{"id":"473dc0c1-3d37-4fb5-ace5-b09169334849","arxiv_id":"2508.07753","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Social bias is a statistically significant cause of faithfulness hallucinations in LLMs, with anti-stereotypical contexts increasing errors and pro-stereotypical contexts decreasing them.","lead":"This paper tests whether social bias, such as gender or income stereotypes, causes large language models to hallucinate in question-answering tasks, rather than merely correlating with errors. It builds a new dataset with three bias states (pro, anti, and non-stereotype) and reports that anti-stereotype text triggers more wrong answers while pro-stereotype text suppresses them.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal claim hinges on the unvalidated 'Precision' criterion of the bias intervention: swapping social-attribute words in BID also changes token frequency, sentiment, and edit distance, which are known hallucination confounders, so the observed Pro-Anti differences may not isolate B→H.","rationale":"The paper is a serious empirical study with an appropriate paired design, and the McNemar test is a reasonable tool for the binary hallucination outcomes. The dataset construction is thoughtful, and the claimed effect is large and consistent across many model/bias combinations. However, the central novelty is explicitly causal: the authors claim to go 'beyond mere correlation analysis' and to demonstrate that biases 'directly cause' faithfulness hallucinations. That claim depends entirely on the intervention being a faithful do-intervention, i.e., changing only the bias variable B while holding all other hallucination-relevant context features Z fixed. The paper states this as the Precision criterion but provides no empirical evidence that it holds. The examples in Figure 7 show that different bias states are realized by different attribute words, which also differ in frequency, connotation, and syntactic position; these are classic confounders for LLM behavior, and the paper itself lists word frequency as a confounder in §3.1. The Consistency criterion is similarly asserted: the Pro-Anti pair, which drives the headline directional result, differs in two lexical tokens while the Non-Pro and Non-Anti pairs differ in only one, so edit distance is not balanced across the three pairwise comparisons. None of this proves the causal claim is false, but it means the paper's strongest claim is not yet supported. A placebo test or a covariate analysis would settle the issue. This is an addressable concern, so the appropriate verdict remains CONDITIONAL, as the reader concluded.","tokens_in":17122,"tokens_out":7829,"duration_ms":96415,"concrete_test":"On a random sample of BID templates, build placebo pairs that keep the bias state fixed (both Non-stereotype) but use the same attribute-word families from Figure 7 (wealthy/rich vs low-income/poor) and run the same greedy-decoding models. If hallucination rates differ significantly between the two Non variants, surface lexical choice alone changes H, directly falsifying the Precision criterion. As a complementary check, compute token-frequency/perplexity distances between Pro-Anti text pairs and test whether model-level ICE correlates with these distances.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 defines do(B=...) as a text edit, and the BID construction (§4, Figure 7) realizes each bias state by replacing attribute words such as wealthy/rich/low-income/poor. Under Pearl's do-calculus, identification requires exchangeability: E[H | do(Pro)] - E[H | do(Anti)] equals E[H(Pro)-H(Anti)] only if no unmeasured confounder Z changes across the intervention. The paper's Precision criterion asserts this but never verifies it. These lexical substitutions are not neutral: 'wealthy' vs 'low-income' differ in token frequency, familiarity, and sentiment, and §3.1 explicitly lists word frequency as a hallucination confounder. Thus the Pro-Anti contrast (Figure 7) cannot distinguish 'bias state causes H' from 'surface form causes H'. The Consistency criterion is also unmet for the central Pro-Anti pair: those texts differ in two attribute tokens while Non-Pro and Non-Anti pairs differ in only one, so the strongest effect may reflect greater textual perturbation rather than bias state. No manipulation check, placebo condition, or covariate analysis is reported. Releasing code/data would help, but as written the headline causal claim reduces to a correlation between template edits and model errors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims that social bias states (anti-stereotype, pro-stereotype, non-stereotype) have a causal effect on faithfulness hallucinations in large language models. It formalizes the problem with a Structural Causal Model, defines bias interventions via template-based text edits, and proposes ICE/UCS statistics tested with McNemar's test. The authors construct a new dataset, BID, from BBQ scenarios and evaluate seven LLMs. They report that anti-stereotype contexts increase hallucinations while pro-stereotype contexts suppress them, and they introduce a new category called 'unfairness hallucination'. The central claim is that this is the first direct causal demonstration, going beyond correlation.","tokens_in":17441,"tokens_out":4862,"duration_ms":59698,"significance":"If the causal claim is established, the paper would be a valuable contribution: it takes a step from correlational studies of hallucination triggers toward interventional analysis, contributes a medium-scale bias intervention dataset, and uses deterministic greedy decoding and paired McNemar tests, which are appropriate for the binary outcome setting. The authors also clearly list the ideal conditions (effectiveness, precision, consistency) that an intervention should satisfy. However, the experiment as reported does not validate those conditions, and the central causal conclusion is currently not supported by the evidence presented.","major_comments":[{"comment":"The intervention do(B=...) is implemented as lexical substitution of social-attribute words. The Precision criterion in §3.2 requires that all contextual elements except B remain unchanged, but no validation is provided. In Figure 7's Pro-Anti pair, the text changes from 'wealthy/poor' to 'low-income/rich'; these tokens differ in frequency, register, and sentiment, which §3.1 itself lists as hallucination confounders. Since no manipulation check, placebo condition, or covariate analysis is reported, the observed Pro-Anti difference cannot be distinguished from an effect of surface-form perturbation. This undermines Eq. (1) and the central causal claim in §5.2.2.","section":"§3.2, Figure 7"},{"comment":"Consistency is not met for the main Pro-Anti comparison. In Figure 7, Non-Pro and Non-Anti pairs change one attribute token, while the Pro-Anti pair changes two attribute tokens. Thus the Pro-Anti ICE may be inflated by larger textual perturbation, and the directional conclusion 'Anti-stereotype increases, Pro-stereotype suppresses' in §5.2.2 is not identified. The authors should either construct all pairs with identical edit distance and matched token-frequency distributions or conduct a control analysis with 'sham' attribute swaps.","section":"§3.2, Figure 7"},{"comment":"The language 'Social bias is a key cause of faithfulness hallucinations' overstates what the experiment establishes. Without verification of the intervention criteria, Table 3 reports associations between template edits and model errors, not causal effects. The authors should add validation experiments (e.g., randomized attribute shuffles, token-frequency-matched controls, and a pre-registered analysis plan) and soften causal claims accordingly.","section":"§5.2.2, Table 3"}],"minor_comments":[{"comment":"Typo: 'we first defines the key concepts' should read 'we first define...'.","section":"§3.1"},{"comment":"The description of one-tailed tests is nonstandard: a chi-square test has no left/right tail as described. Direction should be read from the sign of ∑ICE after a significant two-sided test; please clarify the statistical language.","section":"Appendix B.4"},{"comment":"The text says options are randomly shuffled and multiple generations are performed, but greedy decoding is deterministic. Clarify how randomness affects the reported rates and whether confidence is averaged over option orders.","section":"§5.1"},{"comment":"The reported ordering Pro < Non < Anti is presented as a trend without error bars or variance estimates. Since greedy decoding is deterministic, repeated sampling from option shuffles could provide a check of stability.","section":"Figure 4"},{"comment":"References [24] and [25] are the same paper; [26] appears irrelevant to the word-frequency confounder claim. Also, the Figure 7 caption says the pairs 'differ only in social attributes,' but the lexical realizations differ (e.g., wealthy vs. rich), so the caption overstates the intervention's precision.","section":"References / Figure 7"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is within the journal's scope, and the dataset and evaluation design are potentially useful resources. The main risk is overclaiming causality; I see no citation-pattern concerns. If the authors can add validation of the intervention (placebo pairs, matched token-frequency controls, and a careful rewriting of causal claims), the manuscript could become publishable. As written, the central causal conclusion is not supported by the reported evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this paper has a genuinely useful empirical finding—across seven LLMs, anti-stereotype contexts produce more faithfulness hallucinations than pro-stereotype ones, and the effect concentrates on what they call unfairness hallucinations. The BID dataset and the Unfairness Hallucination category are real contributions regardless of the causal framing. If you work on hallucination evaluation or bias benchmarks, it's worth a read.\n\nWhat's actually new: they frame the bias–hallucination link causally with an SCM and define an intervention (swap social-attribute words) with three explicit criteria: Effectiveness, Precision, Consistency. They then run paired comparisons and McNemar tests. That's an honest attempt to go beyond Ladhak et al.'s correlation.\n\nThe soft spot is exactly where the stress-test note points. The precision criterion is asserted, not demonstrated. In Figure 7, the Pro-Anti pair 'wealthy/poor' vs 'low-income/rich' changes two attribute tokens, and those tokens differ in frequency and sentiment—both listed as confounders in Section 3.1. So the observed difference could be surface-form sensitivity, not bias state. Consistency is also violated for the Pro-Anti pair: they change two tokens while Non-Pro and Non-Anti change one, so the biggest effect is confounded with edit size. No manipulation check, no placebo, no covariate analysis. That doesn't kill the paper, but it means the headline 'bias directly causes hallucinations' is stronger than the evidence. What they've really shown is a robust correlation between stereotype-violating text and hallucination rates, with a do-calculus wrapper.\n\nAlso minor: the SCM machinery is mostly rhetorical—they don't adjust for any observed confounders, and the graph does no work beyond justifying paired edits. The self-citation to their own ICR probe is fine and not load-bearing.\n\nWho this is for: people building bias evaluation suites or studying hallucination triggers. The dataset and the unfairness-hallucination distinction will be reused. The causal language should be softened, and they should release data and code plus a validation that the edits are precision-consistent (e.g., human judgments, token-frequency matching, or a placebo condition). If they fix that, it's a solid paper. As is, it deserves referee time, but the referee should push on the intervention validity.","headline":"Useful dataset and a clean empirical trend, but the causal claim outruns the intervention: swapping attribute words isn't a validated do-operation.","tokens_in":17921,"tokens_out":1723,"would_cite":true,"duration_ms":19326,"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":"Social bias causally shifts whether LLM outputs stay faithful to the input, with anti-stereotype phrasing increasing errors.","keywords":["faithfulness hallucination","social bias","causal inference","structural causal model","bias intervention","unfairness hallucination","large language models","statistical significance testing"],"falsifier":"Take the Pro-Anti pairs from the dataset and replace the social-attribute words with alternative synonyms matched for token frequency, length, and association strength. If the hallucination-rate gap between pro- and anti-stereotype versions disappears or reverses under frequency-matched edits, the intervention is not isolating bias state. A simpler companion check: regress hallucination outcome on bias state and on the token log-frequency of the swapped attribute words; if token frequency absorbs the bias-state effect, the causal claim fails.","tokens_in":17045,"feed_emoji":"⚖️","tokens_out":9531,"duration_ms":103954,"temperature":0.7,"pith_summary":"This paper tries to establish that the social-bias framing of a text—whether it matches, contradicts, or avoids a stereotype—is a cause, not merely a correlate, of faithfulness hallucinations in large language models. The authors build a structural causal model and design controlled bias interventions that rewrite the same scenario into pro-stereotype, anti-stereotype, and non-stereotype versions differing only in social attributes. Across seven models and five bias categories, they find that anti-stereotype wording increases hallucinations, pro-stereotype wording suppresses them, and the effect is concentrated in a newly named class, unfairness hallucinations. If right, this means bias is not a tangential fairness concern but a direct driver of model unfaithfulness, and it can hide inside confident, hard-to-detect errors.","feed_headline":"Anti-stereotype prompts make LLMs hallucinate more","feed_subtitle":"A controlled causal study of seven models ties bias states to faithfulness errors, with pro-stereotype text hiding them.","key_machinery":"The central machinery is a three-state bias intervention built on a structural causal model: each scenario has a bias-state node $B$, a hallucination node $H$, and a confounder node $Z$, and the intervention $\\mathrm{do}(B=\\cdot)$ sets the bias state directly while cutting paths from confounders. Paired templates make the intervention concrete by rewriting the same scene into Pro-, Anti-, and Non-stereotype versions that differ only in the social attributes of the characters. The Individual Causal Effect (ICE) records whether the binary hallucination outcome changes between two versions; summing ICEs and running McNemar's test yields the Unified Causal Significance (UCS), which reports both","core_discovery":"The paper's central claim is that bias states directly cause faithfulness hallucinations in LLMs, not just correlate with them. Using paired texts that differ only in social attributes, it reports that shifting from non-stereotype to anti-stereotype raises hallucination rates significantly in 34 of 35 model-bias combinations, shifting from pro-stereotype to anti-stereotype raises them in 32 of 35, while pro-stereotype phrasing suppresses hallucinations relative to non-stereotype in 19 of 35 combinations. The same significance analysis shows the causal effect falls almost entirely on unfairness hallucinations—errors where the model chooses a person whose social attribute is inconsistent with","pith_inferences":["The causal claim depends on the precision of the intervention, which the paper asserts but does not directly measure. In the worked examples the swapped words are near-synonyms with different lexical frequencies, so part of the effect could come from token-level confounds rather than bias state itself.","A direct test of that concern would be to repeat the comparisons using attribute words matched on frequency and surprisal; if the Pro-Anti gap persists, the causal story is strengthened, and if it shrinks, the bias-state effect is partly lexical.","The consistent direction of the effect suggests a mechanism: stereotype-consistent text aligns with the model's prior and helps it stay on task, while stereotype-contradicting text fights that prior and increases error. This could be probed by tracking internal representations across the three bias states.","The framework could rank other context features by causal significance, producing an ordered map of what makes LLMs unfaithful rather than a single correlation."],"forward_implications":["Bias-induced hallucination is a distinct failure mode: two models with similar hallucination rates can have very different causal sensitivity to bias, so aggregate accuracy is not a reliable proxy.","Deployment of QA systems should treat anti-stereotype contexts as higher-risk inputs for unfaithful answers.","Evaluation sets built from stereotype-aligned text can overstate faithfulness because pro-stereotype prompts suppress errors.","Unfairness hallucinations deserve separate reporting in hallucination benchmarks; confidence-based filters will miss them.","The same intervention-and-test recipe can be applied to other suspected causes of context-driven hallucinations, such as information position or irrelevant content."],"supporting_citations":[{"why":"Supplies the intervention operator and causal-graph formalism used to define do(B=...) and separate causation from correlation.","marker":"[21]"},{"why":"Source question-answering benchmark whose scenarios were reworked into the paper's BID templates.","marker":"[20]"},{"why":"Paired binary-outcome significance test used for the paper's causal-effect significance claims.","marker":"[17]"},{"why":"Documents a correlation between pre-training bias and hallucinations in summarization, the pattern this paper claims to upgrade to causation.","marker":"[12]"},{"why":"Shows LLM output consistency varies across social groups, motivating a direct causal study of bias and faithfulness.","marker":"[29]"},{"why":"Identifies irrelevant context as a confounder that can distract LLMs, justifying the need to control contextual features.","marker":"[24]"},{"why":"Shows knowledge overshadowing in context causes hallucination, part of the contextual-cause background the causal model must control.","marker":"[32]"},{"why":"Shows key-information position in context changes hallucination rates, another confounder the intervention design must hold fixed.","marker":"[15]"},{"why":"Provides the pro/anti/non-stereotype classification convention that the bias-state definitions reuse.","marker":"[35]"}],"fun_headline_variants":["Causal study: social bias directly triggers LLM hallucinations","Pro-stereotype masks LLM errors; anti-stereotype exposes them","Bias states cause LLM hallucinations, not just correlate","Anti-stereotype prompts increase LLM hallucination odds","LLM unfairness errors traced to social bias states"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that swapping only the social attributes in a paired text leaves every other feature—topic, sentiment, word frequency, syntax—unchanged; the paper asserts this precision but never measures it directly, so if the swaps change other things, the causal difference is really a correlation with the whole text edit.","fun_headline_variants_meta":{"raw":{"variants":["Causal study: social bias directly triggers LLM hallucinations","Pro-stereotype masks LLM errors; anti-stereotype exposes them","Bias states cause LLM hallucinations, not just correlate","Anti-stereotype prompts increase LLM hallucination odds","LLM unfairness errors traced to social bias states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001407,"raw_usage":{"total_tokens":5496,"prompt_tokens":694,"completion_tokens":4802,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":4721}},"tokens_in":438,"tokens_out":4802,"duration_ms":37308,"temperature":1.0,"reasoning_tokens":4721,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:52:31.620268+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the Pro-Anti pairs from the dataset and replace the social-attribute words with alternative synonyms matched for token frequency, length, and association strength. If the hallucination-rate gap between pro- and anti-stereotype versions disappears or reverses under frequency-matched edits, the intervention is not isolating bias state. A simpler companion check: regress hallucination outcome on bias state and on the token log-frequency of the swapped attribute words; if token frequency absorbs the bias-state effect, the causal claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the intervention operator and causal-graph formalism used to define do(B=...) and separate causation from correlation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents a correlation between pre-training bias and hallucinations in summarization, the pattern this paper claims to upgrade to causation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Identifies irrelevant context as a confounder that can distract LLMs, justifying the need to control contextual features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pro/anti/non-stereotype classification convention that the bias-state definitions reuse."}],"review_version":1}