acceptodds
Under review as a conference paper at ICLR 2027

Learning to Explain Where Models Disagree

Abstract

Aggregate benchmark performance often does not reveal *where* two classifiers disagree. We study how to turn a small set of labeled agreement and disagreement examples into a concise natural-language description that transfers to new inputs. We introduce **SEAM**, an explanation policy trained with reinforcement learning to maximize pairwise simulatability: the accuracy with which an observer can use an explanation to predict whether the classifiers agree on an unseen input. The reward uses automatically obtained agreement between models and requires no human-written explanations or preference labels. Across 15 structured and text classification tasks held out from training, **SEAM** achieves 0.75 simulatability, compared with 0.69 for GPT-4o and 0.51 for its untrained base model. Its advantage persists for observer models excluded from reward computation. On 4 held-out multimodal tasks, a jointly trained vision-language variant achieves 0.68 simulatability. The advantage measured by the LLM observer also appears for human participants: in a controlled study with 60 participants, **SEAM** explanations support 74.0% agreement-prediction accuracy, compared with 55.4% for GPT-4o explanations. These results show that direct optimization can produce transferable natural-language summaries of pairwise model behavior.

Primary area:
interpretability and explainable AI
Keywords:
Explanations for ML Models, Explainability, Reinforcement Learning, Model Comparison

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_B8dLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework that explains the differences between model behaviors and predicts their agreements on new samples given the explanation. Specifically, it designs specific prompts to generate explanations given a subset of samples where a pair of models agree or disagree with each other, and fine-tunes a LLM/VLM with LORA to predict the agreement labels based on explanations. Results show that the proposed method achieves higher accuracy than generic LLMs, and the generated explanations generalize for both new models and humans.

Critical strengths

(1) Understanding the key differences between models is an interesting and important problem.

(2) The paper takes into account a variety of tasks for evaluations, and also carries out human evaluations to validate the ideas.

(3) Extensive ablation study is provided to facilitate evaluation of contribution for individual components.

Critical weaknesses

(1) The technical contribution of the paper is rather limited. It essentially prompts off-the-shelf LLMs to generate explanations, and fine-tunes another LLM/VLM to perform the binary classification task. Except for the definition of labels, most method design and evaluation metrics are identical to OPeX, e.g., OPeX also optimizes with GRPO and LORA.

(2) While Table 1 shows that the proposed method is better than generic LLMs, it is unclear if the advantage comes from the proposed method or simply additional training. In addition, the method is only marginally better than GPT-4o, and the evaluation sample size is very small (60 samples). For a fair comparison, it would be reasonable to compare other methods that involve optimization on the training data or at least incorporating training data in the context.

(3) Related to the previous comment. The models being compared are inconsistent across different evaluation settings. For instance, GPT-4o is the strongest compared method in text-based tasks, but missing for evaluation on multi-modal tasks.

(4) The paper highlights explanations of model differences as its key contribution. Nevertheless, looking at Table 5, it appears that the proposed method is “memorizing” the detailed differences of individual training samples, and does not generate high-quality explanations as GPT-4o. This again raises questions about overfitting, i.e., whether the improvement in binary classification comes from additional training.

(5) The human evaluation only considers the task where the proposed method shows the largest improvement over GPT-4o (Table 1). It would be reasonable to conduct more comprehensive evaluations over different task settings.

(6) Alongside explaining the samples where a pair of models agree or disagree, it would be more interesting to systematically explain the characteristics of each model. For instance, instead of only feeding 40 training samples, how would the method generate explanations about the defining characteristics of a classifier giving a large dataset (e.g., 10K samples)?

Questions for authors

(1) Please justify the key technical novelty over OPeX, in addition to the difference of problem setup.

(2) How do you disentangle the performance gain from the proposed method and access to additional training data?

(3) Please report the performance of GPT-4o on multi-modal tasks.

(4) Looking at Table 1 and Table 2, it appears that the QWEN family has close-to-chance performance. Any idea why? This is a potential problem because the proposed method also chooses QWEN models as the default backbones.

(5) Please elaborate on the quality of explanation generated by the proposed method (refer to the weaknesses section).

2 comments
AuthorsOpenReview ↗

Response to Reviewer B8dL

We thank the reviewer for their review. We first clarify a fundamental misunderstanding regarding our training pipeline.

Clarification of SEAM's RL Fine-Tuning: SEAM is an explanation generation model trained with RL (GRPO) to capture behavioral differences between two classifiers. Its reward is computed by a separate Observer LLM, which predicts from the explanation whether the classifiers will converge or diverge on an input. RL fine-tuning updates SEAM only; the Observer is a frozen reward evaluator, not the model being trained.

W1/Q1: Technical Contribution and Novelty over OPeX: We do not fine-tune an LLM/VLM for binary classification; we fine-tune SEAM, the explanation generation policy, to improve the contrastive diagnosticity of its explanations. Beyond the problem setup, SEAM differs from OPeX through (1) a pairwise reward computed from agreement between two classifiers rather than a single model's predictions, (2) multimodal model comparison (SEAM-VL), and (3) explanations conditioned on synthetic samples (Table 6). Optimizing for individual model behavior does not transfer to the pairwise problem: OPeX evaluated directly on agreement labels obtains 0.62 recoverability and 0.58 simulatability, compared with 0.85 and 0.75 for SEAM.

W2/Q2: Evaluation and Baselines: Our evaluation is designed to rule out overfitting to training data: evaluation tasks are entirely disjoint from training tasks, the classifiers being compared are new instances never seen during training, and simulatability is measured on held-out samples not used to generate the explanation. The comparison is also controlled: all methods receive the same 40 conditioning samples in context and are evaluated by the same frozen Observer on the same held-out samples. Our baselines span:

  • Prompted: Qwen2.5-VL-32B, Qwen3-32B, Llama-3.1-8B.
  • Larger models: GPT-4o, Llama-3.3-70B-Instruct.
  • Trained: OPeX on agreement labels (W1/Q1); decision trees and random forests fit directly on the data (0.57 and 0.55 simulatability; Table 11).
  • Reasoning: DeepSeek-R1-Distill-Llama-70B (lines 461–468).
  • Most directly related: SLED, i.e., GPT-4o conditioned on SLED-generated samples (0.70 vs. 0.77 for SEAM; Table 6).

The evaluation is not limited to 60 samples: 60 is the held-out set size for a single run, and each reported score for each task is averaged over 80 runs with freshly sampled conditioning and held-out sets. SEAM's improvement over GPT-4o is statistically significant over these runs and grows to 18.6 percentage points with human participants.

W3/Q3: GPT-4o Baseline for Multimodal Classifiers: Agreed. Please see below. We will include these results in the paper.

Multimodal Tasks (Average)RecoverabilitySimulatability
GPT-4o0.66 ± 0.100.59 ± 0.11
SEAM0.67 ± 0.100.63 ± 0.14
SEAM-VL0.78 ± 0.110.68 ± 0.16
Qwen2.5-VL-32B0.58 ± 0.110.53 ± 0.12
Qwen3-VL-32B0.59 ± 0.110.53 ± 0.12

W4/Q5: Explanation Quality and Generalization Beyond Training Examples: The goal is not verbose explanations, but precise, actionable patterns that generalize beyond the conditioning examples. Memorization yields high recoverability but low simulatability, as with random forests (0.96 vs. 0.55; Table 11); SEAM drops only from 0.85 to 0.75. SEAM produces more rules, but its maximum number of conjuncts per rule (3.47) matches Llama-3.3-70B (3.43), reflecting broader coverage rather than instance-specific rules (Table 13). Table 5 shows that GPT-4o's concise explanations often miss diagnostic distinctions, though more detail does not always help, as on Car Evaluation.

W5: Human Evaluation: Human evaluations across multiple tasks and modalities are resource-intensive, limiting the feasibility of large-scale studies. We therefore focus on the Rice task, which has a moderate number of features (7), avoiding excessive cognitive load on participants.

W6: Scope: We agree this is an interesting but distinct problem: characterizing a single classifier over a large dataset, which prior work addresses (OPeX, MaNtLE, LIME, Anchors). Our work focuses on explaining behavioral differences between classifier pairs. For large datasets, informative samples can be selected with sample-discovery methods such as SLED, on whose samples SEAM can be conditioned (Table 6).

Q4: Choice of Backbone: The untuned Qwen models are near chance because they are not optimized for contrastive diagnosticity. RL fine-tuning raises the same backbone from 0.51 to 0.75 simulatability, showing that the gains come from our method rather than the backbone. This holds across backbones: SEAM-Llama improves Llama-3.1-8B from 0.49 to 0.67 (Figure 3a). We chose Qwen2.5-VL-32B-Instruct for its model capacity and multimodal capabilities.

We value your review, which has improved the clarity of our paper. Given these clarifications, we invite you to reconsider your evaluation scores.

I thank the authors for the additional results and clarification. While it addresses a few concerned I raised in the initial review, I am still not fully convinced about the technical contribution of the paper. Specifically,

(1) I understand that the paper is not finetuning a LLM. Nevertheless, the proposed policy model is essentially a finetuned version of Qwen2.5-VL-32B-Instruct. The training algorithm and model design largely follows previous studies (e.g., OPex), with the major difference being the definition of labels. While the authors highlight a few subtle design choices and experiment setups, I don't think it bring sufficient technical novelty.

(2) I am aware that the paper uses performance averaged across multiple repeated runs. However, my point is that each run only considers a limited set of samples (40-60), which can lead to large variance in performance. For a more comprehensive evaluation, running on larger set is more preferrable.

(3) The rebuttal argues that the generated explanations focus on broader coverage. However, given the examples in the paper, I think this choice considerably deteriorate the interpretability of the explanations, as it reads like reciting every detailed element in the scene.

(4) The rebuttal does not address my concerns on human evaluation and the choice of backbone. For instance, if QWEN models have the worst performance in generating reasonable explanations, why should we use it as the backbone for policy model and observer?

As a result, I am leaning toward keeping my original score.

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

The paper proposes SEAM, a language model trained to describe in natural language where two classifiers agree and where they disagree. Given 20 inputs on which two classifiers agree and 20 on which they disagree, SEAM writes a short paragraph like "The models diverge when ... The models converge when ...". The policy (Qwen2.5-VL-32B with LoRA) is trained with GRPO, a reinforcement learning method, using as reward the accuracy of a separate LLM (Qwen3-VL-32B) that reads the explanation and predicts agreement on 40 new inputs. The method requires no human-written explanations, since agreement labels come directly from the two classifiers. On 15 held-out structured and text tasks, SEAM is compared on simulatability with five LLM-based baselines (GPT-4o, its untrained base model, Qwen3-32B, Llama-3.3-70B, and concatenated single-model explanations from OPeX). It has the highest average (0.75 vs. 0.69 for GPT-4o) and the best score on 11 of the 15 tasks. On 4 multimodal tasks, a variant trained with image-text data (SEAM-VL) is compared only with other SEAM variants and two untrained Qwen models. In a human study with 60 participants on one task, people using SEAM explanations predicted agreement with 74.0% accuracy, compared with 55.4% with GPT-4o explanations.

Critical strengths
  • Explaining model differences in natural language form is a practical problem for model selection, and agreement labels make the training signal cheap to obtain.
  • The evaluation uses tasks held out from training, checks that gains hold when other LLMs make the predictions, and includes a human study in which the gain carries over to people.
Critical weaknesses
  • The quality of the explanations themselves is not studied in depth, although producing them is the purpose of the paper. Nearly all evidence comes from one automatic measure: how well an LLM can use an explanation to predict agreement. The human study covers one task and a comparison with GPT-4o, and it does not measure how understandable or useful participants find the explanations. The qualitative analysis is limited to four examples and automatic counts of rules, and some of these examples look fitted to the 40 conditioning examples rather than general: on Rice, the models "diverge when Major Axis Length is between approximately 185 and 195" with three further narrow ranges; on Car Evaluation, SEAM lists many door and passenger combinations and is beaten by GPT-4o's single rule on safety; and one Wine explanation notes that its conditions hold "with no exceptions in the provided data." The drop from recoverability (0.85) to simulatability (0.75) points the same way, and the paper never checks whether the described regions actually produce disagreement.
  • The method is close to prior work, and the comparison does not isolate what is new. The observer-based objective and both metrics come from OPeX (Malladi and Srivastava 2026), and the task setup comes from SLED (Malladi et al. 2025). The main differences are the prediction target (the agreement of two models), training on agreement-labelled and synthetic data, and the multimodal extension. The OPeX baseline concatenates two single-model explanations rather than applying OPeX to agreement labels, and GPT-4o is the strongest LLM baseline, so it is not clear how much of the gain comes from SEAM's training.
Questions for authors

Can you evaluate explanation quality across several tasks, for example how understandable and useful the explanations are, and compare against GPT-4o and a simple interpretable baseline such as a shallow decision tree trained on the same 40 examples (or a similar setup)? It'd be especially interesting and in to dig deeper into when the natural language explanations seem uninterpretable or unintuitive to humans.

4 comments
AuthorsOpenReview ↗

Response to Reviewer RXis

We thank the reviewer for their feedback.

W1: Reliance on Recoverability and Simulatability: Following prior work (MaNtLE [1], OPeX [2], SLED [3]), explanations are evaluated on two complementary dimensions: faithfulness to the conditioning samples (recoverability) and generalization to unseen samples (simulatability). These metrics enable scalable training and evaluation across tasks, modalities, and model families. Our findings are not tied to a single LLM: SEAM's gains hold across four observers that provided no training reward (Fig. 3b) and transfer to human participants.

W1: Human Study: Human evaluations across multiple classifiers are resource-intensive, limiting the feasibility of large-scale studies. We thus focus on the Rice task, which has a moderate number of features (7), avoiding high cognitive load on participants. While the study measures user simulatability, participants in the GPT-4o condition also reported that the explanations "could be far clearer" and required multiple readings (lines 1013–1017).

W1: Qualitative Analysis Limited to Four Examples: The examples in the main body are representative cases while observations were made from multiple explanations. The rule-count statistics (Table 13) are averaged over all generated explanations.

W1: Generalization Beyond Conditioning Examples: The RL reward is computed on a disjoint set of samples the policy never observes, so explanations that merely fit the conditioning examples cannot obtain high rewards. Simulatability, our primary metric, measures prediction on unseen inputs: SEAM achieves 0.75 on structured and text tasks and 0.68 on multimodal tasks, outperforming all baselines, including GPT-4o (0.69 and 0.59). The examples highlighted by the reviewer generalize: Rice and Wine reach held-out simulatability of 0.88 and 0.77 (Table 1), and Rice is also the task of our human study, where SEAM explanations improved human accuracy over GPT-4o. Precise thresholds that hold on unseen inputs reflect precision, not overfitting. The phrase "no exceptions in the provided data" stems from our prompt, which requires conditions to hold for all conditioning samples (Table 7). Car Evaluation is an exception (see Q1).

W1: Drop from Recoverability to Simulatability: Some drop is expected, since recoverability is measured on the conditioning samples and simulatability on unseen ones. Fitting the conditioning samples looks like random forests: 0.96 recoverability but only 0.55 simulatability (Table 11). What matters is performance on unseen inputs: SEAM's simulatability (0.75) exceeds every baseline's simulatability.

W1: Whether Described Regions Produce Disagreement: Simulatability directly tests this. The Observer applies the regions described in the explanation to each held-out input and is scored against whether the classifiers actually disagree on that input.

W2: Relation to Prior Work: Explaining model differences splits into two sub-problems (lines 38–39): sample discovery and explanation generation. SLED [3] addresses sample discovery. We address generation: given the conditioning samples, we optimize the explanation directly for contrastive diagnosticity. The two are complementary, since SEAM conditioned on SLED's synthetic samples transfers to real data (Table 6). OPeX [2] also optimizes explanations, but only for a single model's predictions. Agreement is a relation between two models, not the output of either, so explanations of each model alone do not yield an explanation of their differences. We also extend explanation optimization to vision-language classifiers, a setting neither prior work addresses. Each contribution is isolated by a controlled comparison:

  • RL fine-tuning vs. the base model: 0.75 vs. 0.51 (Table 1).
  • Pairwise objective vs. concatenated OPeX: 0.85 vs. 0.54 recoverability and 0.75 vs. 0.54 simulatability (Tables 1, 9).
  • Synthetic-data training vs. SEAM-R: 0.77 vs. 0.69 (Table 6).
  • Multimodal training vs. SEAM: 0.68 vs. 0.63 (Table 2).

W2: OPeX on Agreement Labels: We additionally evaluate OPeX directly on agreement labels: 0.62 recoverability and 0.58 simulatability, vs. 0.85 and 0.75 for SEAM.

Q1: Explanation Quality: Against decision trees and random forests trained on the same examples, SEAM achieves higher simulatability on structured tasks (Table 11). Against GPT-4o, SEAM performs better on average (Table 1), and helps humans more (W1: Human Study). SEAM explanations become harder to follow when they list many feature combinations, as on Car Evaluation, where GPT-4o's single rule works better (Table 5). Limiting explanation length helps: a 256-token SEAM still outperforms GPT-4o (Table 4). We agree that human ratings of understandability across more tasks would be valuable and will add this as a limitation.

We value your review, which has improved the clarity of our paper. Given these clarifications, we invite you to reconsider your evaluation scores.

AuthorsOpenReview ↗

References

References

[1] Menon, R. R., Zaman, K., and Srivastava, S. (2023). MaNtLE: Model-agnostic Natural Language Explainer. EMNLP 2023.

[2] Malladi, A., and Srivastava, S. (2026). Directly Optimizing Natural Language Explanations for Behavioral Faithfulness: Simulatability and Recoverability. ICML 2026.

[3] Malladi, A., Menon, R. R., Jain, Y., and Srivastava, S. (2025). Explaining Differences between Model Pairs in Natural Language through Sample Learning. EMNLP 2025.

As we approach the end of the author response period, we would much appreciate any follow-up thoughts you might have regarding our clarifications, and if they have have helped resolve outstanding concerns and influenced your overall assessment.

Thank you for the rebuttal; it partially clears my concerns. What remains is the in-depth human and qualitative study of the explanation quality, which is left for future work. Since producing explanations is the goal of the paper, I will keep my rating.

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

The paper studies how to describe in plain language where two classifiers agree or disagree. Its method, SEAM, is an explanation model trained by reinforcement learning, rewarded by how well an LLM observer can use the explanation to predict agreement on held-out inputs. The authors find that, on 15 held-out structured and text tasks, SEAM reaches 0.75 simulatability compared with 0.69 for GPT-4o.

Critical strengths
  1. The experimental design is relatively broad and comprehensive. SEAM is tested on 15 held-out structured and text tasks and 4 multimodal tasks against several open and closed LLM baselines, with ablations on token budget and input modality, transfer to four observer models that gave no training reward, and a user study with 60 participants.

  2. Table 1 and Table 2 show consistent gains and positive experimental results. In Table 1, SEAM has the highest average simulatability (0.75, against 0.69 for GPT-4o and 0.51 for its base model) and the best score on 11 of the 15 tasks; in Table 2, SEAM-VL is best on every multimodal task, with 0.78 recoverability and 0.68 simulatability on average.

Critical weaknesses
  1. It would be better to justify why recoverability and simulatability are the right targets and what is gained when they improve. Each score asks only whether an observer can tell if the two classifiers agree, not which of them is right, and recoverability is computed on the very examples that were used to generate the explanation.

  2. It would be better to reduce the dependence on a single LLM observer. Because Qwen3-VL-32B-Instruct provides both the training reward and the main evaluation scores, the explanations may be tuned toward what this type of observer finds easy to use.

  3. It would be better to extend the human study. It covers one task (Rice), 30 participants per condition, and 30 test samples, and it reports accuracy plus a few participant comments, so it is hard to know how the result carries over to text or multimodal tasks or to long rule lists such as the Car Evaluation example, where GPT-4o scored higher.

  4. It would be better to fix a few presentation details. Appendix A.4 refers to "Table ??", and the introduction reports an "18%" human accuracy gain while the reported values (74.0% versus 55.4%) differ by 18.6 percentage points. The statistical test is also described as a "two-way t-test" without stating whether it is paired or two-sample.

Questions for authors
  1. Can the authors describe a concrete use, such as model selection, routing, or finding failure slices, where a higher simulatability score leads to a better outcome, and report any evidence for it?

  2. Why is recoverability reported as a main metric, given that it is computed on the conditioning examples and could be high for explanations that simply restate the shown cases?

  3. When an explanation predicts disagreement well, how often does it also identify inputs on which one classifier is wrong, which is frequently what practitioners need?

3 comments
AuthorsOpenReview ↗

Response to Reviewer RNYq

We thank the reviewer for their feedback.

W1: Why Recoverability and Simulatability?: SEAM's goal is to compare classifiers, not to evaluate them, so agreement is the target by design; which classifier is correct is outside our scope (lines 78–80). Recoverability and simulatability capture two complementary dimensions of explanation quality: faithfulness to observed behavior and generalization to unseen inputs. (1) Recoverability: Building on prior work on explanation faithfulness (Jacovi and Goldberg, 2020; MaNtLE, Menon et al., 2023; OPeX, Malladi and Srivastava, 2026), recoverability measures whether an observer can recover the classifiers' agreement and disagreement behavior on the conditioning examples used to generate the explanation. It is a necessary condition: an explanation that cannot account for its own conditioning examples is unfaithful. (2) Simulatability: Following Hase and Bansal (2020), simulatability measures whether an observer can use the explanation to predict agreement and disagreement on previously unseen inputs, testing whether the explanation captures generalizable patterns rather than summarizing the conditioning examples.

W2: Observer Dependence: This concern is already addressed in our “Transfer across Observer Models” experiment (Figure 3b; lines 430–439). SEAM's explanations transfer to observers outside the reward model's family and to humans. Under GPT-4o and Llama-3.3-70B-Instruct, recoverability and simulatability stay within two percentage points of the training observer (Figure 3b), and SEAM's advantage over baselines persists. Human participants using SEAM explanations reach 74.0% accuracy versus 55.4% with GPT-4o explanations.

W3: Human Evaluation: Human evaluations across multiple tasks and modalities are resource-intensive, limiting the feasibility of large-scale studies. We therefore focus on a controlled evaluation on the Rice task, where SEAM demonstrates substantially higher human simulatability than GPT-4o. We discuss tasks with long rule lists, such as Car Evaluation, in Table 5 and will note this limitation in the revision.

W4: Presentation: We will fix the unresolved table reference (Table ??), report the gain as 18.6 percentage points, and specify the t-test type.

Q1: Practical Applications of SEAM Explanations: The quality of an explanation is measured by the downstream decisions and predictions it enables through simulatability. While evaluating practical utility is a promising research direction, it remains outside the scope of this work. SEAM explanations enable several practical uses: (1) Failure explanation: SEAM explanations are built from samples labeled by f1 and f2. If f2 is replaced with ground-truth labels, divergence identifies classifier errors, allowing SEAM to describe failure regions. (2) Model selection: SEAM explanations identify input regions where model choice matters and where either model can be used without changing predictions. (3) Cost-aware routing: When one model is cheaper, route inputs predicted to fall in the convergence region to it and use the more expensive model otherwise. We will clarify these applications in the revision.

Q2: Simulatability over Recoverability: Agreed. Simulatability is our primary metric and is reported in every main results table; recoverability is complementary.

Q3: Model Failure Explanations: By replacing f_2 labels with ground-truth labels, divergence directly identifies classifier errors, allowing SEAM to explain the regions where the classifier fails.

We thank the reviewer for recognizing the breadth of our experimental evaluation and the consistent gains achieved by SEAM. We believe the current score/rating does not fully reflect the strengths acknowledged in the review. Given these new clarifications, we invite you to reconsider your evaluation scores.

References

  • Hase, P., and Bansal, M. (2020). Evaluating Explainable AI: Which Algorithmic Explanations Help Users Predict Model Behavior? ACL 2020.
  • Jacovi, A., and Goldberg, Y. (2020). Towards Faithfully Interpretable NLP Systems: How Should We Define and Evaluate Faithfulness? ACL 2020.
  • Malladi, A., and Srivastava, S. (2026). Directly Optimizing Natural Language Explanations for Behavioral Faithfulness: Simulatability and Recoverability. ICML 2026.
  • Menon, R. R., Zaman, K., and Srivastava, S. (2023). MaNtLE: Model-agnostic Natural Language Explainer. EMNLP 2023.

As we approach the end of the author response period, we would much appreciate any follow-up thoughts you might have regarding our clarifications, and if they have have helped resolve outstanding concerns and influenced your overall assessment.

Thank you for the clarification. I will keep my previous rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.