected samples in Houses have positive comprehensiveness and less than 10% have negative comprehensiveness. In addition, the evidence extracted on Houses recovers the sentences that occlusion identifies as supporting the score better than random selection at k = 3 and k = 7 (new Table 17 below). These per-sample results, rather than the averages alone, are the basis for the statement about Houses.

More generally, on every dataset except Health and Social Good, fewer than half of the rejected samples have negative comprehensiveness. These results are given in Tables 15 and 16 of the revised manuscript. To test whether the extracted evidence reflects the sentences the score depends on, we identified by leave-one-sentence-out occlusion the sentences whose removal lowers the score and measured how many of them the evidence recovers (capped recall@k), against the expected value for k randomly selected sentences.

Capped recall@kk = 3, extractedk = 3, randomk = 7, extractedk = 7, randomk = 10, extractedk = 10, random
Airbnb0.1540.1970.1740.2500.3450.397
Beer0.7390.6410.7880.9270.8080.993
Houses0.4300.2830.4790.4010.6380.726
Laptops0.0520.1410.0640.2080.0690.378
Mercari0.7410.6300.8800.9050.9350.988
Permits0.3110.4970.4980.8060.5990.999
Wine0.2490.3540.3000.5260.7190.912

On Beer, Houses, and Mercari, the extracted evidence recovers score-supporting sentences better than random selection at k = 3 (and at k = 7 on Houses), whereas on the other four datasets it does not. We therefore conclude that the evidence is faithful to the scoring on some datasets and not on others. This is consistent with prior findings that LLM self-explanations can be unfaithful and that faithfulness varies across datasets (Turpin et al., 2023; Madsen et al., 2024).

Q2-1. The risk comparison in Tables 2 and 4

We agree with the reviewer. Comparing selective risks at different empirical coverages is not meaningful, and the sentence claiming lower risk "in most settings" was not justified. Tables 2 and 4 now mark in bold the lowest risk only among methods whose empirical coverage is within 0.05 of the target, and mark coverages that deviate by more than 0.05 with ∗\ast. The time-series paragraph of Section 5.2.1 was rewritten accordingly: it now reports the coverage deviations and attributes them to a shift between the calibration and test score distributions. We also added risk–coverage curves for all datasets and methods (Appendix C.3, Figs. 7 and 8) so that methods can be compared at equal empirical coverage. NAURC, on which the main conclusions rest, integrates over all thresholds and is unaffected by calibration.

Q2-2. Calibration method

For simplicity of implementation, we follow the calibration method of Li et al. (2024), which attains the target coverage in finite samples when calibration and test samples are exchangeable. The same rule is applied to every method except TEM, which uses the bin-based procedure of its original paper, so the comparison is consistent across methods. We recognize that exchangeability does not hold for chronologically split time series, which is why coverage drifts there for every method. We nevertheless kept this common rule so that all methods are compared under the same procedure. Calibration methods designed for non-exchangeable data could reduce these violations. We leave them to future work, and Appendix C.4 mentions recalibration on recent observations as one option.

Q3-1. What does the generative LLM add over a frozen embedding?

We added the suggested baseline, Qwen3-Embedding-8B + GBDT, to Table 1: a frozen Qwen3-Embedding-8B embedding of the text, concatenated with X and Y^\hat{Y}, is fed to LightGBM. The proposed rejector outperforms it on 6 of the 7 datasets and on average. A frozen LLM representation alone therefore does not reach the performance of the generative, fine-tuned rejector, and it provides no explanation for rejected samples.

AuthorsOpenReview ↗

Response to Reviewer Ktyr (3/3)

Q1 and other requests answered by revision

  • Number of runs and per-split training (Q1). Tabular: 4 random splits per dataset. Time series: one chronological split, 4 prediction lengths. Every method, including SFT and GRPO, is trained separately for each split or prediction length, and the tables report the average. This is now stated in Section 5.1 and in the captions of Tables 1 and 3.
  • Confidence intervals (Q1, W1). Added for the tabular datasets in Table 8 (Appendix C.1). For time series we use a single chronological split to avoid temporal leakage, and the reported values average over prediction lengths rather than over repeated splits, so we do not report intervals there.
  • Training time and parameter counts (Q1, W1). Added in Tables 18 and 19 (Appendix D) and reproduced in W1-2 above. The parameter count does not depend on the dataset.
  • Y^\hat{Y} as an input of the baselines (Q1). In the submission, only TEM used Y^\hat{Y}, and the other baselines followed their original implementations. We added "w/ Y^\hat{Y}" variants of SR-Kernel, SR-kNN, and SR-TabSTAR to Tables 1 and 3 (Section 5.1). SR-MMTSFlib and SR-TaTS have no such variant because MMTSFlib and TaTS take inputs of the look-back length, whereas Y^\hat{Y} has the prediction length. Adding Y^\hat{Y} improves the baselines, but the proposed rejector keeps the best average rank.
  • Error bars in Figure 9 (W2). The standard deviations in Figure 9 (Fig. 11 in the revision) are sample-wise standard deviations, averaged over the four splits for the tabular datasets and over the four prediction lengths for the time-series datasets. Section 5.2.2 and Appendix C.5 now say so.
  • Provenance of the multimodal baselines (W1). Section 5.1 now states that SR-TabSTAR, SR-MMTSFlib, and SR-TaTS are our constructions, not prior methods, and that no baseline produces explanations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.