{"id":"46587c48-f0ee-4f6c-a938-cc9d6b5f99a2","arxiv_id":"2504.12898","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An automatic debiasing framework rewrites instruction-tuning data to make identified biased features statistically independent of answers, improving zero-shot transfer on NLI and paraphrase tasks.","lead":"This paper proposes a framework that automatically removes dataset biases from instruction-tuning data by rewriting examples until a biased feature, such as a person's popularity, carries no information about the answer. The authors show small but consistent gains on transfer and challenge tests for Llama and Gemma models, with the largest improvements on natural language inference tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal attribution of the transfer gains is not established: no control condition isolates the rewriting step from the information-gain balancing objective, and the reported gains are small enough to be explained by generic augmentation or unintended semantic drift.","rationale":"The reader's weakest assumption already identified semantic preservation as the vulnerable point; my reading converges on the same locus but sharpens it: even full semantic preservation would not establish the causal role of Eq. 6 without a rewriting-only control. The paper's own evidence includes standard deviations over three runs and cross-model transfer, which is real support, but the central comparison is Vanilla SFT vs ICD (rewriting + balancing). The missing cell is 'rewriting without balancing'. The small gains on IMDB and Squadv2 further show the effect is not large, so confounding is a serious risk. This is not an internal inconsistency or an appeal to consensus; it is an identification problem in the experimental design. Because the missing control is straightforward and the paper's claims are otherwise coherent, the appropriate verdict remains conditional rather than rejection.","tokens_in":18318,"tokens_out":5355,"duration_ms":56021,"concrete_test":"On the NLI and PI tasks, build a control dataset by taking exactly the samples ICD rewrites and applying the same few-shot LLM rewriting procedure with the instruction modified to preserve the biased feature value (or to make a superficial synonym substitution), while keeping the same number of rewrites and the same acceptance/rejection loop. Fine-tune the same base models with the same hyperparameters and evaluate on SNLI, MRPC, HANS, and PAWS. If the control transfer deltas match ICD's deltas within one point, the gains are attributable to rewriting/augmentation rather than to satisfying IG(Y,B)=0; if the control is clearly worse, ICD's criterion is supported. As a secondary check, have an independent NLI model or annotators judge 100 original/rewritten pairs; a high rate of non-equivalence would indicate the intervention changed more than B.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline effect is attributed to Eq. 6 (N(yi,bk)=N(yj,bk)): balancing answer counts within each biased-feature value should remove the information gain of B for Y and thereby improve transfer. But the only manipulation that implements this is the LLM few-shot rewrite in Section 3.3, and the paper never isolates that manipulation. Appendix D checks only that the targeted feature changed (e.g., lexical overlap rate, negation-word presence, or popularity as judged by the biased LLM), not that other task-relevant semantics were preserved; footnote 3 concedes that semantics tightly tied to the biased feature must change. The reported transfer gains are small (e.g., SNLI +4.0/+2.2, MRPC +2.1/+2.0, HANS +2.9/+1.5, PAWS +3.1/+3.4 for Llama/Gemma; IMDB and Squadv2 gains below 0.5 points). These deltas are within the range that a generic paraphrase-augmentation effect or a shift in example difficulty could produce. Without a control that performs the same number of LLM rewrites while not enforcing Eq. 6, or a semantic-equivalence check on rewritten pairs, the experiments cannot distinguish 'removing B–Y association improves generalization' from 'LLM rewriting improves generalization'. The absence of a CAL-based control (the method whose features are used) makes the attribution even less secure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ICD, an information gain-guided causal intervention debiasing framework for LLMs. The method first fine-tunes a pretrained LLM on the original instruction-tuning data, then uses the CAL method to identify biased features, and rewrites a subset of the training data with the LLM itself so that, within each value of the biased feature, the answer counts are approximately balanced. This realizes the stated goal that the biased feature provides zero information gain for the answer (IG(Y,B)=0). The debiased dataset is then used for standard supervised fine-tuning. Experiments across NLI, paraphrase identification, sentiment analysis, and QA report improved zero-shot transfer on SNLI, MRPC, HANS, and PAWS, along with gains on MMLU, BBH, and TruthfulQA, while in-domain accuracy is roughly maintained. The paper also shows that data debiased by Llama3.1-8B can be used to fine-tune Gemma2-9B with similar benefits.","tokens_in":18588,"tokens_out":4506,"duration_ms":44008,"significance":"If the causal attribution is established, the paper would provide a principled, automatic debiasing method with a clean information-theoretic objective: a derivation from IG(Y,B)=0 to the count-balancing rule N(y_i,b_k)=N(y_j,b_k). The strengths include the algebraic clarity of the derivation, the use of standard deviations over three runs, the automatic nature of the pipeline, and the demonstration that a debiased dataset transfers across different base LLMs. The main empirical claim, however, is broader than the evidence: the transfer gains on sentiment and QA are within noise, and the experiments do not isolate the information-gain balancing from the generic effect of LLM rewriting. The paper is therefore of interest but requires additional controls and a more measured claim before it can be accepted.","major_comments":[{"comment":"The experiments do not include a control that isolates the causal intervention (Eq. 6 balancing) from the rewriting operation itself. The only manipulation is the few-shot LLM rewriting described in Section 3.3, and the paper compares it only against vanilla SFT and a prior-knowledge baseline. Without a condition that performs the same number of LLM rewrites while not enforcing the information-gain goal (e.g., random paraphrasing or rewriting to balance a non-identified feature), the improvements on SNLI, MRPC, HANS, and PAWS could be attributed to generic data augmentation or to a shift in example difficulty rather than to the removal of the B-Y association. This is load-bearing because the paper's central claim is specifically about the information gain-guided criterion.","section":"Section 3.3 and Section 5.1"},{"comment":"There is no semantic-equivalence check for the rewritten examples. Footnote 3 concedes that semantics tightly tied to the biased feature must change (e.g., substituting the person in the answer), and Appendix D checks only that the targeted feature changed (lexical overlap rate, negation presence, or popularity as judged by the biased LLM). The paper should verify that the rewritten samples retain their gold answer and remain plausible, for example through human evaluation or a second-model consistency check, and should report the rate at which rewriting had to change the answer. Without such a check, the observed transfer gains may reflect changes in content difficulty or new artifacts rather than debiasing.","section":"Section 3.3, footnote 3, and Appendix D"},{"comment":"The abstract and Section 5.1 (point 2) claim that ICD achieves consistent improvement across all transfer test sets, but the SA and QA transfer gains are negligible: Llama IMDB +0.1, Gemma IMDB +0.3, Llama Squadv2 +0.4, Gemma Squadv2 +0.3, several of which are within one standard deviation of the vanilla baseline. The claim of improved generalizability should be tempered to the NLI and PI tasks (SNLI, MRPC, HANS, PAWS), or significance testing should be provided to support the cross-task claim.","section":"Table 1"},{"comment":"The comparison with Razor is not a fair automatic-debiasing comparison for QA, because Appendix E states that Razor cannot be applied to QA and is therefore only applied to MNLI, QQP, and SST2; the Razor rows for QA are effectively identical to vanilla. The text should explicitly state this limitation when claiming that ICD outperforms Razor on transfer test sets. Additionally, since the proposed method relies on CAL for biased-feature identification, the paper should include a CAL-based control (e.g., CAL's in-context learning debiasing or a CAL-identified feature without the count-balancing rewrite) to separate the contribution of the new information-gain objective from the contribution of feature identification.","section":"Section 5.1 and Appendix E"}],"minor_comments":[{"comment":"The step from Eq. (4) to Eq. (5) relies on the additional assumption that P(Y) is uniform (\"there is no prior knowledge about Y\"); this is a modeling choice, not a consequence of IG(Y,B)=0, and should be stated as an assumption.","section":"Section 3.2, Eq. (5)"},{"comment":"The text refers to \"IDB\" in the discussion of general ability; this should be \"ICD\".","section":"Section 5.1"},{"comment":"There are typos in the lexical overlap classes: \"low lexical overelap\" appears twice, and the third class should read \"high lexical overlap\" instead of \"low lexical overelap (bigger than 0.6)\".","section":"Appendix C"},{"comment":"The details of the few-shot rewriting prompt and the retry mechanism are deferred to the supplemental material; including the prompt (or a pointer to it) in an appendix would improve reproducibility.","section":"Section 3.3"},{"comment":"For popularity bias, checking the feature using the biased LLM itself is circular; the paper should validate this popularity judgment against human annotation or another reference.","section":"Appendix D"},{"comment":"The claim of \"maintaining performance on the in-domain datasets\" should acknowledge the small drops on MNLI (Llama 84.3 to 83.8, Gemma 76.2 to 75.7) and QQP (Llama 86.3 to 85.7, Gemma 87.0 to 86.5), even if these are within the reported standard deviations.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper shares authors with the CAL method (Sun et al., 2024) and uses CAL as the automatic feature identifier; this is legitimate but makes the absence of a CAL-based control more conspicuous. The central idea is promising and the derivation is clean, but the empirical support for the cross-task claim and the attribution to information-gain balancing needs strengthening. The scope of the paper is suitable for the venue, and the issues raised are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. The paper has a clean, formal idea: debias an instruction-tuning set by making each detected biased feature statistically independent of the answer, operationalized as IG(Y,B)=0 and then as a count-balancing rule. The derivation is direct and correct. The empirical setup is also honest in one important way: the comparisons are against external held-out transfer and challenge sets, so the results are not fitted to those benchmarks. And the dataset-level debiasing is reusable: debias with Llama, fine-tune with Gemma, and the gains persist. That is a genuine practical asset.\n\nThe weak point is exactly what your stress-test note says. The only manipulation implementing the balancing rule is LLM few-shot rewriting, and the paper never runs a control that does the same rewriting without the information-gain constraint. Without that, the transfer gains on NLI and paraphrase sets could come from generic data augmentation or unintended semantic drift, not from removing the B–Y association. The negligible gains on IMDB and Squadv2 reinforce that the 'across different tasks' claim in the abstract is too strong. There is also no comparison with CAL as a debiasing method, even though CAL is the pre-existing automatic approach the authors build on; using it only as a feature detector leaves the 'better than automatic debiasing' claim unsupported.\n\nSome smaller issues: the linear programming step for choosing intervention direction and counts is underspecified; the checklist says code/data are in supplemental material, but the public preprint does not give access; and the semantic-preservation check is limited to verifying that the targeted feature changed. The limitations section is honest — standard answers only, one biased feature per task — but that honesty also narrows the scope.\n\nI would not dismiss the paper. The framework is coherent, the math is fine, and the positive results on NLI and paraphrase are consistent across two models. The missing control and baseline are fixable with a revision. My recommendation for review: send it out, but ask the authors to add a rewriting-only control, a CAL debiasing comparison, a semantic-equivalence check, and release the data/code and LP details. A reader working on data-centric debiasing will get value from the formalization even if this version does not settle the causal claim.","headline":"A coherent debiasing recipe with a clean criterion, but the paper never isolates the rewriting step from the balancing objective, so the 'across tasks' claim is only partly supported.","tokens_in":19147,"tokens_out":3509,"would_cite":false,"duration_ms":38693,"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":"An LLM can be debiased by rewriting its training data until the biased feature yields zero information gain about the answer, which improves zero-shot transfer across tasks.","keywords":["dataset bias","large language models","information gain","causal intervention","automatic debiasing","instruction tuning","zero-shot transfer","data rewriting"],"falsifier":"A control experiment that rewrites the same number of samples without targeting any biased feature; if transfer gains match ICD's, the improvement is not due to bias removal. A second check is human annotation of rewritten samples for whether the gold answer still holds; if a substantial fraction change meaning, the debiasing explanation is unsupported.","tokens_in":18078,"feed_emoji":"⚖️","tokens_out":6126,"duration_ms":53652,"temperature":0.7,"pith_summary":"The paper argues that large language models trained on instruction-tuning data learn dataset biases—correlations between non-causal features and answers—and that these biases hurt generalization to new datasets of the same task. It claims that a dataset is debiased with respect to a biased feature exactly when that feature contributes zero information gain about the answer, and that this condition can be met by rewriting examples so that each value of the biased feature appears with each answer equally often. The proposed ICD framework identifies biased features automatically, rewrites a minimal set of examples using the LLM itself, and then fine-tunes on the rebalanced data. The reported result is consistent improvement on zero-shot transfer sets (SNLI, MRPC, IMDB, Squadv2) and challenge sets (HANS, PAWS) with no loss on in-domain accuracy. If correct, the work turns debiasing from a per-task manual exercise into a data-level goal that any LLM can apply, and debiased datasets transfer to other models.","feed_headline":"LLM transfer scores rise when data rewrite kills biased cues","feed_subtitle":"Fine-tuning on rebalanced data lifts zero-shot transfer to SNLI, MRPC, IMDB, and Squadv2.","key_machinery":"The load-bearing object is the information gain-guided data debiasing goal: for any biased feature B, the dataset must satisfy IG(Y,B)=0, equivalently N(y_i,b_k)=N(y_j,b_k) for every feature value and answer pair. To reach this state, the framework first identifies biased features automatically with a causal-guided active learning method, then performs a causal intervention do(B=b_k) by using the LLM to rewrite samples via few-shot prompting, and selects which samples to rewrite by solving a linear program whose objective is minimal rewriting subject to the count-balance constraint. The rewriting step is what carries the empirical weight: it alters the joint distribution of biased feature and answer without, by assumption, changing the semantics that determine the correct answer.","core_discovery":"The central claim is that predictive but non-causal features ('biased features') in instruction-tuning datasets can be eliminated by enforcing the information-theoretic condition IG(Y,B)=0—the biased feature B provides no information about the answer Y. The paper derives the operational form P(Y|B)=P(Y), realizes it as the count-balancing requirement N(y_i,b_k)=N(y_j,b_k), and achieves this by a causal intervention do(B=b_k) that rewrites a fraction of examples with the LLM itself, solving a linear program to minimize the number of rewrites. Trained on the rewritten data, the LLM no longer relies on the biased feature and transfers better to held-out datasets of the same task, while keeping in-domain performance and even improving general benchmarks such as TruthfulQA. The authors frame this as the first strict criterion for an unbiased instruction-tuning dataset.","pith_inferences":["If the count-balancing criterion is the right target, the same principle could be applied to any biased feature in any dataset, turning debiasing into a data-curation step that does not require knowing the downstream task.","Because the paper reports no check that rewritten examples remain semantically equivalent, some of the transfer gains may come from data augmentation rather than bias removal; a random-rewriting control would separate these explanations.","The framework inherits the limitation that bias identification requires a task with standard answers, so open-ended or generative tasks still lack an automatic debiasing route within this approach.","The linear-programming formulation suggests a general principle: among all datasets satisfying the independence condition, prefer the one closest to the original; this framing could support future work on minimal-intervention debiasing."],"forward_implications":["ICD improves zero-shot transfer accuracy over vanilla fine-tuning on SNLI, MRPC, IMDB, and Squadv2 while preserving in-domain accuracy on MNLI, QQP, SST2, Squadv1, and TriviaQA.","On challenge sets HANS and PAWS, which remove dataset biases, ICD outperforms both vanilla fine-tuning and the prior-knowledge baseline Razor.","Datasets debiased by one LLM (Llama3.1-8B) can be used to fine-tune a different LLM (Gemma2-9B) with the same debiasing benefit.","General ability benchmarks MMLU, BBH, and TruthfulQA do not decline after debiasing; TruthfulQA and BBH improve.","The IG=0 condition provides a formal, task-agnostic definition of what an unbiased instruction-tuning dataset must satisfy."],"supporting_citations":[{"why":"Supplies the causal-guided active learning method that automatically identifies interpretable biased features, the entry point of the ICD pipeline.","marker":"[35]"},{"why":"Provides the information-theoretic definitions of entropy and information gain used to state the debiasing goal IG(Y,B)=0.","marker":"[4]"},{"why":"Supplies the do-operator and causal intervention framework used to justify rewriting data to change the joint distribution of bias and answer.","marker":"[25]"},{"why":"The prior-knowledge debiasing baseline (Razor) that ICD must beat on transfer and challenge sets.","marker":"[47]"},{"why":"Provides the HANS challenge set used to test whether models stop relying on lexical-overlap bias.","marker":"[20]"},{"why":"Provides the PAWS challenge set used to test paraphrase-identification bias.","marker":"[48]"},{"why":"Llama3.1-8B, the model used for both debiasing the dataset and fine-tuning in the main experiments.","marker":"[8]"},{"why":"Gemma2-9B, the second model fine-tuned on the debiased dataset to test cross-model transferability of the debiased data.","marker":"[38]"}],"fun_headline_variants":["Debias LLMs by rebalancing data via causal intervention","Zero information gain: causal rewrite to debias LLMs","Causal intervention to make LLMs drop spurious cues","Information-guided data rewriting for unbiased LLMs","Autonomous LLM debiasing via information-guided rewrites"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that few-shot rewriting changes only the targeted biased feature and leaves every other property of the sample that determines the answer unchanged.","fun_headline_variants_meta":{"raw":{"variants":["Debias LLMs by rebalancing data via causal intervention","Zero information gain: causal rewrite to debias LLMs","Causal intervention to make LLMs drop spurious cues","Information-guided data rewriting for unbiased LLMs","Autonomous LLM debiasing via information-guided rewrites"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001121,"raw_usage":{"total_tokens":4656,"prompt_tokens":931,"completion_tokens":3725,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":3645}},"tokens_in":547,"tokens_out":3725,"duration_ms":28048,"temperature":1.0,"reasoning_tokens":3645,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:19:53.295601+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A control experiment that rewrites the same number of samples without targeting any biased feature; if transfer gains match ICD's, the improvement is not due to bias removal. A second check is human annotation of rewritten samples for whether the gold answer still holds; if a substantial fraction change meaning, the debiasing explanation is unsupported.","supporting_citations":[{"cited_title":"Causal-guided active learning for debiasing large language models","cited_arxiv_id":null,"evidence_quote":"Supplies the causal-guided active learning method that automatically identifies interpretable biased features, the entry point of the ICD pipeline."},{"cited_title":"Information Theory: Coding Theorems for Discrete Memoryless Systems","cited_arxiv_id":null,"evidence_quote":"Provides the information-theoretic definitions of entropy and information gain used to state the debiasing goal IG(Y,B)=0."},{"cited_title":"Razor: Sharpening knowledge by cutting bias with unsupervised text rewriting","cited_arxiv_id":null,"evidence_quote":"The prior-knowledge debiasing baseline (Razor) that ICD must beat on transfer and challenge sets."},{"cited_title":"Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference","cited_arxiv_id":null,"evidence_quote":"Provides the HANS challenge set used to test whether models stop relying on lexical-overlap bias."},{"cited_title":"Paws: Paraphrase adversaries from word scrambling","cited_arxiv_id":null,"evidence_quote":"Provides the PAWS challenge set used to test paraphrase-identification bias."}],"review_version":1}