{"id":"5b519127-6b3e-4afa-88c6-32f50b41b9fc","arxiv_id":"2412.14050","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"English-only debiasing and detoxification partly transfer to other languages, but transfer is weaker for low-resource languages and typically degrades non-English generation quality.","lead":"Researchers finetuned multilingual AI models in English to reduce bias and toxicity, then tested whether the fixes transfer to 19 other languages. They found that transfer is real but costly: models often lose fluency or switch languages, so English-only safety training is not a free lunch.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'predicted by pretraining data' claim rests on an untested proxy and is contradicted by the paper's own known-mixture results for Aya models; it needs out-of-sample testing before it can be called a prediction.","rationale":"I read the paper as an empirical study with two layers: a robust descriptive finding (English-only debiasing and detoxification transfer unevenly and often degrade non-English generation) and a more fragile explanatory claim (transfer is predicted by the amount of pretraining data per language). The descriptive layer is well supported: the benchmark suite is large, the models and languages are varied, the code is public, and the per-language patterns in Figure 5 are concrete and internally consistent. The explanatory layer is where the argument is least secure. The reader identified the Common Crawl proxy as the weakest assumption, and I agree; my analysis sharpens that concern with internal evidence. Table 11 is the paper's own test using known language shares for the Aya models, and it yields no significant correlations with transfer. That is not an external critique of the proxy; it is a contradiction inside the paper's evidence. In addition, the abstract's word 'predicted' overstates what an in-sample Spearman correlation can establish, especially with 8 to 17 languages per model and no held-out evaluation. A leave-one-language-out prediction test would settle whether the proxy has any genuine predictive value. I do not think this concern warrants rejection: the empirical transfer results and the practical recommendation to mitigate in the evaluation language remain credible, and the authors already hedge in places with 'often' and 'to some extent.' The appropriate outcome is the reader's CONDITIONAL verdict, with the predictive claim reframed as a correlation and the proxy assumption either validated or dropped. Since my concern reinforces the reader's verdict rather than moving it, I set verdict_should_be to UNCHANGED.","tokens_in":30672,"tokens_out":8405,"duration_ms":81370,"concrete_test":"For the RTP-LX detoxification transfer scores, run a leave-one-language-out test for each model: fit a rank regression of transfer magnitude on Common Crawl language share using 16 languages, predict the held-out language's transfer rank, and average the Spearman correlation of predicted versus actual ranks across folds. Separately, recompute the Table 10 DATA correlations for Aya 23 and Aya Expanse using exact Aya Collection language proportions on the same RTP-LX language set. If out-of-sample prediction is at chance, or if the known-mixture correlations remain non-significant, the headline should be downgraded from 'predicted by pretraining data' to 'correlated with a Common Crawl proxy in some models.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central explanatory claim—that cross-lingual transfer 'can be predicted by the amount of data in a given language present in the model's pretraining data'—rests entirely on the correlation analysis in Table 10 and Appendix J. Three problems make this the weakest load-bearing link. First, the predictor is a proxy: Common Crawl language percentages are assumed to approximate the actual pretraining mixture of Gemma 2, Llama 3.1, and Mistral, none of which disclose per-language pretraining data; Appendix J states this assumption explicitly. Second, the only place where the paper has a known language mixture, the Aya Collection, the correlations with transfer are small and non-significant (Table 11: -0.19, -0.35, -0.10, 0.08). If pretraining data share were the driver, the strongest signal should appear where the mixture is actually known; it does not. Third, the supporting statistics are in-sample Spearman correlations computed per model over small language samples (8 languages for CrowS-Pairs, 17 for RTP-LX), and the word 'predicted' in the abstract implies out-of-sample validity that is never tested. The transfer asymmetry itself is well evidenced, but the explanatory claim that pretraining data share predicts it is not established; the observed correlations could be confounded by collinearity with subword overlap, typological similarity, or artifacts of English-centric translated benchmarks.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether English-only debiasing and detoxification finetuning transfers to non-English languages in generative LLMs, and whether such transfer comes at the cost of language generation ability. The authors finetune four instruction-tuned models (Aya 23, Aya Expanse, Gemma 2 9B IT, Llama 3.1 Instruct) plus their base variants on four English datasets (Panda, BiasDPO, Jigsaw, DetoxDPO) using SFT and DPO. They measure bias with CrowS-Pairs, StereoSet, and MBBQ; toxicity with RTP-LX; and language generation ability via language consistency, fluency, diversity, and Global-MMLU question answering. The main empirical findings are that SFT on Panda reduces bias, DPO on DetoxDPO reduces toxicity, transfer to non-English languages is uneven and generally weaker for lower-resource languages, and successful transfer is often accompanied by degraded language generation in the target language. The paper also reports correlations suggesting that transfer is predicted by a language's share of Common Crawl data.","tokens_in":30946,"tokens_out":3830,"duration_ms":33827,"significance":"If the empirical pattern holds, the paper makes a useful contribution to the practical question of when English-only safety tuning is sufficient for multilingual deployment. Its strengths include the broad model and language coverage, the systematic comparison of SFT and DPO, the joint evaluation of harmfulness and generation quality, the use of established external benchmarks, and the public release of code. The learning trajectories in Appendix F and the qualitative analysis are also valuable. However, the paper's central predictive claim about pretraining data share is only supported by in-sample correlations with a Common Crawl proxy and is contradicted by the known-mixture results for the Aya models; this weak link needs to be addressed before the conclusion can be accepted as stated.","major_comments":[{"comment":"The only place where the paper has access to a known pretraining mixture is the Aya models against the Aya Collection, and there the correlations between transfer and language data percentage are small and non-significant (-0.19, -0.35, -0.10, 0.08). If pretraining data share were the driver of transfer, the strongest signal should appear where the mixture is actually known, not only in the Common Crawl proxy analysis. The manuscript does not discuss this discrepancy, yet the abstract and conclusion assert that transfer 'can be predicted' by pretraining data share. Please either provide out-of-sample evidence from a model with disclosed per-language pretraining proportions, or explicitly temper the claim to state that transfer correlates with a Common Crawl-based proxy for resource availability.","section":"Section 5, Appendix J, Table 11"},{"comment":"The word 'predicted' in the abstract and conclusion implies out-of-sample validity, but the supporting statistics are in-sample Spearman correlations computed on small language samples (8 languages for CrowS-Pairs, 17 for RTP-LX), with no held-out evaluation or cross-validation. A correlation observed on the same languages used to estimate it does not establish predictive power. Please replace 'predicted' with 'correlates with' or add a proper out-of-sample test, such as leave-one-language-out prediction of transfer magnitude.","section":"Abstract, Section 5, Section 6"},{"comment":"The analysis does not control for plausible confounds. Common Crawl language share is likely collinear with subword overlap, bilingual sentence similarity, and general linguistic resource availability. Indeed, in Table 10 several predictors attain comparable or larger magnitudes than DATA for some models (e.g., RTP-LX Gemma 2 9B base: SIM -0.67 vs DATA -0.60; CrowS-Pairs Gemma 2 9B IT: OVER -0.64 vs DATA 0.40). To support the claim that pretraining data share is the best predictor, please report partial correlations or a regression with nested model comparisons that show DATA adds explanatory power beyond subword overlap and similarity measures. Additionally, the table tests many model-predictor-task combinations without any multiple-comparison correction; only a handful of correlations reach p<0.01, and some are positive (e.g., Aya Expanse RTP-LX SIM 0.47), which should be discussed.","section":"Section 5, Table 10"},{"comment":"The conclusions about toxicity transfer are inconsistent across model families. For Aya 23 and Aya Expanse, the correlations between transfer and Common Crawl data share are -0.12 and 0.04, respectively, and for Aya Expanse the correlation with bilingual sentence similarity is positive (0.47), opposing the direction reported by Li et al. (2024). The paper notes that these correlations are not significant 'for all except the Aya models,' but the abstract and conclusion still generalize the pretraining-data-share explanation to all models. Please either reconcile these inconsistencies explicitly or restrict the claim to the models for which the correlation actually holds.","section":"Appendix J, Table 10, RTP-LX rows"}],"minor_comments":[{"comment":"There is a typo: 'significant modererate correlation' should read 'significant moderate correlation.'","section":"Section 5"},{"comment":"Several numerical entries contain stray spaces (e.g., '51 .89', '0 .059', '57 .37'), likely from formatting or extraction artifacts; please standardize the table formatting.","section":"Tables 2, 9, 12, 13"},{"comment":"The choice of Common Crawl week 30 of 2024 is clearly motivated by model release timing, but the analysis would be more robust if the authors reported sensitivity to the snapshot, or at least acknowledged that different snapshot choices could change the correlation values.","section":"Appendix J"},{"comment":"The qualitative analysis is described as conducted by 'one author and an annotator external to the project.' For reproducibility, please specify the annotation instructions and whether any agreement metric was computed, or explicitly label this part as illustrative rather than systematic.","section":"Section 4.2, Qualitative Analysis"},{"comment":"The 'Mean absolute change in score' panels combine metrics with very different scales (e.g., diversity as a percentage, fluency as negative perplexity, language consistency as a percentage). Consider presenting these on separate subplots or with standardized scales to avoid visual distortions.","section":"Figures 2c and 3b"},{"comment":"The limitations paragraph mentions the black-box nature of Perspective API, but it does not mention that the RTP-LX evaluations only consider completions detected as being in the prompt language; although Appendix E shows minimal differences for the 'all completions' comparison, this filtering decision and its implications for language-confused generations deserve a sentence in the main text.","section":"Limitations"}],"recommendation":"major_revision","confidential_remarks":"The paper self-cites MBBQ and the authors' own subword-overlap method, but the core transfer results are demonstrated on several independent benchmarks, so the self-citation pattern is not a concern. The main issue for a revision is the overstatement of the pretraining-data-share predictor: the manuscript has the data to present this as a resource-proxy correlation, but not as an established predictive mechanism. The manuscript is a good fit for the journal's scope, and the empirical contribution is substantial; the revision should focus on the correlation-to-prediction gap and the confound analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading: it gives the most thorough empirical picture I know of what happens when you finetune multilingual LLMs in English for debiasing and detoxification. Transfer is uneven, often comes with a real drop in non-English generation quality, and the SFT-vs-DPO contrast is informative. The experiments are careful: seven models, twenty languages, base and instruction-tuned variants, confidence intervals, a control that SFTs the DPO preferred completions, learning trajectories, and a qualitative pass. Code is released. The finding that SFT on Jigsaw or DetoxDPO increases toxicity is counterintuitive and important, and the DPO diversity gains contradict prior claims that DPO reduces diversity. The central practical message—do not rely on English-only safety tuning for low-resource languages—is well supported.\n\nThe soft spot is the explanatory claim. The abstract says transfer 'can be predicted by the amount of data in a given language present in the model's pretraining data,' but the evidence is in-sample Spearman correlations with Common Crawl percentages, over 8 to 17 languages, and the paper's own known-mixture results for Aya models (Table 11) show small, non-significant correlations. That is a proxy, not a test. The transfer asymmetry is real; the pretraining-share predictor is not established. Also, the conclusion that 'both methods come at the cost of a decrease in language modeling ability' is too strong—their DPO runs often improved diversity. The summary should say 'often' or separate fluency and consistency from diversity.\n\nMinor point, acknowledged by the authors: the bias benchmarks are translations of English-centric datasets, so the cross-lingual story is about transferred English stereotypes, not culturally specific bias.\n\nWho gets value: anyone working on multilingual safety, preference tuning, or deployment of multilingual LLMs. The empirical core is solid and reproducible. It deserves a serious referee; the revision should focus on reframing the predictive claim as a correlation hypothesis and softening the cost conclusion, not on redoing the experiments.","headline":"A solid empirical study of cross-lingual transfer of debiasing/detoxification; the transfer-asymmetry finding holds, but the 'predicted by pretraining data' claim is oversold.","tokens_in":31469,"tokens_out":2211,"would_cite":true,"duration_ms":20299,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Finetuning an LLM in English to reduce bias and toxicity does transfer to other languages, but only as far as the model's pretraining data reaches, and every successful transfer measurably weakens the model's generation in that language.","keywords":["cross-lingual transfer","debiasing","detoxification","multilingual LLMs","direct preference optimization","supervised finetuning","language generation ability","pretraining data share"],"falsifier":"Directly measuring each model's actual pretraining data mix (e.g., through a data-mixture inference attack on the tokenizer) and recomputing the correlation between language share and transfer would settle the predictive claim: if the correlation disappears under true data shares, the paper's predictor is an artifact of the proxy. A second falsifier is a controlled experiment finetuning two identically architected models pretrained with deliberately different language shares; if transfer does not track the manipulated shares, the causal claim fails.","tokens_in":30469,"feed_emoji":"🌐","tokens_out":13213,"duration_ms":92970,"temperature":0.7,"pith_summary":"Prompted in languages other than English, large language models tend to show stronger social bias and toxicity, and the usual fix—safety finetuning in English—is assumed to carry over. This paper tests that assumption by finetuning seven multilingual LLMs in English with four datasets (two for debiasing, two for detoxification) and evaluating them in up to twenty languages on bias, toxicity, and generation quality. The authors find that transfer is real but uneven: the amount of bias and toxicity mitigation a language receives is predicted, for most models, by that language's share of the model's pretraining data, with low-resource languages gaining the least. They also find that both debiasing and detoxification, whenever they succeed, come at the cost of the model's ability to generate fluent, diverse, correctly language-matched text in that language. The paper concludes that English-only safety tuning is not a sufficient substitute for mitigation done in the target language.","feed_headline":"English-only safety tuning transfers only to well-resourced languages","feed_subtitle":"Low-resource languages gain little, and every successful transfer hurts fluency and correct-language output.","key_machinery":"The machinery is a matched comparison of two finetuning families: supervised finetuning on curated non-harmful sentences (Panda for bias, Jigsaw and DetoxDPO-preferred for toxicity) and direct preference optimization (BiasDPO for bias, DetoxDPO for toxicity), applied in English to seven instruction-tuned LLMs and then evaluated per language. The load-bearing comparison is the correlation between the per-language change in bias or toxicity and the percentage of that language in Common Crawl, a proxy for the model's pretraining data share; this correlation carries the paper's predictive claim. The evaluation stack—CrowS-Pairs, StereoSet, and MBBQ for bias, RTP-LX for toxicity, and language consistency, mT5 fluency, unigram diversity, and Global-MMLU for generation ability—is what converts the finetuning comparison into a cross-lingual claim.","core_discovery":"The paper's central finding is that English-only safety finetuning is a partial, not free, solution. Supervised finetuning (SFT) on the Panda corpus reliably reduces stereotypical bias scores in non-English languages across the CrowS-Pairs, StereoSet, and MBBQ benchmarks, while direct preference optimization (DPO) on BiasDPO works only for the two Aya models. For toxicity, the situation is sharper: SFT on either Jigsaw or DetoxDPO increases toxicity, and only DPO on DetoxDPO reliably reduces it in both English and non-English languages. When mitigation does transfer, the amount of transfer across languages correlates with the percentage of the language in Common Crawl data, used as a proxy for pretraining data share, and not consistently with typological similarity or subword overlap. And in every case of successful transfer, at least one language-generation metric degrades—language consistency, fluency, diversity, or question-answering accuracy—so the authors argue that relying on transfer from English is especially risky for lower-resource languages.","pith_inferences":["Editorial extension: a deployer could use per-language pretraining share, estimated by tokenizer-based mixture inference rather than Common Crawl proxies, to decide in advance whether English-only safety tuning will transfer to a given language.","Editorial extension: the finding that SFT on 'non-toxic' comments increases toxicity suggests that sentiment-level data is not a safe detoxification signal, and that preference-based objectives should be preferred whenever generation quality matters.","Editorial extension: the paper's per-language correlation analysis invites a causal test—finetuning two models with identical architecture but deliberately different per-language data shares would confirm whether pretraining share actually drives transfer rather than merely correlating with it."],"forward_implications":["Supervised finetuning on Panda debiases models in non-English languages for most models, but DPO on BiasDPO is effective only for the two Aya models, showing that the winning debiasing recipe depends on dataset size and method.","SFT on Jigsaw or DetoxDPO increases toxicity, while DPO on DetoxDPO reduces toxicity in all models and both English and non-English languages, making DPO the only effective detoxification method tested.","The degree of toxicity-mitigation transfer correlates negatively with language share in Common Crawl for all models except the Aya models, meaning languages with more pretraining data benefit more from transfer.","Every successful debiasing or detoxification transfer is paired with a decrease in at least one language-generation metric, and SFT often reduces language consistency, diversity, and question-answering accuracy.","The authors recommend performing bias and toxicity mitigation in the evaluation language rather than relying on English transfer, particularly for lower-resource languages underrepresented in pretraining data."],"supporting_citations":[{"why":"Supplies the Panda corpus used for SFT debiasing; finetuning on it produces the paper's consistent bias reductions.","marker":"(Qian et al., 2022)"},{"why":"Supplies BiasDPO, the small DPO dataset whose debiasing works only for Aya models, anchoring the method-dependence finding.","marker":"(Allam, 2024)"},{"why":"Provides the Jigsaw civil-comments corpus used as SFT detoxification data, which the paper shows increases toxicity.","marker":"(cjadams et al., 2019)"},{"why":"Supplies DetoxDPO, the DPO dataset that yields the only reliable toxicity reduction in the study.","marker":"(Lee et al., 2024)"},{"why":"Prior work showing DPO toxicity mitigation transfers and proposing bilingual sentence similarity as a predictor; this paper extends and partially replicates that result against pretraining-data share.","marker":"(Li et al., 2024)"},{"why":"Provides RTP-LX, the multilingual toxicity benchmark used to measure whether detoxification transfers.","marker":"(de Wynter et al., 2025)"},{"why":"Provides CrowS-Pairs, one of the bias benchmarks used to measure debiasing transfer across languages.","marker":"(Nangia et al., 2020)"},{"why":"Provides MBBQ, the multilingual question-answering bias benchmark used to test downstream transfer.","marker":"(Neplenbroek et al., 2024)"},{"why":"Supplies the language-confusion pipeline used to measure language consistency, the key generation-ability metric that degrades with transfer.","marker":"(Marchisio et al., 2024)"}],"fun_headline_variants":["Safety tuning in English helps few languages, hurts many","Debiasing transfers, but fluency drops in low-resource tongues","Only DPO cuts toxicity; SFT makes it worse in LLMs","Cross-lingual safety transfer: partial, expensive, and rare","English debiasing: transfer only for pretraining-rich languages"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Common Crawl language percentages accurately proxy each model's actual pretraining data mix, which is unverified for the Gemma, Llama, and Mistral models; if this proxy is wrong, the predictive claim about transfer loses its main support.","fun_headline_variants_meta":{"raw":{"variants":["Safety tuning in English helps few languages, hurts many","Debiasing transfers, but fluency drops in low-resource tongues","Only DPO cuts toxicity; SFT makes it worse in LLMs","Cross-lingual safety transfer: partial, expensive, and rare","English debiasing: transfer only for pretraining-rich languages"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1543,"prompt_tokens":944,"completion_tokens":599,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":511}},"tokens_in":560,"tokens_out":599,"duration_ms":6115,"temperature":1.0,"reasoning_tokens":511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:31:24.625479+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Directly measuring each model's actual pretraining data mix (e.g., through a data-mixture inference attack on the tokenizer) and recomputing the correlation between language share and transfer would settle the predictive claim: if the correlation disappears under true data shares, the paper's predictor is an artifact of the proxy. A second falsifier is a controlled experiment finetuning two identically architected models pretrained with deliberately different language shares; if transfer does not track the manipulated shares, the causal claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies BiasDPO, the small DPO dataset whose debiasing works only for Aya models, anchoring the method-dependence finding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides RTP-LX, the multilingual toxicity benchmark used to measure whether detoxification transfers."}],"review_version":1}