{"id":"97f6794b-f8ae-4a4c-8497-4961c5244add","arxiv_id":"1908.10763","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"DRiFt, a residual-fitting debiasing algorithm, improves NLI model accuracy on challenge sets like HANS by training on examples a biased model cannot solve.","lead":"The paper presents DRiFt, a two-step training method that first learns a biased model from known superficial cues in language inference data, then trains the main model on the remaining examples the biased model cannot solve. It matters because it offers a simple way to harden NLP systems against known dataset biases without changing the model architecture.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline claim is scoped too broadly: on STRESS, DRiFt often decreases macro-F1, and in-domain drops of 20+ points for CBOW/HAND undermine 'reasonable performance.'","rationale":"The reader's weakest_assumption focuses on the mechanism: if I(x) contains useful semantics, DRiFt trains away useful signal. That is supported by Table 3 and is a real risk. My stress-test points at a related but more direct problem with the headline: the evaluation metrics used to support the central claim are class-wise F1 scores, and aggregating them shows that the claimed gains are not consistent across configurations, especially on STRESS. This is not an objection to the method's idea, which is clearly presented and has credible support from the synthetic experiment and HANS; it is an objection to the scope of the empirical claim. The released code and reproducible setup are positive evidence, but the absence of error bars and aggregate metrics makes the headline assertion hard to verify as stated. A conditional verdict with a request for aggregate metrics and a narrowed claim is appropriate, matching the reader's verdict. I do not see grounds for rejection, because the core residual-fitting idea is sound under its stated assumptions and there are configurations with large, robust-looking improvements.","tokens_in":25999,"tokens_out":14397,"duration_ms":158325,"concrete_test":"Recompute Table 5 and Table 4 using macro-F1 over all three labels for every model x biased-classifier configuration, and report paired bootstrap confidence intervals for the DRiFt-minus-MLE difference. If DRiFt is not better than MLE for a majority of configurations on STRESS, the abstract and conclusions should be narrowed to HANS or to the challenged class only; if the confidence intervals are wide, the word 'significant' should be replaced with a quantitative statement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract asserts 'significant gains on two challenge test sets, while maintaining reasonable performance on the original test sets,' but the reported evidence supports this only for selected configurations. Table 5 reports per-class F1 and highlights gains on the challenged entailment class, while the same table shows large simultaneous drops on contradiction and neutral. For example, on STRESS-Negation, DA with DRiFt-CBOW has macro-F1 (28.4+21.4+39.5)/3 = 29.8 versus MLE (17.4+47.3+55.3)/3 = 40.0; on STRESS-Overlap, DA DRiFt-CBOW has macro-F1 40.2 versus MLE 55.0. Even BERT loses on STRESS-Overlap with DRiFt-CBOW and DRiFt-HAND (macro-F1 51.6 and 51.4 versus MLE 53.9). Thus 'significant gains on two challenge test sets' depends on focusing on one class rather than an overall metric. In-distribution, Table 3 shows DRiFt-CBOW lowers SNLI accuracy by 22.7 points for DA and 25.7 points for ESIM, and DRiFt-HAND lowers it by 10.3 and 8.8 points, respectively. That makes 'reasonable performance on the original test sets' also configuration-dependent. The central empirical claim is therefore accurate only for specific model/bias pairs, particularly BERT with a matched bias, and should be narrowed accordingly.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DRiFt, a two-stage training procedure for NLI under known dataset bias. A biased classifier is first trained on insufficient features I(x) (hypothesis-only, bag-of-words, or handcrafted overlap/negation features); a debiased classifier is then trained to fit the residual of the biased classifier under a joint softmax objective. At test time only the debiased classifier is used. The authors derive a gradient reweighting interpretation (Eqs. 9-13) and evaluate on SNLI and MNLI with DA, ESIM, and BERT, measuring in-distribution accuracy, synthetic bias robustness, and performance on HANS and STRESS challenge sets. The central empirical claim is that debiased models achieve significant gains on the two challenge sets while maintaining reasonable in-distribution performance.","tokens_in":26266,"tokens_out":6238,"duration_ms":56272,"significance":"If the stated empirical claim held across the reported configurations, DRiFt would be a practical and widely applicable debiasing method. The algorithm is simple, the derivation is self-contained, and the code is released; the best configuration (BERT with HAND) raises HANS non-entailment F1 from 7.8 to 70.9, which is a strong falsifiable demonstration. However, the results are configuration-dependent: macro-F1 on STRESS often decreases, and in-distribution drops of over 20 points for DA/ESIM with CBOW undermine the 'reasonable performance' phrasing. The contribution is therefore real but more modest than the abstract states.","major_comments":[{"comment":"The abstract claims 'significant gains on two challenge test sets,' but on STRESS the gains are concentrated in the entailment class while macro-F1 often decreases. For example, DA with DRiFt-CBOW yields macro-F1 (28.4+21.4+39.5)/3 = 29.8 on STRESS-Negation versus MLE (17.4+47.3+55.3)/3 = 40.0, and a similar drop occurs on STRESS-Overlap (40.2 vs 55.0). Even BERT's macro-F1 on STRESS-Overlap drops from 53.9 to 51.6 with DRiFt-CBOW and 51.4 with DRiFt-HAND. The paper should report aggregate metrics and significance tests, and the abstract should be narrowed to note that gains are configuration- and class-dependent.","section":"Section 4.6, Table 5"},{"comment":"The claim in the abstract of 'maintaining reasonable performance on the original test sets' is not supported for DA and ESIM with CBOW or HAND biased classifiers. SNLI accuracy drops 22.7 points for DA with DRiFt-CBOW (85.3 to 62.6) and 25.7 for ESIM (88.0 to 62.3); MNLI drops 15.9 points for DA with DRiFt-CBOW and 9.2 with DRiFt-HAND. The paper acknowledges degradation in Section 4.3, but the abstract's 'reasonable' is too strong; the claim should be qualified to specific configurations or the method should be evaluated with selection criteria that preserve in-domain accuracy.","section":"Section 4.3, Table 3"},{"comment":"The text says 'DRiFt improves performance on non-entailment in all cases with little degradation on entailment,' but the DA rows do not support 'little degradation': DRiFt-HAND lowers entailment F1 from 66.6 to 60.5 on lexical, 66.6 to 61.4 on subsequence, and 66.5 to 55.9 on constituent. Also DRiFt-CBOW lowers entailment F1 on lexical from 66.6 to 65.3 and on constituent from 66.5 to 65.1. This is a load-bearing discrepancy because the conclusion that DRiFt is preferable to RM relies on preserving entailment performance while improving non-entailment; for DA, RM-HYPO preserves entailment better than DRiFt-HAND. The claim should be revised to specify for which models it holds.","section":"Section 4.5, Table 4"},{"comment":"The paper assumes without proof that discarding the biased classifier at test time is sound; the derivation in Eqs. 9-13 shows that DRiFt trains the sum f_s + f_d to be a good model, not that f_d alone is calibrated for p(y|x). The synthetic experiment in Figure 2 itself shows the cost: DRiFt-HYPO is below RM-cheat, which the text attributes to HYPO capturing unbiased information. This is an inherent limitation that should be stated as an assumption and tested by comparing DRiFt with an ensemble (f_s + f_d) at test time, or by ablating the biased classifier with features that are known to be purely biased.","section":"Section 3.1-3.2, Figure 2"}],"minor_comments":[{"comment":"The constant C is written as sum_{(x,y) in D} log p_s^*(k|I(x)) with a free index k; it should be p_s^*(y|I(x)) over the true labels.","section":"Section 3.2, Eq. (10)"},{"comment":"The baseline is typeset as 'R M' in prose (e.g., Sections 3.1 and 4.3) but as 'RM' in tables; please unify the notation.","section":"Notation throughout"},{"comment":"The Antonym and Length stress subsets have missing entries (shown as '-') in several rows; the caption should state that these classes are absent from those subsets.","section":"Appendix A, Table 6"},{"comment":"There is a typo in the related work paragraph on reading comprehension: 'reading comparehension' should be 'reading comprehension'.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid contribution and the derivation is sound, but the empirical claims need to be scoped honestly and supported with significance testing. The authors should be encouraged to report macro-F1 and per-configuration breakdowns rather than only the entailment-class F1 that favors their method. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core of this paper is a simple residual-fitting procedure for debiasing NLI models, and it mostly works as advertised when the biased classifier is matched to the bias. The label-shift formalization is clean, the gradient analysis in Section 3.2 is genuinely illuminating, and the synthetic-bias experiment is a nice controlled validation. The paper also does the right thing by acknowledging the concurrent work of Clark et al. and by reporting the full STRESS results even where they are not flattering.\n\nThe biggest issue is the abstract. It claims 'significant gains on two challenge test sets, while maintaining reasonable performance on the original test sets.' That is only true for specific model/bias pairs. On STRESS, macro-F1 frequently drops relative to MLE: for DA with DRiFt-CBOW, macro-F1 on negation goes from 40.0 down to 29.8, and on overlap from 55.0 down to 40.2. The in-domain numbers are worse: DRiFt-CBOW drops DA from 85.3 to 62.6 and ESIM from 88.0 to 62.3 on SNLI. The paper does acknowledge the trade-off in Section 4.6, but the abstract overstates the case.\n\nThe other soft spots are real but less serious. There are no error bars or significance tests, which matters when some gains are small. The method's core assumption is that the chosen biased features isolate dataset bias and carry no useful semantics; the large drops with CBOW and HAND show that this assumption fails in practice for weaker models. The HAND features were built from the same error analysis that produced the STRESS test, so there is a mild circularity concern on that evaluation. None of this is fatal, but it should be confronted directly.\n\nWho is this for? Anyone working on NLI robustness, dataset bias, or distribution shift will get value from the formalization and from the careful dissection of when residual fitting helps. It deserves a serious peer review; I would not desk reject it. The authors should be asked to narrow the abstract, report variance, and add an analysis of the trade-off between in-domain accuracy and robustness. I would cite this as a baseline debiasing method, and it is a reasonable reading-group paper for a robustness seminar.","headline":"A solid, honestly-reported debiasing method with a clean derivation and real gains in matched settings, but the abstract's 'significant gains' and 'reasonable performance' are too broad and need narrowing.","tokens_in":26840,"tokens_out":2511,"would_cite":true,"duration_ms":28666,"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":"A two-stage procedure that first learns a shallow biased model and then fits the residual can unlearn dataset bias in natural language inference, improving scores on challenge sets while keeping most in-distribution accuracy.","keywords":["natural language inference","dataset bias","debiasing","residual fitting","distribution shift","label shift","annotation artifacts","HANS challenge set"],"falsifier":"Compute the biased classifier's accuracy on both training and a held-out set drawn from the same distribution. If $I(x)$ remains highly predictive on the held-out set, there is no label shift, and DRiFt should underperform MLE by the amount of useful signal in $I(x)$; the paper's own in-distribution results for DRiFt-CBOW and DRiFt-HAND on DA and ESIM already show drops of more than twenty points, the predicted failure mode. A sharper test: construct a challenge set where the known bias association is preserved rather than reversed—if DRiFt still improves over MLE there, residual fitting is doing something other than removing label shift.","tokens_in":25724,"feed_emoji":"🎯","tokens_out":11102,"duration_ms":106746,"temperature":0.7,"pith_summary":"This paper aims to establish that known dataset bias in natural language inference—spurious cues such as negation words or high word overlap that correlate with labels in crowd-sourced training data—can be unlearned without deleting the cues from the input. It formalizes bias as label shift on a known insufficient feature set $I(x)$, then proposes DRiFt: first train a biased classifier on $I(x)$, freeze it, and train the full model to fit the residual, so examples the biased model already explains contribute little gradient. On SNLI and MNLI, DRiFt improves scores on the HANS and STRESS challenge sets relative to maximum-likelihood training while keeping most in-distribution accuracy, with the trade-off largest for weaker models and biased features that carry real semantics. The reason to care is that this is a general recipe: it converts prior knowledge about a specific annotation artifact into a training objective that focuses the model on examples the artifact cannot solve.","feed_headline":"A biased-model residual recipe beats language-AI dataset bias","feed_subtitle":"Train a weak biased model first, then fit its residual; HANS and STRESS scores rise while in-domain accuracy mostly holds.","key_machinery":"The load-bearing object is the additive residual model $f^*(x)=f_s(I(x);\\theta^*)+f_d(x;\\phi^*)$, with the biased classifier $f_s$ fixed after being trained on insufficient features $I(x)$ (hypothesis-only text, bag-of-words embeddings, or handcrafted overlap and negation features). The residual fitting is what does the debiasing: it turns maximum-likelihood training of $f_d$ into a per-example gradient reweighting in which confidently biased examples are down-weighted toward zero gradient, forcing $f_d$ to learn from examples that cannot be solved by the bias. At test time the biased component is discarded, so the method never tries to remove or project out the biased features from the input.","core_discovery":"The central claim is that a standard NLI model trained by maximum likelihood absorbs the label association carried by shallow input features, and that this association can be removed by fitting the residual of a biased model. Given a feature map $I(x)$ believed to contain dataset bias, DRiFt first solves $\\theta^* = \\arg\\min_\\theta \\mathbb{E}_P[L(f_s(I(x);\\theta),y)]$, then trains a full model $f_d$ on the additive objective $f_s(I(x);\\theta^*) + f_d(x;\\phi)$, and at test time uses only $f_d$. Under cross-entropy, this objective is equivalent to reweighting each training example's gradient by $p_a(y\\mid x)\\propto p_s(y\\mid I(x))p_d(y\\mid x)$: when the biased classifier confidently predicts the true label, the gradient on that example is cancelled, and when the biased classifier is uninformative the gradient reduces to the MLE gradient. The paper argues this corrects the label shift $p(y\\mid b(x))\\neq q(y\\mid b(x))$ on known biased features, and reports that DRiFt improves non-entailment F1 on HANS and entailment F1 on the negation and overlap STRESS sets while degrading in-distribution accuracy only modestly for BERT, though more sharply for the weaker DA and ESIM models.","pith_inferences":["The same two-stage recipe should transfer to any task with a known shortcut, such as reading comprehension, visual question answering, or paraphrase detection, as long as one can build a cheap classifier on the suspected shortcut; the paper notes this direction but does not run those experiments.","Because DRiFt's definition of bias is whatever $I(x)$ captures, a practitioner using a feature set whose label association is actually stable across train and test would train away useful signal; checking the biased classifier's accuracy on a target distribution before applying DRiFt would be a cheap safeguard.","The gradient-reweighting view suggests a post-hoc variant: rather than retraining a full model, one could reweight the gradient contributions of a trained model's training examples, or use the residual objective as a fine-tuning loss; the paper does not explore either.","The label-shift framing leaves covariate shift unaddressed; the paper's STRESS results hint at this because distractor phrases shift the input distribution as well, so a natural extension is to combine residual fitting with covariate-shift robust training."],"forward_implications":["On HANS, DRiFt raises non-entailment F1 across all three base models and all three biased feature sets, with the handcrafted overlap features giving the largest non-entailment gains; MLE-trained models score near zero on non-entailment.","On the MNLI STRESS negation and overlap subsets, DRiFt improves entailment F1 in most configurations, though weaker models such as DA and ESIM lose F1 on one or both of the other classes.","Under synthetic label-leaking bias, DRiFt keeps accuracy approximately flat as the cheating rate rises to 0.9, whereas MLE accuracy falls by about 20 points; with an oracle biased classifier the method is equivalent to training only on the unbiased examples.","In-distribution accuracy is generally lower under DRiFt than under MLE, and the drop is largest when the biased features also encode useful semantics or the base model is weak; BERT retains accuracy close to MLE.","The paper's analysis implies that examples with zero gradient under the biased model contribute nothing to the debiased model, so when nearly all training examples are strongly biased, DRiFt alone cannot recover and additional data or augmentation is needed."],"supporting_citations":[{"why":"Shows hypothesis-only classifiers reach unexpectedly high accuracy, the evidence that shallow cues carry dataset bias and the origin of the RM baseline.","marker":"Gururangan et al. (2018)"},{"why":"Documents high pointwise mutual information between hypothesis words and labels, identifying the artifact DRiFt targets.","marker":"Poliak et al. (2018)"},{"why":"Contributes the HANS challenge set and diagnoses word-overlap heuristics, the main out-of-distribution evaluation.","marker":"McCoy et al. (2019)"},{"why":"Provides the STRESS test sets and the error analysis that motivates the HAND biased features.","marker":"Naik et al. (2018)"},{"why":"Supplies the SNLI training and test data used for the main experiments.","marker":"Bowman et al. (2015)"},{"why":"Supplies the MNLI training and matched development data.","marker":"Williams et al. (2017)"},{"why":"Supplies the label-shift formalism the paper uses to define dataset bias.","marker":"Lipton et al. (2018)"},{"why":"Concurrent work proposing an ensemble-based alternative to residual fitting, used as a comparison point.","marker":"Clark et al. (2019)"}],"fun_headline_variants":["Fit the residual: a simple recipe to unlearn NLI bias","Residual fitting debiases NLI without losing in-domain accuracy","DRiFt: train a biased model, then fit its residual","Weak biased model's residual lifts HANS and STRESS scores"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that examples the biased classifier predicts well are biased: the shallow features $I(x)$ must mark exactly the label association that changes between training and test, and they must carry no useful semantic signal that would be lost when the biased classifier is discarded at test time.","fun_headline_variants_meta":{"raw":{"variants":["Fit the residual: a simple recipe to unlearn NLI bias","Residual fitting debiases NLI without losing in-domain accuracy","DRiFt: train a biased model, then fit its residual","Weak biased model's residual lifts HANS and STRESS scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000518,"raw_usage":{"total_tokens":2555,"prompt_tokens":1033,"completion_tokens":1522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":1448}},"tokens_in":649,"tokens_out":1522,"duration_ms":12037,"temperature":1.0,"reasoning_tokens":1448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:35:18.736611+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the biased classifier's accuracy on both training and a held-out set drawn from the same distribution. If $I(x)$ remains highly predictive on the held-out set, there is no label shift, and DRiFt should underperform MLE by the amount of useful signal in $I(x)$; the paper's own in-distribution results for DRiFt-CBOW and DRiFt-HAND on DA and ESIM already show drops of more than twenty points, the predicted failure mode. A sharper test: construct a challenge set where the known bias association is preserved rather than reversed—if DRiFt still improves over MLE there, residual fitting is doing something other than removing label shift.","supporting_citations":[],"review_version":1}