{"id":"853ac047-7404-4fc3-b91d-63cac367e543","arxiv_id":"2411.16527","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Twelve LLMs show significant gender bias in contextual embeddings, with female names skewed toward warmth and male names toward competence on stereotype-content dimensions.","lead":"This paper measures gender stereotypes in twelve large language models by projecting their word embeddings onto seven social-psychology dimensions such as warmth and competence. It produces visual bias profiles showing that female-associated names skew warm and male names skew competent across models, offering a communication tool for AI fairness.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Direction-prediction validation is partly circular; a held-out external validation of the warmth/competence axes is needed to rule out dictionary-internal geometry.","rationale":"The paper's contribution is a measurement framework plus an empirical finding: consistent warmth/competence gender associations across twelve LLMs. The measurement framework's validity rests on two legs: the projection procedure and the stereotype dictionaries. The projection procedure is a direct transfer of SensePolar/POLAR, which prior work validated with human judgments; the reader's verdict treated this as plausible. The weaker leg is the construct validity of the axes themselves, because the Sec 4.1 evaluation uses the same dictionary family that defines the poles. I read the strongest claim and found that the gender difference itself is supported by the paper's t-tests on 100 names per gender, and the observed asymmetry aligns with prior human-stereotype literature, so I do not think the central empirical finding is fabricated or likely to vanish entirely. However, the empirical finding is only as strong as the interpretation that the projected axes measure warmth and competence rather than a general valence or a dictionary-specific factor. The paper's own discussion and limitations concede that the method is not a complete bias benchmark and that context choices matter, which is honest but does not resolve the validation gap. The reader's weakest_assumption identified exactly this circularity, and my independent reading of the manuscript text (Sec 3.3 pole construction from seed dictionary; Sec 4.1 evaluation on extended dictionary terms from the same source) confirms it. I therefore agree with the reader's verdict of CONDITIONAL, and my concrete test is the single check that would settle whether the concern lands: external validation with independent stereotype-rated words. I am not raising a separate concern about the post hoc cut-off adjustment, the layer-wise means without error bars, or the small gendered-term samples, because those are secondary and would not by themselves overturn the main claim; the circular validity check is the load-bearing issue.","tokens_in":13646,"tokens_out":1689,"duration_ms":15570,"concrete_test":"Reproduce the Sec 3.3 projection and Sec 4.1 direction-prediction task, but construct the evaluation terms from an external source independent of Nicolas et al. (2021), for example single-word English translations of Fiske et al. (2002)'s original warmth/competence scale items or affective norms rated on warmth and competence. If accuracy on this held-out eval set is materially lower (e.g., below 0.70 for both the 2D and 7D spaces across the twelve models), then the axes do not generalize beyond the dictionary and the gender profiles should not be interpreted as stereotype-content measurements.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim—twelve LLMs show significant gender bias along warmth and competence—depends on the polar projection actually recovering the intended stereotype dimensions. The validation in Sec 4.1 projects additional terms from the same Nicolas et al. (2021) dictionary that defined the poles in Sec 3.3, so high accuracy on this task mainly shows the projection separates the seed words from near-synonyms collected by the same semi-automated method. In particular, the sign-based accuracy measures whether projected values agree with the dictionary's low/high labels, not whether the axes correspond to independently established stereotype content. If the axes capture valence (good/bad) or some other general semantic factor shared by the seed lists, the warmth/competence asymmetry for gender-associated names could still be real, but it would not be licensed by the stated validation. The paper's own caveat that GPT-Neo-125M and others need mean-centered cut-offs to reach reasonable accuracy further shows the zero-sign threshold is not well calibrated even for the internal task, weakening the justification for interpreting projected signs as stereotype direction. Thus the profiles are plausible and consistent with prior work, but their construct validity as stereotype dimensions, rather than dictionary-internal geometry, is the least secure load-bearing assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for profiling gender bias in large language models by projecting contextual word embeddings onto stereotype dimensions from the stereotype content model (warmth, competence, and seven granular facets). The authors construct polar axes from seed terms in the Nicolas et al. (2021) dictionary, apply a linear projection to embeddings of gender-associated names and gendered terms, and generate 2D and 7D stereotype profiles for twelve open-source LLMs. They report statistically significant gender differences in the projected warmth and competence values, largely consistent with human stereotype content studies, and analyze how the results vary across context types, model layers, and model families. The paper also makes available code for the method.","tokens_in":13817,"tokens_out":5122,"duration_ms":55461,"significance":"If the stereotype dimensions are accepted as valid, the paper offers a practical, interpretable tool for exposing and communicating representational bias in LLMs, which is valuable for both AI audit and public communication. The authors provide a broad empirical sweep across twelve models, use theory-grounded dictionaries, make their code available, and back the headline gender-bias claim with t-tests on a set of 100 names per gender. The layer-wise and context-type analyses add useful nuance. The main caveat is that the validation of the stereotype axes relies on the same dictionary family that defines the axes, so the construct validity of the dimensions is the least secure part of the argument.","major_comments":[{"comment":"The validation of the stereotype dimensions is partly circular. The polar axes are defined by seed terms from the Nicolas et al. (2021) dictionary, and the direction-prediction task in Sec. 4.1 evaluates projection accuracy on \"additional terms from the extended stereotype dictionary,\" which is the same dictionary, built semi-automatically as synonyms of the seed terms via WordNet. High accuracy on this internal task mostly shows that the projection separates seed words from their near-synonyms; it does not independently establish that the axes correspond to the psychological warmth and competence constructs. This matters because the paper interprets the gender profiles as evidence about SCM-aligned bias. I recommend an external validation, for example predicting human-rated stereotype content for social groups (as in Fraser et al. 2021) or using a dictionary from a different source, or at least adding an explicit statement that the validation is internal consistency rather than independent construct validation.","section":"Sec. 3.3 and Sec. 4.1"},{"comment":"Equation (5) writes d = (a^T)^{-1} x, but a is an h×d matrix with h=2 or 7 and d equal to the embedding dimension (e.g., 768), so a^T is not square and has no inverse. The authors must state whether they use the Moore-Penrose pseudo-inverse or a least-squares solution, and define the projection accordingly. As written, the formula is mathematically undefined and the projected values are not reproducible without the code's specific numerical routine.","section":"Sec. 3.3, Eq. (5)"},{"comment":"The mean-centered cutoff for GPT-Neo, GPT2, and the Gemma models is introduced after observing poor accuracy with the zero cutoff. This is a post hoc adjustment that affects the accuracy numbers in Table 2 and the claim that \"all models can reasonably discriminate between low and high-labeled terms.\" The paper should either motivate the cutoff a priori (e.g., because the projection has no built-in centering) or report both zero-cutoff and mean-centered results without presenting the latter as the default. This does not necessarily invalidate the gender-bias t-tests in Sec. 4.2, since a constant shift would not change the between-group difference, but the validation evidence for those models is weakened.","section":"Sec. 4.1"}],"minor_comments":[{"comment":"The construction of the warmth and competence directions from subordinate dimensions should be stated explicitly: are the word lists simply concatenated before averaging, or are the subordinate direction vectors averaged? This affects the exact definition of the axes.","section":"Sec. 3.3"},{"comment":"The sentence \"For all twelve studied LLMs, we find statistically significant bias for gender-associated names\" is followed by the caveat that GPT-Neo shows associations of both dimensions with female names and OLMo has an insignificant competence difference. Please clarify that the claim refers to at least one dimension being significantly different, and quantify the p-value threshold and correction method used for the t-tests.","section":"Sec. 4.2"},{"comment":"There is a typo \"Lama-3-8B\" in Figure 1 and \"T erms\" in Figure 3; please fix these spelling errors.","section":"Fig. 1 and Fig. 5"},{"comment":"The number of context examples is set to five, but the paper does not report how sensitive the results are to this choice. A brief analysis or discussion would help assess robustness.","section":"Sec. 3.2"},{"comment":"For the religion dimension, the imbalance between high-labeled (142) and low-labeled (6) additional terms is noted; this makes the reported accuracy less informative. Please also report per-class or balanced accuracy for that dimension.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable fit for a computational social science or NLP-oriented venue. The central empirical finding—gender-associated names differ in the projected warmth and competence dimensions in a direction consistent with human stereotypes—is plausible and well-supported by the t-tests on held-out names. The main issues are the circularity of the dimension validation and the undefined projection formula, both of which are fixable in revision. The post hoc cutoff adjustment should also be handled transparently. I would not reject, but the manuscript needs a careful revision before it can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: the paper delivers what it promises—a clear, reproducible way to visualize gender bias in contextual embeddings along warmth/competence and finer stereotype dimensions, applied to twelve open models. The headline result, female names shifted toward warmth and male names toward competence, holds across models and is tested on held-out names, which is a real empirical finding.\n\nWhat's new is the breadth: twelve models, seven dimensions, layer-wise analysis, and multiple context types (generated, dictionary, Reddit, none). The method itself is borrowed—SensePolar projection plus the Nicolas et al. dictionaries—but the combination with a multi-model audit is useful and the code is released. The t-tests on 100 names per gender are appropriate, and the authors are honest about small populations and the limits of embedding-based bias measures. The paper's own limitations section covers most of the scope caveats.\n\nThe soft spots are the ones you'd expect. First, the dimension validation in Sec 4.1 projects additional terms from the same Nicolas et al. dictionary that defined the poles, so accuracy on that task mostly shows internal consistency, not construct validity. The warmth/competence axes could partly capture valence or another general factor. The gender analysis uses names that are not in the dictionary, so the gender asymmetry is not itself circular, but the interpretation that it's specifically warmth vs. competence leans on an axis identity that is not independently established for contextual embeddings. The mean-centered cutoff adjustment for GPT-Neo, GPT2, and Gemma is post hoc; without a pre-registered threshold it weakens the sign-based interpretation for those models. The layer-wise plots lack error bars, which is minor.\n\nI don't think these issues sink the paper. The gender profiles are consistent with prior work on static embeddings and with the SCM literature, and the held-out names give the central comparison independent standing. But for a general bias audit tool, the dimensions need external validation—human ratings or an independent lexicon—and the cutoffs need to be justified rather than fit.\n\nWho should read it: people working on bias auditing, model documentation, or interpretability. It deserves a serious referee; the revision path is clear. I'd engage with it.\n\nBest","headline":"Useful empirical extension of stereotype-content profiling to twelve LLMs; the gender asymmetry is robust across models, but the dimension validation is partly circular and the cutoff handling is post hoc.","tokens_in":14378,"tokens_out":2248,"would_cite":true,"duration_ms":21528,"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":"The paper claims that projecting contextual embeddings onto stereotype dimensions from social psychology reveals significant gender bias in all twelve studied LLMs, with most models placing female names at warmth and male names at…","keywords":["bias in LLMs","stereotype content model","gender bias","contextual word embeddings","bias profiles","polar projection","warmth and competence","embedding interpretability"],"falsifier":"Rebuild the warmth and competence pole vectors from an independently sourced, human-rated antonym list, re-project the same 100 male and female names, and check whether the female-warmth/male-competence separation persists; if it disappears or flips, the bias profile is an artifact of the seed-word choice.","tokens_in":13416,"feed_emoji":"⚖️","tokens_out":6875,"duration_ms":61977,"temperature":0.7,"pith_summary":"This paper argues that the hidden embedding spaces of large language models can be made legible by projecting them onto stereotype dimensions from social psychology, and that doing so reveals a consistent gendered pattern. Across twelve open LLMs, historically female-associated names project toward the warmth pole and male-associated names toward the competence pole; under gender-neutral generated contexts, every model shows statistically significant differences for names, and most show the same female-warmth/male-competence alignment. The authors present these projections as bias profiles: visual, intuitive descriptions of where social groups sit on dimensions like sociability, morality, ability, and agency, intended for communicating risk to non-specialists and for guiding debiasing. The claim matters because it connects the opaque internal states of LLMs to a well-established human psychology framework, giving bias measurement a theoretical grounding rather than a purely correlational one.","feed_headline":"Bias check: 12 LLMs split genders along warmth and competence","feed_subtitle":"Social-psychology stereotype axes turn opaque embeddings into readable bias profiles.","key_machinery":"The carrying mechanism is a polar projection onto stereotype axes, following the SensePolar approach. For each dimension, the averaged contextual embeddings of the seed dictionary's high-pole words ($p_{\\mathrm{high}}$) and low-pole words ($p_{\\mathrm{low}}$) define a direction vector $a_h = p_{\\mathrm{high}} - p_{\\mathrm{low}}$; stacked, these form a change-of-basis matrix $A$, and a target term's averaged embedding $x$ is projected as $d = (A^T)^{-1}x$, so the sign of $d$ indicates association with the high or low pole. The warmth and competence axes aggregate the finer dimensions (sociability plus morality, and ability plus agency), while the 7D profile keeps these subdimensions separate. This transformation is what turns opaque high-dimensional embeddings into the two- or seven-number profiles the paper visualizes.","core_discovery":"The core claim is that the stereotype content model's two organizing dimensions, warmth and competence, are present in the contextual embeddings of current LLMs and that gender-associated names are positioned along them in a way that mirrors human stereotype research. Using gender-non-specific generated contexts, the 100 historically female and male names separate significantly on these axes in all twelve models, with female names associated with warmth and male names with competence in nearly all cases; the two exceptions are one model where both dimensions favor female names and one where the competence difference is not significant. The seven-dimensional profiles add detail: sociability and morality drive the female warmth association in ten or eleven models, while ability, agency, and sometimes status drive the male competence association. The authors further report that these associations are distributed across network layers rather than concentrated in any single layer, and that a smaller set of nine gendered terms shows a reliable warmth bias for the newest Llama models, with the direction of the term differences consistent with the name profiles.","pith_inferences":["The dimensions' validation is dictionary-internal; a natural extension is to rebuild the axes from an independent, human-rated antonym list and test whether the female-warmth/male-competence asymmetry survives.","The method's statistical power comes from large name sets, so profiling individual terms or small gender groups will require many more context samples or a Bayesian treatment to be trustworthy.","The lower warmth of nonbinary and transgender terms in newer models is suggestive; with more contexts it could be developed into a quantitative diagnostic for harms beyond binary gender.","Because the dimensions come from social psychology rather than from the target group, the same projection could be turned toward ethnicity, age, or other stereotype dimensions without changing the measurement pipeline."],"forward_implications":["Bias in these models is not only an output-level artifact: it is encoded in the geometry of hidden representations and visible through a simple linear projection.","The same stereotype axes can serve as a group-agnostic measurement device across model families, sizes, and generations, without needing examples of the discriminated group in the test data.","Choice of context changes the measured profile, so gender-neutral generated contexts provide a baseline of pre-existing bias while domain-specific contexts can probe task-relevant bias.","Because stereotype associations persist across layers, mitigation that targets only the final layer would leave most of the representational bias in place.","Bias profiles give a non-expert-accessible visualization of representational harm, suitable for model cards and public accountability."],"supporting_citations":[{"why":"Supplies the stereotype content model with warmth and competence as the two primary dimensions.","marker":"Fiske et al., 2002"},{"why":"Provides the seed and extended dictionaries defining high and low poles for seven stereotype dimensions.","marker":"Nicolas et al., 2021"},{"why":"Supplies the SensePolar projection method for contextual embeddings used to build the stereotype space.","marker":"Engler et al., 2022"},{"why":"Establishes the direction-prediction evaluation and the precedent of reproducing stereotype content model dimensions in static embeddings.","marker":"Fraser et al., 2021"},{"why":"Provides the nine gendered terms per gender used as the smaller vocabulary population.","marker":"Caliskan et al., 2017"},{"why":"Introduces the polar-opposite transformation that SensePolar extends.","marker":"Mathew et al., 2020"},{"why":"Provides the neutral template contexts (e.g., 'This is [NAME]') used for names and gendered terms.","marker":"May et al., 2019"},{"why":"Supplies the Reddit-based context sampling used as the natural-data condition.","marker":"Guo and Caliskan, 2021"}],"fun_headline_variants":["LLMs encode gender stereotypes: female names warmth, male names competence","Warmth and competence axes expose gender bias in 12 LLMs","Bias profiles: How LLMs place gendered names on warmth and competence","LLM embeddings show female names skew warmth, male names skew competence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stereotype axes are validated with test words drawn from the same dictionary that defined the poles, so the measured gender asymmetry could reflect the dictionary's internal word geometry rather than stereotype content that holds independently.","fun_headline_variants_meta":{"raw":{"variants":["LLMs encode gender stereotypes: female names warmth, male names competence","Warmth and competence axes expose gender bias in 12 LLMs","Bias profiles: How LLMs place gendered names on warmth and competence","LLM embeddings show female names skew warmth, male names skew competence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2871,"prompt_tokens":819,"completion_tokens":2052,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":1975}},"tokens_in":435,"tokens_out":2052,"duration_ms":15754,"temperature":1.0,"reasoning_tokens":1975,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:59:47.219441+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rebuild the warmth and competence pole vectors from an independently sourced, human-rated antonym list, re-project the same 100 male and female names, and check whether the female-warmth/male-competence separation persists; if it disappears or flips, the bias profile is an artifact of the seed-word choice.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the stereotype content model with warmth and competence as the two primary dimensions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the polar-opposite transformation that SensePolar extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Reddit-based context sampling used as the natural-data condition."}],"review_version":1}