{"id":"93b00d58-1ce2-4edb-ab90-1a62f2a5fa94","arxiv_id":"1908.05601","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Regularizing a deep network's local explanations to match human rationales during training improves accuracy on out-of-distribution text data without hurting in-distribution test accuracy.","lead":"This paper proposes CREX, a training method that uses expert-highlighted words, called rationales, to encourage deep learning text classifiers to base predictions on relevant evidence instead of dataset biases. The method improves accuracy on new, out-of-distribution review datasets, while leaving test-set accuracy roughly unchanged.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OOD generalization claim rests on untested faithfulness of the Eq. (2) omission explanation; if those scores are unfaithful, CREX's credibility metric is self-referential and the rationale-regularization story is unsupported.","rationale":"The paper's contribution is a training regularizer plus an empirical claim about credibility and OOD generalization. I read the strongest claim as the causal story in the abstract: CREX makes DNNs avoid dataset-specific artifacts, hence generalize beyond the test set. The most load-bearing premise is that the explanation used for regularization is faithful; if it is not, the loss may be doing something useful but the paper's interpretation is unsupported. The paper explicitly asserts this premise in Sec. IV-A rather than checking it: phrase and sentence omission are said to increase faithfulness relative to word omission [29], but faithfulness of the final trained model's explanations is never measured. The credibility metric in Eq. (11) compares the same explanation scores against rationales, so it is partly self-referential; a model could game the regularized loss by reducing all non-rationale omission scores without depending on rationales. The OOD experiments are a useful step, but they are averages over three runs without error bars, and hyperparameters were selected per model and dataset. I did not find an internal mathematical contradiction that would warrant rejection; the framework is coherent, and the ablations suggest that rationale content matters beyond pure sparsity. Thus the appropriate verdict remains conditional, pending an independent check of explanation faithfulness and statistical robustness. I agree with the reader's weakest assumption, and my concern adds the concrete point that Eq. (6) uses an absolute value and therefore does not even enforce the intended direction of the uncertainty penalty.","tokens_in":14885,"tokens_out":8798,"duration_ms":93116,"concrete_test":"Run one controlled synthetic text classification experiment with known ground-truth rationale tokens and an injected spurious token that is correlated with the label in the training split but uncorrelated in the test split. Train Vanilla, sparse-only, and CREX using the true signal tokens as rationales. The central claim is supported only if CREX beats Vanilla on the balanced clean test set and if, on the same models, an independent attribution method such as integrated gradients or leave-one-out ranks the true rationale tokens above the spurious token. If either condition fails, the reported credibility and OOD gains are tied to the chosen Eq. (2) explanation and do not demonstrate learning of the rationales.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is causal: aligning local explanations with expert rationales makes the model rely on rationales, which yields better OOD accuracy. This requires s_n(t,c) in Eq. (2) to be a faithful measure of what the DNN actually uses while it is being optimized. The paper lists faithfulness as a desideratum in Sec. IV-A and cites [29] for phrase omission, but it never verifies faithfulness for the trained models. Credibility is then measured only as agreement between the same regularized explanation and rationales via Eq. (11), so part of the reported improvement is built into the objective. The confident loss in Eq. (3) penalizes the L1 magnitude on non-rationale words, which can be minimized by making the model insensitive to those words without forcing positive use of rationale words. The uncertainty term in Eq. (6) maximizes an absolute probability difference, so even its directional claim, that deleting rationales should reduce f_yn, is not guaranteed. If deletion scores are unfaithful, CREX could improve OOD accuracy through generic sparsity or robustness rather than through the stated credibility mechanism. Tab. IV and Tab. V report only averages over three runs, with no error bars or significance tests, so the magnitude and reliability of the effect are hard to assess. This is not a fatal flaw, but it is the load-bearing assumption: without faithfulness, neither the credibility metric nor the explanation of the OOD gains is established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CREX, a training-time regularizer for text-classification DNNs that aligns omission-based local explanations (contribution scores over input words) with expert rationales. When rationales are absent, CREX applies an L1 sparsity penalty to the explanation. The authors instantiate the method with CNN, LSTM, and self-attention architectures and evaluate on movie-review and beer-review datasets. They report that CREX improves the credibility score (symmetric KL between normalized explanation and rationale) on the test set, preserves test accuracy, and improves accuracy on external or adversarially constructed datasets not used in training or hyperparameter tuning. The paper also includes ablations, hyperparameter sensitivity analyses, and robustness checks for rationale quantity and quality.","tokens_in":15198,"tokens_out":7181,"duration_ms":74807,"significance":"If the out-of-distribution generalization result is reliable, the paper makes a practically useful contribution: it suggests a simple, architecture-agnostic way to use sparse human rationales to improve DNN robustness beyond the i.i.d. test set. The external evaluation is a genuine strength: the Kaggle and Polarity datasets are not used for fitting or tuning, and the improvement appears across three different architectures. The sparse self-guidance variant without rationales is also a useful extension. However, the central mechanistic claim—that aligning explanations with rationales makes the model actually rely on those rationales—is not established, because the faithfulness of the omission-based explanation is asserted but never verified, and the credibility metric is essentially the training objective. The statistical evidence for the OOD claim is also thin (three runs, no error bars).","major_comments":[{"comment":"The explanation method in Eq. (2) is the foundation of both the regularization losses and the credibility metric, but the manuscript never verifies that these omission-based contribution scores faithfully reflect the features the trained model actually uses. The paper lists faithfulness as a desideratum and cites [29], but no empirical faithfulness check is provided for the trained vanilla or CREX models. If the scores are unfaithful, then aligning them with rationales does not by itself imply the model relies on rationales, and the reported OOD gains could stem from generic sparsity or implicit regularization rather than the stated credibility mechanism. Please add a faithfulness evaluation, such as deletion/insertion fidelity curves for top-ranked words or a comparison with an independent attribution method (e.g., gradient-based or Integrated Gradients), for both vanilla and CREX models.","section":"Section IV-A and Eq. (2)"},{"comment":"The confident-explanation loss only penalizes the L1 magnitude of contributions on non-rationale words; it contains no term that increases the magnitude of contributions on rationale words. The text claims that this 'implicitly' encourages dense explanation scores on relevant features, but that is not guaranteed by the loss form: with the supervised loss fixed, the model could satisfy Eq. (3) by becoming generally insensitive to all words rather than by focusing on the rationale. Please either add an explicit positive-attraction term for rationale words or provide direct evidence (e.g., attribution mass on rationale versus non-rationale words before and after training) that the intended focusing behavior actually occurs.","section":"Section IV-B1 and Eq. (3)"},{"comment":"The uncertain-explanation loss is not aligned with the stated intuition. Eq. (6) contains the absolute value |f_yn(x_n) - f_yn(x'_n)|, so the loss rewards any change in the ground-truth probability after deleting rationale words, including an increase; the directional claim in the same section that the probability for the label 'should be much smaller' is not enforced by the objective. In addition, Eq. (4) constructs the negative input by zeroing the embeddings of rationale words, which is a different perturbation from the token deletion used in Eq. (2) and in the text's example. Please correct the loss to a directional form (e.g., max(0, f_yn(x_n) - f_yn(x'_n))) and align the negative-input construction with the perturbation used by the explanation operator.","section":"Section IV-B2 and Eq. (6)"},{"comment":"The credibility metric is symmetric KL divergence between the normalized absolute explanation and the normalized rationale, which is essentially the same quantity minimized by the training losses in Eqs. (3) and (6). Therefore, the improved credibility scores in Table II are partly expected by construction and do not provide independent evidence that the trained models are more credible. This issue does not invalidate the external OOD accuracy results, but it does weaken the paper's interpretation of Table II. Please either report an independently defined credibility measure that is not directly optimized (e.g., agreement with held-out human rationales on a separate set, or consistency with an alternative attribution method) or explicitly acknowledge the circularity and rely on the OOD results for empirical support.","section":"Section V-B1 and Eq. (11)"},{"comment":"The central empirical claim that CREX 'significantly increases' accuracy on data beyond the test set is not supported by the reported statistics. All results are averages over three runs with no standard deviations, confidence intervals, or significance tests, and several improvements are small in magnitude (e.g., 0.5 percentage points on Kaggle for the attention model; 0.2 to 0.8 percentage points on the PR adversarial dataset). Please report run-level results with error bars and perform appropriate significance tests, and clarify whether the same hyperparameters were used across runs.","section":"Section V-C, Tables IV and V"}],"minor_comments":[{"comment":"The phrase length m used in Eq. (2) is not defined in the notation section; please introduce it explicitly before first use.","section":"Section III and Eq. (2)"},{"comment":"The normalization in Eq. (5) divides by the sum of absolute explanation scores, which is undefined if all scores are zero; please specify a small epsilon or a fallback for degenerate cases.","section":"Section IV-B2 and Eq. (5)"},{"comment":"The back-translation data augmentation uses the Google Translate API, which is non-deterministic and not reproducible; please specify translation settings or note the reproducibility limitation.","section":"Section V-A3"},{"comment":"The text reports 'relative accuracy improvement of 4.1%, 2.1%, 0.5%' for Kaggle, but the numbers in Table IV are percentage-point differences (e.g., 78.4 - 74.3 = 4.1 percentage points); use 'percentage points' or compute relative improvements correctly.","section":"Section V-C1"},{"comment":"Hyperparameters are tuned on the development set using both accuracy and credibility; because the credibility metric is the same as the training objective, this selection procedure may further inflate the test credibility results. Please clarify whether the reported test credibility was used in any form during model selection.","section":"Section V-A4"},{"comment":"The two qualitative examples are selected cases and cannot by themselves support the general credibility claim; consider adding a quantitative aggregate of explanation quality on a random sample of test instances.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is in scope and the external OOD evaluation is a genuine strength. The main risk is that the mechanistic story is under-supported: the faithfulness of the explanation operator is unverified, the credibility metric is nearly the training objective, and Eq. (6) has a directional flaw. A revision that adds faithfulness checks, fixes the uncertain-loss formulation, and reports error bars would make the contribution much stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CREX regularizes an omission-based local explanation to match expert rationales and adds a sparsity penalty when rationales are absent. The combination is new: prior work regularized intrinsic attention or used rationales as extra input, while CREX is architecture-agnostic and applies to post-hoc explanations. The paper's real strength is the out-of-distribution check: Kaggle and Polarity are genuinely external to the MR training set, and the consistent improvements over vanilla training across three architectures are worth taking seriously. The ablation study and the rationale quantity/quality analysis are also useful, and the authors are honest that in-distribution test accuracy does not always improve.\n\nThe soft spots are real but not fatal. The credibility metric in Eq. (11) is symmetric KL between the explanation and the rationale, which is essentially the same objective used in training. So the reported credibility gains are partly by construction; the accuracy gains on external data are the non-circular evidence. The bigger issue is faithfulness: Eq. (2) assumes deletion scores reflect what the model actually uses, but the paper never checks this on the trained models. If those scores are unfaithful, the regularization could improve OOD accuracy through generic sparsity or robustness rather than through the stated rationale-alignment mechanism. That alternative is plausible, not proven.\n\nTwo technical points: the absolute value in Eq. (6) contradicts the stated direction (deleting rationales should reduce f_yn, but the loss penalizes any change); and the report uses averages over three runs without error bars or significance tests, with hyperparameters tuned per model-dataset pair on the dev set. These are fixable in revision. No code is released, which makes the empirical claims harder to verify.\n\nBottom line: this is a solid paper for the interpretable-ML and domain-knowledge-guided learning community. The central mechanism is not fully established, but the OOD evaluation is a step in the right direction. I would send it to peer review with a request for faithfulness checks, error bars, and code. It deserves a serious referee.","headline":"CREX is a reasonable rationale-regularization method whose OOD generalization claim is worth testing, but the credibility metric is partly self-referential and the omission explanation's faithfulness is unverified.","tokens_in":15702,"tokens_out":1713,"would_cite":false,"duration_ms":17448,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a rationale-alignment penalty during training makes deep neural networks generalize better to new data.","keywords":["credible deep neural networks","rationale regularization","local explanations","text classification","out-of-distribution generalization","explanation alignment","sparse explanations","CREX"],"falsifier":"Take a CREX-trained model and a vanilla model, delete exactly the rationale tokens from held-out test instances, and compare the resulting prediction changes. If the CREX model is not substantially more sensitive to the removal of rationale tokens, then its improved explanation scores are not evidence that its decisions now depend on those tokens.","tokens_in":14677,"feed_emoji":"🎯","tokens_out":9074,"duration_ms":76072,"temperature":0.7,"pith_summary":"The paper sets out to make deep neural networks credible in a concrete sense: the words a model relies on for a prediction should match the evidence a human expert would cite. To achieve this, it adds a training loss called CREX that penalizes models whose local explanations disagree with expert rationales, plus a sparsity penalty for training instances that have no rationale. The central empirical claim is that this regularization does not reliably improve accuracy on the held-out test set, but it does increase accuracy on new, previously unseen datasets drawn from related tasks. The authors argue this matters because test-set accuracy can hide overfitting to dataset-specific bias, and a model that uses correct evidence is more likely to transfer and to earn trust.","feed_headline":"Match AI explanations to expert rationales for better generalization","feed_subtitle":"Keeping model explanations aligned with expert rationales boosts accuracy on new data while test accuracy holds steady.","key_machinery":"The mechanism is a differentiable, architecture-agnostic explanation score. The contribution of word $x_n^{(t)}$ to class $c$ is computed by deleting phrases containing that word and averaging the drop in prediction score, $$$s_n^{{(t,c)}}$ = \\frac{1}{m}\\sum_{j=1}^{m} [f_c(x_n)-f_c($x_n^{{(1:t-1-m+j)}}$ \\oplus $x_n^{{(t+j:T)}}$)].$$ CREX minimizes a combined loss: a confident-explanation term that penalizes explanation mass on features outside the rationale, an uncertain-explanation term that feeds the model the rationale-suppressed input $x_n' = x_n \\odot (1-r_n)$ and asks for lower confidence plus a more uniform explanation, and a sparsity term for instances without rationales. Expert knowledge flows into the network parameters through backpropagation, and the same loss applies to any differentiable text classifier.","core_discovery":"On the paper's own terms, the discovery is that a model's local explanation—the per-word contribution scores produced by deleting phrases from the input—can be used as a training signal, and steering those scores toward expert rationales makes the model rely on the evidence experts consider important. Across CNN, LSTM, and self-attention architectures on two text classification datasets, CREX-trained models show lower symmetric KL divergence between their explanations and expert rationales, and their accuracy on out-of-distribution datasets (Kaggle movie reviews, the sentence polarity dataset, and an adversarially filtered beer review set) rises by up to roughly four percentage points relative to vanilla training while held-out test accuracy remains comparable. The paper presents this as evidence that credible models have better real-world generalization, not merely better explanation scores.","pith_inferences":["Editorial inference: if faithful explanation is the active ingredient, the same recipe should transfer to image classification, where rationales could be object masks or bounding boxes and deletion-based attribution is already standard.","Editorial inference: the observed correlation between credibility and out-of-distribution accuracy suggests that explanation-to-rationale divergence could serve as a model-selection signal when a target distribution is unavailable.","Editorial inference: since the regularizer forces explanations to be sparse, a natural next test is whether CREX-trained models are also more stable under adversarial edits that remove rationale words; the paper does not report this.","Editorial inference: the conclusions are tied to one omission-based attribution method, so a direct check would be to repeat the experiments with gradient-based attribution to see whether the out-of-distribution gain is caused by explanation alignment generally rather than by the specific deletion procedure."],"forward_implications":["If the paper's claim is right, test-set accuracy alone is a poor yardstick for text classifiers meant to be deployed; out-of-distribution accuracy should be part of evaluation.","Because the benefit appears across CNN, LSTM, and self-attention models, the regularization is not tied to one network family and can be applied to other differentiable classifiers.","Sparse explanations alone, without any expert rationales, already beat vanilla training on both credibility and out-of-distribution accuracy, so the approach is usable when annotation is not available.","A few hundred rationales give nearly the full benefit of the complete rationale set, which keeps annotation cost down.","CREX adds no inference-time cost; the same trained model answers as quickly as a vanilla model, so the credibility gain does not slow deployment."],"supporting_citations":[{"why":"Supplies the Movie Review dataset with expert rationales and earlier evidence that annotator rationales improve text categorization.","marker":"[15]"},{"why":"Supplies the product review dataset's rationales and the extractive-rationale formulation used for comparison.","marker":"[16]"},{"why":"Provides the representation-erasure method that underlies the omission-based explanation score in Eq. (2).","marker":"[28]"},{"why":"Supports the paper's choice of phrase and sentence omission over single-word deletion for more faithful attribution.","marker":"[29]"},{"why":"Defines model credibility as agreement between explanation and human knowledge and supplies the symmetric-KL evaluation used here.","marker":"[5]"},{"why":"The closest prior work, deriving machine attention from human rationales; CREX extends this to black-box architectures.","marker":"[27]"},{"why":"Provides the rationale-augmented CNN baseline that trains on rationale snippets as extra data.","marker":"[14]"},{"why":"The source of the original Movie Review corpus that the MR dataset and its rationales are built on.","marker":"[37]"}],"fun_headline_variants":["Align AI reasoning with expert rationales to boost generalization","Expert rationales as training signal: credible AI generalizes better","Teach deep nets to justify like experts, gain on new domains","Rationale regularization: credible nets that excel on unseen data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that deleting phrases from the input and measuring the drop in prediction score faithfully reveals which words the model actually uses; if that attribution is inaccurate, matching it to expert rationales may not change what the model truly relies on.","fun_headline_variants_meta":{"raw":{"variants":["Align AI reasoning with expert rationales to boost generalization","Expert rationales as training signal: credible AI generalizes better","Teach deep nets to justify like experts, gain on new domains","Rationale regularization: credible nets that excel on unseen data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000941,"raw_usage":{"total_tokens":3998,"prompt_tokens":898,"completion_tokens":3100,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":3032}},"tokens_in":514,"tokens_out":3100,"duration_ms":21107,"temperature":1.0,"reasoning_tokens":3032,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:37:19.421231+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a CREX-trained model and a vanilla model, delete exactly the rationale tokens from held-out test instances, and compare the resulting prediction changes. If the CREX model is not substantially more sensitive to the removal of rationale tokens, then its improved explanation scores are not evidence that its decisions now depend on those tokens.","supporting_citations":[{"cited_title":"Using annotator rationales to improve machine learning for text categorization,","cited_arxiv_id":null,"evidence_quote":"Supplies the Movie Review dataset with expert rationales and earlier evidence that annotator rationales improve text categorization."},{"cited_title":"Rationalizing neural predictions,","cited_arxiv_id":null,"evidence_quote":"Supplies the product review dataset's rationales and the extractive-rationale formulation used for comparison."},{"cited_title":"Representation of linguistic form and function in recurrent neural networks,","cited_arxiv_id":null,"evidence_quote":"Supports the paper's choice of phrase and sentence omission over single-word deletion for more faithful attribution."},{"cited_title":"Learning credible models,","cited_arxiv_id":null,"evidence_quote":"Defines model credibility as agreement between explanation and human knowledge and supplies the symmetric-KL evaluation used here."},{"cited_title":"Deriving machine attention from human rationales,","cited_arxiv_id":null,"evidence_quote":"The closest prior work, deriving machine attention from human rationales; CREX extends this to black-box architectures."},{"cited_title":"Rationale-augmented convo- lutional neural networks for text classification,","cited_arxiv_id":null,"evidence_quote":"Provides the rationale-augmented CNN baseline that trains on rationale snippets as extra data."},{"cited_title":"A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts,","cited_arxiv_id":null,"evidence_quote":"The source of the original Movie Review corpus that the MR dataset and its rationales are built on."}],"review_version":1}