{"id":"eb1fd53a-ba0e-4178-beb9-e8b68f0d4ab7","arxiv_id":"2505.02252","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Country and language personas degrade LLM hate speech detection F1 scores, and a custom reweighted fine-tuning loss reduces the degradation for Llama and Nemo, but less for Phi.","lead":"Giving a hate speech detection LLM a one-line persona such as 'someone from Saudi Arabia' changes its verdicts, sharply increasing false negatives for some countries. The authors fine-tune models with a reweighted loss and report that it partly restores accuracy under personalized prompts.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The debias-tuning contribution is not yet supported: Eq. 2 is a reweighting/scaling of cross-entropy, not a consistency penalty, and no α=0 control rules out ordinary fine-tuning on CountryHate.","rationale":"The reader's weakest-assumption pick, label noise, is genuine and is supported by the paper's own error analysis (74% of sampled false negatives are 'doubtful labels'). But the core bias observation—same text, different country/language persona, different verdict—does not require gold labels to demonstrate that context changes behaviour; labels are needed to call those changes bias, and noisy labels weaken that interpretation. The debias-tuning claim is more directly load-bearing and more easily falsifiable: the custom loss in Eq. 2 does not implement a consistency penalty, and no control separates the proposed mechanism from ordinary fine-tuning. If the code is exactly as written, the headline claims about 'debias tuning' are not supported; if the code differs, the paper needs to correct the description. In either case the conditional verdict stands, with the conditions including a corrected loss description, the α value, and an α=0 control. This is why I set agreement_with_reader to partial: the label-noise point is valid, but I would put the loss/control issue first.","tokens_in":17808,"tokens_out":9895,"duration_ms":121942,"concrete_test":"Open the GitHub repository (palomapiot/geographic-bias) and check the custom loss implementation against Eq. 2. If it matches, run the same debias-tuning pipeline on the same four countries with α fixed to 0 (no penalty), keeping all other hyperparameters fixed. If the α=0 model reproduces the Table 6 and Figure 2 gains, the improvements are caused by standard fine-tuning on CountryHate, not by the consistency penalty; if the α=0 model is substantially worse, the penalty is doing real work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's main contribution is the debias-tuning mitigation (Section 'Custom loss'), but the loss as written cannot support the claimed mechanism. Equations 1-2 give L_loss = L_avg_class + α·L_avg_class, with α>0 only when the context prediction is wrong/invalid while the non-context prediction matches the gold label. This is not a consistency penalty: no term compares the with-context and without-context predictions or logits, and the equation is just a (possibly per-example) reweighting of the standard cross-entropy loss; if α is a fixed scalar it is global loss scaling. The reported F1/FNR gains over the base models (Table 6, Figure 2) could therefore come entirely from standard supervised fine-tuning on 19,306 CountryHate training instances, and no control fine-tune without the penalty is reported. The value of α is also omitted, so even the reweighting is not reproducible from the text. Because the abstract and conclusions attribute the gains specifically to penalising inconsistencies, the central mitigation claim is unsupported as written unless the released code implements a different loss.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper investigates whether country and language personalization, simulated through persona prompts, introduces bias in hate speech classification performed by five open-source LLMs (Llama 3.1, Nemo, Gemma, DeepSeek, Phi 4). Using 'CountryHate,' a subset of the MetaHate dataset augmented with twelve countries selected from low-scoring human-rights, LGBTQ+, and women's-rights indices, the authors prompt models with and without country personas, in English and in the personas' official languages. They report that country context generally lowers F1 and raises false-negative rates, with the largest shifts for Llama 3.1 (FNR rising from 32.37% to over 71%). To mitigate this, they propose 'debias tuning,' a fine-tuning procedure with a custom loss intended to penalize inconsistent classifications with and without context, applied to Llama 3.1, Nemo, and Phi 4. Results in Table 6 and Figure 2 show improved F1 scores and reduced country-specific FNRs for the debiased models, with the Llama variants showing the largest gains. The paper concludes that memory personalization features can introduce geographic and language bias in hate speech detection, and that debias tuning significantly improves model behavior.","tokens_in":18025,"tokens_out":4968,"duration_ms":66299,"significance":"The paper addresses a timely and important problem: memory-enabled personalization in LLMs may alter decisions in sensitive tasks such as hate speech detection. The zero-shot evaluation across five open-source models on a large aggregated dataset provides credible evidence that contextual prompts can shift classification outcomes, and the country-level FNR analysis is a useful diagnostic. The proposed debias tuning, if it worked as described, would be a practical contribution, and the authors provide a public code repository and explicitly acknowledge label noise in their error analysis. However, the central mitigation claim is not yet supported: the loss function as written does not implement a consistency penalty, no α=0 control is reported, and the debias results lack uncertainty quantification. In addition, the bias measurement is weakened by the removal of invalid responses without re-basing denominators and by the paper's own finding that 74% of sampled false negatives are 'doubtful labels.' These issues prevent the conclusions from being accepted as stated.","major_comments":[{"comment":"The loss defined in Equations (1)-(2) is not a consistency penalty. Equation (2) is L_loss = L_avg_class + α·L_avg_class, which equals (1+α)·L_avg_class whenever the condition on the predictions holds. This is simply a scaled version of the same average cross-entropy; no term compares the context and non-context predictions or logits, so the model is not penalized for inconsistency in any mechanism distinct from ordinary classification. The abstract and Conclusions attribute the gains to 'penalising inconsistent classifications,' but the equation cannot support that mechanism. Furthermore, the value of α is never reported, and no α=0 control (ordinary fine-tuning on CountryHate without the penalty) is run. Without that control, the F1/FNR improvements in Table 6 could be entirely due to standard supervised fine-tuning on 19,306 training instances. The authors should either provide the actual implemented loss (e.g., with a KL-divergence term between the two output distributions), report α, and add an α=0 condition, or revise the claim that the method is a consistency-based debiasing procedure.","section":"Custom loss, Equations (1)-(2)"},{"comment":"The paper removes invalid responses (e.g., 'I cannot perform this action') before computing metrics, but it does not report how many were removed per condition or re-base the denominators. For Llama 3.1, Gemma, and DeepSeek, the country-lang variant produced more than 7,000 invalid generations, which is a substantial fraction of the 4,826 test instances per condition. If the F1 and FNR values in Tables 2-4 are computed only over valid responses, then differences across conditions may reflect variation in the models' refusal rates rather than in classification accuracy. The authors should report the number and rate of valid responses for each model and condition, and recompute the main comparisons either on the full set (treating invalid outputs as errors) or with appropriate per-condition denominators.","section":"Output processing"},{"comment":"The paper's own error analysis states that 74% of 100 sampled false negatives are 'doubtful labels,' with examples such as 'I hate racist people very much' being labelled as hate speech. This directly undermines the reliability of the gold labels used to establish bias in RQ1-RQ3 and to evaluate debias tuning in RQ4. If a large fraction of the labels are questionable, then the FNR and F1 gaps between contexts (Tables 2-5) are partly disagreements with noisy annotations rather than clean measures of model bias. The authors should quantify the impact of label noise, for example by manually validating a random sample of test instances per condition, reporting the agreement rate, or re-running the key comparisons on a cleaned subset.","section":"Error Classes of False Negatives"},{"comment":"The claim that debias tuning 'significantly improved' model behaviour is not supported by any significance testing or uncertainty quantification. The fine-tuning procedure is stochastic (LoRA, batch sampling), yet only a single run appears to be reported, with no confidence intervals, repeated-seed variance, or paired tests. The chi-squared tests in Tables 5 and 7 assess whether country-context predictions differ from baseline (or from the UK), not whether the differences between the base and debiased models are statistically significant. The authors should report variance across seeds and test the improvement in F1/FNR for significance, or soften the 'significantly' language in the Abstract and Conclusions.","section":"Debias Tuning Results, Table 6"}],"minor_comments":[{"comment":"The text says the twelve countries were 'randomly selected' from the bottom-25 indices, but the list appears to be a purposive sample. Please clarify the random procedure (e.g., seed and sampling frame) or state that the selection was purposive to achieve regional coverage.","section":"Country augmentation selection"},{"comment":"The paper does not name the translation tool used to translate the English posts into the personas' official languages, nor does it discuss translation quality. This is a reproducibility issue; please identify the tool and any quality checks.","section":"Languages"},{"comment":"The sentence 'We performed a chi-squared test (X2) at p > 0.01' appears to be a typo; the subsequent discussion says the null hypothesis is rejected at p < 0.01. Please correct the notation.","section":"Analysis Results, chi-squared test"},{"comment":"The FNR for Australia in Table 5 is 69.64%, but the text just below the table says '69.00%.' Please ensure the numbers are consistent.","section":"Table 5"},{"comment":"The word 'significantly' is used in the Abstract and Conclusions (e.g., 'debias tuning significantly improved LLMs'' behaviour') without a statistical test. If no significance test is added, please replace 'significantly' with 'substantially' or another neutral qualifier.","section":"Abstract and Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The paper has a genuinely interesting research question and the country-level bias evidence is worth publishing if the methodological issues are addressed. The main load-bearing problem is the custom loss: as written it cannot support the consistency-penalty mechanism, and without an α=0 control the debiasing contribution is indistinguishable from ordinary fine-tuning. The label-noise issue is acknowledged but under-treated; given that 74% of false negatives are 'doubtful labels,' the authors should demonstrate that their bias conclusions are robust to label quality. I also recommend insisting on reporting valid-response rates and uncertainty estimates, as these are standard expectations in this area. No concerns about novelty or scope; the related work is adequate but could be strengthened by comparing against other debiasing baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague — quick take: the empirical core is worth taking seriously; the mitigation is not yet supported as described.\n\nWhat's genuinely new: this is the first study I know that puts geographic persona prompting together with hate speech classification, and the effect sizes are hard to dismiss. Llama 3.1's false-negative rate jumping from ~32% to over 71% under country personas is a real behavioral shift, not a rounding error. The pattern—Western personas producing lower FNR than the targeted countries—also lines up with prior geographic-bias work, so I believe the RQ1–RQ3 result.\n\nWhere it gets soft: the custom loss in Equations 1–2 is not a consistency penalty. It is (1+α) times the average cross-entropy, with α>0 when a hard prediction is inconsistent. No term compares with-context and without-context logits or predictions. As written, it is per-example upweighting of the standard loss, and if α is a fixed scalar it is global scaling. That matters because the abstract and conclusions credit the gains to penalising inconsistencies. The stress-test note is right: without an α=0 fine-tune, the reported improvements could come from ordinary supervised fine-tuning on 19,306 CountryHate instances. The α value is also not reported, so even the reweighting is not reproducible.\n\nOther soft spots, in proportion: invalid responses are dropped without re-basing denominators; the translation tool is unnamed; LoRA rank is missing; there are no confidence intervals or repeated-seed variance. These are all fixable with a revision. The label-noise issue is more substantive. The error analysis says 74% of 100 sampled false negatives are 'doubtful labels,' e.g. 'I hate racist people very much.' That does not kill the bias finding—the same items get different verdicts under different personas, so there is inconsistency regardless of label quality—but it does weaken the absolute F1/FNR claims and suggests some of the 'bias' is disagreement with questionable annotations rather than clean miss-classification.\n\nBottom line: this paper deserves a serious referee. The geographic-bias result is novel and deployment-relevant, and the authors are transparent about limitations and error analysis. But the mitigation section needs major revision: either correct the loss description and report α, or add the α=0 control, plus full hyperparameters, translation details, and a label-sensitivity analysis. I'd send it to review, not desk reject.","headline":"Country and language personas really do shift LLM hate-speech verdicts, but the debias-tuning loss as written is just reweighted cross-entropy, not a consistency penalty, so the mitigation needs an ablation before it is taken at face value.","tokens_in":18557,"tokens_out":3940,"would_cite":true,"duration_ms":48293,"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":"The paper claims that LLM memory features that add country or language context introduce geographic bias into hate speech detection, and that a debias-tuning fine-tune with a consistency penalty reduces that bias while improving baseline…","keywords":["hate speech detection","geographic bias","LLM personalisation","debias tuning","country personas","multilingual bias","consistency loss","false negative rate"],"falsifier":"Take the paper's own 100 false-negative sample, have annotators re-label each post under a strict hate-speech definition, and recompute per-country false-negative rates on the cleaned labels; if the large gaps between high-bias countries and the UK or US collapse or shrink to noise, the claimed geographic bias is largely a label-quality artefact.","tokens_in":17576,"feed_emoji":"⚖️","tokens_out":5669,"duration_ms":64640,"temperature":0.7,"pith_summary":"The paper tries to establish that the memory and personalisation features now built into large language models—which inject user details such as country and language into the context—change how those models classify hate speech, creating systematic geographic and language bias. It claims that adding a country persona lowers F1 scores and raises false-negative rates for most tested models, with especially large effects for some countries, and that translating the same posts into the persona's language often worsens the gap. To fix this, the paper fine-tunes models with a custom loss that penalises predictions that flip between a no-context baseline and a country-persona version, forcing consistency. If right, personalisation features can silently degrade a sensitive moderation task, and a relatively cheap debias-tuning step can recover much of the lost performance while also improving the no-context baseline.","feed_headline":"Country context flips LLM hate speech verdicts","feed_subtitle":"A consistency-penalty fine-tune shrinks the gap between personalised and neutral hate speech detection.","key_machinery":"The central mechanism is 'debias tuning', a fine-tuning procedure built on a custom consistency loss. For each post, the model sees both the plain text and the same text wrapped in a country-persona prompt; the loss combines ordinary cross-entropy on both views with an extra penalty term, active only when the context-free prediction is correct but the persona-context prediction is wrong or the output is invalid, scaled by a hyperparameter $\\alpha$. This forces the model toward answers that do not flip because of location or language context. The paper also builds CountryHate, a country-augmented subset of the MetaHate collection, and uses a fixed persona template plus a strict True/False task prompt as the test harness.","core_discovery":"The paper's central claim is that injecting geographic and linguistic context into an LLM's prompt—the way memory features do—biases hate speech detection: the same text can be labelled hate speech for one country persona and not for another. Across five open-source LLMs, the no-context baseline consistently outperforms the country-persona setting on F1, and false-negative rates rise sharply for countries the model associates with negative stereotypes; for Llama 3.1, the false-negative rate jumps from about 32% to over 71% under country persona prompts. The paper further claims that debias tuning—fine-tuning with a loss that penalises inconsistency between context-free and country-context predictions when the context-free prediction is correct—reduces these discrepancies, improves F1 in both personalised and non-personalised settings, and generalises to countries not seen during tuning.","pith_inferences":["Beyond the paper: a natural extension is to apply the same consistency-penalty tuning to other sensitive tasks, such as sentiment or content moderation for self-harm, to test whether personalisation bias is a general phenomenon rather than specific to hate speech.","The paper's own error analysis implies the measured bias magnitudes may be upper bounds; re-benchmarking with adjudicated labels could separate genuine country bias from disagreement with questionable annotations.","The debias loss penalises only the direction where the context-free prediction is correct and the persona-context prediction is wrong; adding symmetric penalties for flips in the opposite direction might change which errors are suppressed.","Because memory features in commercial models are hidden, a fixed public benchmark of country-persona prompts could serve as a low-cost monitoring tool for new model releases."],"forward_implications":["If LLM memory personalisation includes country or language, hate speech detection can vary by user, so systems relying on such personalisation may under-protect users from some regions.","Debias tuning recovers performance not only in personalised contexts but also in the no-context baseline, so fine-tuning for consistency does not trade away general accuracy.","The debiasing effect generalises to countries not used in tuning, though unevenly; broader country coverage during tuning would likely improve fairness further.","Multilingual debias tuning, using persona languages during fine-tuning, can generalise to unseen languages, though not uniformly across all languages.","Residual errors concentrate in doubtful labels and implicit or sarcastic hate speech, so benchmark label quality becomes the next bottleneck for measuring bias."],"supporting_citations":[{"why":"Supplies the debias tuning method: fine-tuning with a loss that penalises inconsistent predictions with and without sensitive context.","marker":"(Dong et al. 2024)"},{"why":"Provides MetaHate, the 1.2M-instance hate speech collection from which the country-labelled subset CountryHate is built.","marker":"(Piot, Martín-Rodilla, and Parapar 2024)"},{"why":"Defines personalisation and the persona concept that motivates simulating memory features through country-persona prompts.","marker":"(Zhang et al. 2024)"},{"why":"Provides the persona template ('Adopt the identity of...') used to create country-specific prompts.","marker":"(Gupta et al. 2024)"},{"why":"Demonstrates geographic bias with nationality-assigned personas, the prior finding this paper extends to hate speech detection.","marker":"(Kamruzzaman and Kim 2024)"},{"why":"Provides evidence that LLMs are geographically biased, motivating the comparison across low- and high-ranked countries.","marker":"(Manvi et al. 2024)"},{"why":"Supplies the error taxonomy used in the manual error analysis of false negatives and false positives.","marker":"(van Aken et al. 2018)"},{"why":"Provides the hate speech definition and the problem of offensive language that shapes the classification task and error analysis.","marker":"(Davidson et al. 2017)"}],"fun_headline_variants":["LLM hate speech calls flip with country persona","Debias tuning cuts geographic bias in hate speech detection","Country context skews LLM hate speech judgments","Fine-tune shrinks LLM hate speech bias by persona","Context personalization tilts LLM hate speech calls"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gold labels in the country-augmented hate speech dataset are reliable enough that differences in false-negative rates across personas measure model bias rather than disagreement with questionable annotations.","fun_headline_variants_meta":{"raw":{"variants":["LLM hate speech calls flip with country persona","Debias tuning cuts geographic bias in hate speech detection","Country context skews LLM hate speech judgments","Fine-tune shrinks LLM hate speech bias by persona","Context personalization tilts LLM hate speech calls"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00049,"raw_usage":{"total_tokens":2381,"prompt_tokens":887,"completion_tokens":1494,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":1418}},"tokens_in":503,"tokens_out":1494,"duration_ms":11820,"temperature":1.0,"reasoning_tokens":1418,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:57:23.447734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the paper's own 100 false-negative sample, have annotators re-label each post under a strict hate-speech definition, and recompute per-country false-negative rates on the cleaned labels; if the large gaps between high-bias countries and the UK or US collapse or shrink to noise, the claimed geographic bias is largely a label-quality artefact.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides MetaHate, the 1.2M-instance hate speech collection from which the country-labelled subset CountryHate is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the persona template ('Adopt the identity of...') used to create country-specific prompts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the error taxonomy used in the manual error analysis of false negatives and false positives."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the hate speech definition and the problem of offensive language that shapes the classification task and error analysis."}],"review_version":1}