{"id":"c0750044-baa2-4ccc-9e03-d568e6901a64","arxiv_id":"2501.02295","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"LLMs show strong stereotype-consistent choices on an indirect fill-in-the-blank task but weak agreement with stereotypes on a direct rating task.","lead":"This paper evaluates social bias in LLMs with two different prompt-based tests, a fill-in-the-blank task it calls implicit bias and a Likert agreement task it calls explicit bias, and finds large gaps between them. The finding matters because it suggests alignment training may hide, rather than remove, bias, but the validity of the implicit measure is not established.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The implicit-bias measure is not validated: masked word choice is a deliberative task, and the explicit measure does not actually use self-reflection, so the explicit-implicit dissociation may be a task-format artifact.","rationale":"The reader's weakest_assumption identifies the same core issue, and I agree. The central claim depends on two measurement constructs: the implicit completion score and the explicit self-reflection score. The implicit construct is unvalidated, and the explicit construct is mislabeled because no self-reflection actually occurs. A falsifiable test—adding an explicit anti-stereotyping instruction—can distinguish an automatic association from a task-framing effect. Because the paper reports no such validation, the rejection is warranted; my review does not change the reader's verdict. I am not attacking the authors; the issue is that the operationalization does not support the psychological construct claimed.","tokens_in":12044,"tokens_out":5942,"duration_ms":65430,"concrete_test":"On a representative subset of models (e.g., GPT-4o and LLaMA-3.1-405B), rerun the implicit completion templates with one added instruction: 'Choose the two words without relying on stereotypes.' Keep candidate lists and template order identical. If the stereotypical score drops substantially (e.g., by more than half) relative to the original protocol, the original score is a controllable response to task framing rather than an automatic association, undercutting the implicit-bias claim. If the score does not drop, the interpretation gains support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the masked sentence-completion task in Section 3.1 measures implicit bias—automatic, unconscious associations. The prompt asks the model to deliberately select two names to fill '<mask> is attrX as <mask> is attrY'; this is a conscious, rule-governed word choice, and no independent evidence shows it taps unconscious cognition rather than corpus co-occurrence or instruction-following. The self-reflection description in Section 3.2 is also inaccurate: the explicit probe does not present the model with its own implicit-phase completions for evaluation; it simply asks for Likert agreement with a generic stereotype sentence. Therefore the central 'explicit-implicit inconsistency' may be an artifact of comparing an indirect completion probe with a direct agreement probe, especially because aligned models are explicitly trained to disagree with stereotype statements. The scaling and alignment conclusions inherit this problem; no validity check or per-instance failure rate is reported for the implicit measure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a two-phase, 'self-reflection-based' framework for measuring explicit and implicit social bias in LLMs. In the implicit phase, models fill masked analogy templates such as '<mask> is often to CEO, as <mask> is often to secretary' by choosing candidate names. In the explicit phase, models rate agreement with stereotype statements (e.g., 'Men are often to CEOs as women are to secretaries') on a 5-point Likert scale. The authors define a Stereotypical Score (SC) as the proportion of stereotype-consistent responses and report experiments on six prominent LLMs across six social dimensions (age, disability, gender-career, gender-occupation, race, science). They find high implicit SC and low explicit SC across all models. Scaling experiments on LLaMA families show that explicit bias decreases while implicit bias increases with model size and training data, and DPO-based alignment reduces explicit bias but leaves implicit bias stable. The paper concludes that LLMs exhibit an explicit-implicit bias inconsistency analogous to that found in humans.","tokens_in":12237,"tokens_out":5848,"duration_ms":59424,"significance":"The paper addresses a timely and important question: whether direct agreement probes and indirect association probes reveal different levels of stereotyping in LLMs, and how those levels change with model scaling and alignment. The use of multiple LLaMA model families with documented token counts is a strength, and the paper's core empirical observation—that indirect completion templates yield much higher stereotype-consistent rates than direct Likert items—is reproducible in principle. The framing is also valuable because it jointly studies explicit and implicit bias, which is rare in the LLM bias literature. However, the significance depends entirely on interpreting the indirect measure as 'implicit bias' in the psychological sense. That interpretation is not validated, and the explicit phase is not actually based on self-reflection as described. Without addressing these issues, the central conclusion is unsupported.","major_comments":[{"comment":"The implicit bias measure is not validated. The masked sentence-completion task is a deliberative choice among candidate names; no evidence is given that it measures automatic, unconscious associations. The psychological IAT relies on response-time differences under speeded conditions, which a word-choice task does not reproduce. The authors do not provide convergent validity (e.g., correlation with human IAT D-scores or established bias benchmarks), test-retest reliability, or per-item consistency checks. Without such evidence, the high 'implicit' SC may simply reflect corpus co-occurrence or instruction-following, and the observed explicit-implicit dissociation may be a task-format artifact. This concern directly undermines the central claim in Sections 5.1 and 6 that LLMs exhibit implicit bias that persists after alignment.","section":"Section 3.1"},{"comment":"The explicit measure is not self-reflection as described. Section 3.2 states that the model evaluates 'its potential attitudes demonstrated during the implicit bias measurement phase,' but the actual prompt in Figure 1(b) is a generic Likert rating of a stereotype statement. The model is never shown its own implicit-phase completions. Moreover, the explicit task differs from the implicit task in multiple ways beyond direct versus indirect measurement: it uses group labels (men/women) instead of names (John/Lisa) and an agreement scale instead of a forced choice. These differences alone could produce the score gap. The paper therefore does not implement the claimed self-reflection methodology, and the comparison is not a clean test of explicit versus implicit attitude.","section":"Section 3.2 and Figure 1"},{"comment":"The scaling analysis cannot separate training-data scale from model size. In the LLaMA family, model size and pre-training token count are nearly perfectly confounded: LLaMA-2 7B/13B/70B all use 2T tokens, LLaMA-3 8B/70B and LLaMA-3.1 405B all use 15T, and LLaMA-3.2 1B/3B use 9T. The claim that 'increasing training data and parameters' drives the observed trends is therefore not supported by these data. Disentangling the two effects would require models trained with varying token counts at a fixed parameter count, or vice versa. This confound also limits the generality of the conclusion that implicit bias increases with scaling.","section":"Section 5.2.1 and Figure 3"}],"minor_comments":[{"comment":"The abstract says explicit bias is evaluated by 'prompting LLMs to analyze their own generated content,' but the explicit prompt does not use content generated in the implicit phase; it uses fixed stereotype statements. Please align the abstract with the actual procedure.","section":"Abstract and Section 3.2"},{"comment":"The claimed total of 14,400 experiments counts only the six models in Table 1, but the scaling and alignment experiments on additional LLaMA base and tuned models are not included in this arithmetic. Please clarify the total number of experiments actually run.","section":"Section 4.1"},{"comment":"The figure should include a legend or point labels identifying model family and token count; without this, the confound between model size and training data is not visually transparent.","section":"Figure 3"},{"comment":"The Stereotypical Scores are reported as point estimates without confidence intervals or standard errors, even though N=200 per cell; the table would benefit from uncertainty quantification to prevent over-reading of small differences.","section":"Table 3"},{"comment":"The template phrase 'is often to' is ungrammatical; consider using a grammatical analogy format and reporting whether results are robust to template wording.","section":"Section 3.1 and Appendix A"},{"comment":"The Limitations section does not acknowledge the absence of validity evidence for the implicit measure or the fact that the explicit phase is not self-reflection; these are central limitations and should be discussed.","section":"Limitations"}],"recommendation":"reject","confidential_remarks":"The paper has a clear core idea and a large empirical effort, but the construct-validity problem is fundamental: the indirect measure is not validated as 'implicit bias,' and the explicit measure is not self-reflection. The scaling analysis also conflates model size with training-data scale. These are load-bearing issues rather than local presentation problems, so I recommend rejection. That said, the empirical materials—particularly the LLaMA-family comparisons and the template-based scoring—could be useful if the authors reposition the work as measuring 'indirect versus direct stereotype expression' and substantially temper the psychological claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline finding — that LLMs give near-neutral Likert ratings on stereotype statements while producing stereotypical completions under a fill-in-the-blank probe — is real and worth a look. The paper's scale is a plus: six bias dimensions, six frontier models, and a systematic sweep across the LLaMA family with alignment-training ablations. The cross-model consistency, especially the sharp gap between direct and indirect measures, is striking, and the scaling result (explicit bias down, indirect bias up) is provocative. That part of the paper is a solid empirical contribution for the bias-evaluation community.\n\nBut the central framing is not supported. The 'implicit' measure is a masked sentence-completion task in which the model deliberately selects two names from a candidate list. That is a conscious, rule-governed word choice, not an automatic or unconscious association. No evidence is given that it taps implicit cognition rather than corpus co-occurrence or instruction-following. The explicit measure is also misdescribed: Section 3.2 never feeds the model its own implicit-phase completions; it simply replaces masks with group labels (\"Men are often to CEOs as women are to secretaries\") and asks for Likert agreement. The abstract's claim about \"analyzing their own generated content\" is inaccurate. The dissociation may therefore be a task-format artifact, not a finding about two forms of bias.\n\nThe scaling analysis has a confound: all LLaMA-2 models use 2T tokens, all LLaMA-3 models use 15T, so Figure 3 conflates training data scale with model size. There are also no error bars or significance tests; with 200 runs per condition, binomial confidence intervals would have been trivial to compute. Minor point: the reader's report mentions factor analysis, but the paper's three 'factors' are just three experimental manipulations — there is no statistical factor analysis.\n\nThe paper is clearly written, the templates are in the appendix, and the authors honestly list some limitations. But the load-bearing validity issue means the headline claim is overstated. I would send this to peer review, because the question matters and the data are substantial, but I would expect major revision: either validate the indirect measure against something external, or reframe the contribution as direct-versus-indirect probe sensitivity and drop the 'implicit/unconscious' language. Add uncertainty quantification and disentangle data size from model size. As it stands, it is a useful empirical resource whose interpretation is not yet earned.","headline":"A broad and provocative empirical study, but the 'implicit vs. explicit' framing overclaims: the implicit probe is an unvalidated completion task, the 'self-reflection' step never uses the model's own outputs, and the scaling analysis conflates data size with model size.","tokens_in":12745,"tokens_out":2360,"would_cite":false,"duration_ms":25933,"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":"This paper claims that large language models exhibit a systematic split between explicit and implicit bias, with explicit bias appearing mild and implicit bias strong, and that scaling and alignment training widen this split rather than…","keywords":["implicit bias","explicit bias","self-reflection","Implicit Association Test","social bias in LLMs","model scaling effects","alignment training","stereotypical score"],"falsifier":"If a probe that measures association strength without requiring a conscious choice — for example, comparing the model's token probabilities for stereotype-consistent versus stereotype-inconsistent completions — fails to reproduce the strong implicit scores, or if a model can be induced to fill the blanks stereotype-consistently while showing no corresponding probability difference, the claimed implicit–explicit distinction would be an artifact of the completion prompt.","tokens_in":11877,"feed_emoji":"⚖️","tokens_out":3814,"duration_ms":35820,"temperature":0.7,"pith_summary":"Large language models, like people, may show a gap between what they openly state and what they reveal under indirect testing. The authors measure implicit bias by asking models to fill blanks in analogy statements that pair social groups with attributes, and explicit bias by asking the same models to rate whether such statements express stereotypes on a Likert scale. Across six social dimensions and six models, they report that explicit stereotyping is mild while implicit stereotyping is strong, and that this gap widens as models grow and as alignment training is applied.","feed_headline":"LLMs hide strong implicit bias behind mild explicit bias","feed_subtitle":"Across six social dimensions, masked-choice tests reveal stereotypes that direct self-report hides while scaling and alignment make the…","key_machinery":"The self-reflection evaluation framework. Implicit bias is measured by adapting the Implicit Association Test to prompt templates of the form \"<mask> is attrX as <mask> is attrY\", where the model must choose two names or group stimuli to fill the blanks; explicit bias is measured by presenting the same association with explicit target groups and asking the model to rate agreement on a 5-point Likert scale, instructing it to reflect on the statement it had just completed. The pairing of the two measures on identical attribute pairs is what allows the explicit–implicit comparison.","core_discovery":"The paper's central discovery is a systematic explicit–implicit bias inconsistency in large language models: when a model is asked directly whether men are to CEOs as women are to secretaries, it strongly disagrees, but when the same association is probed through masked word-choice completions, it reliably selects the stereotypical pairing. The authors contend this mirrors the human pattern documented in social psychology, and that the two measures diverge with respect to scaling: more training data and parameters reduce self-reported stereotyping while increasing stereotype-consistent completions, and preference alignment reduces the former while leaving the latter largely unchanged.","pith_inferences":["An untested reading is that the 'implicit bias' here is really task-format bias: masked completion may simply reflect corpus co-occurrence statistics rather than an attitude, while the explicit rating may reflect instruction-following; if so, the central gap is a difference between a cloze task and a refusal task, not between two kinds of cognition.","A testable extension would be to correlate the masked-completion score with an independent response-time or log-probability measure of association strength; if they do not track each other, the 'implicit' label is doing too much work.","The scaling results suggest a possible 'bias scaling law' opposite to the one for capabilities: implicit association strength rises with scale while explicit compliance rises, so any safety evaluation that only checks stated preferences would be blind to the trend.","Because the alignment experiments use only DPO on a single dataset, the conclusion that alignment cannot touch implicit bias should be read as specific to that method and data, not as a general property of all alignment techniques."],"forward_implications":["If the claim holds, bias evaluations that rely on self-reported or refusal-based tests underestimate the stereotypes that language models actually operate with.","Scaling up models and training data would make the explicit–implicit discrepancy worse, not better, so mitigation would need to target implicit associations directly rather than relying on scale.","Alignment methods such as DPO can train models to deny stereotypes while leaving the underlying statistical associations intact, implying that surface behavior can be made non-stereotypical without changing the model's internal associations.","The reported positive correlation between model capability ranking and implicit bias would mean that the most capable models are also the most implicitly biased.","Any bias benchmark that uses only explicit probes would be measuring a compliance behavior rather than the strength of the underlying association."],"supporting_citations":[{"why":"Supplies the theoretical distinction between explicit and implicit bias that the whole evaluation framework is built on.","marker":"(Greenwald and Banaji, 1995)"},{"why":"Provides the Implicit Association Test method that the implicit-bias prompt template is adapted from.","marker":"(Greenwald et al., 1998)"},{"why":"Supplies validated target words, attribute words, and stimuli for four of the six stereotype categories.","marker":"(Caliskan et al., 2017)"},{"why":"Supports the social-psychology claim that implicit bias forms early and persists, used to explain the observed human-like inconsistency.","marker":"(Baron and Banaji, 2006)"},{"why":"Provides the account that social norms suppress explicit expression, which the paper analogizes to alignment training.","marker":"(Crandall et al., 2002)"},{"why":"Prior work adapting IAT to measure implicit bias in LLMs, which this paper extends and compares against.","marker":"(Bai et al., 2024)"},{"why":"Supplies the DPO algorithm used in the alignment-training experiments on LLaMA-3.2 models.","marker":"(Rafailov et al., 2024)"},{"why":"Defines the LLaMA-2 model family used as baselines and in the scaling analysis.","marker":"(Touvron et al., 2023)"},{"why":"Defines the LLaMA-3 and LLaMA-3.1 model families used for the data-scale and parameter-size experiments.","marker":"(Dubey et al., 2024)"}],"fun_headline_variants":["Self-reflection exposes LLMs' hidden implicit bias","LLMs pass explicit bias tests, fail implicit ones","Masked tests reveal LLMs' real stereotypes","Scaling cuts explicit bias, fuels implicit bias","LLMs claim fairness, yet show implicit bias"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results stand on the assumption that filling blanks in a masked analogy sentence captures the same automatic, unconscious associations that the Implicit Association Test captures in people, rather than merely reflecting surface word co-occurrence in training data.","fun_headline_variants_meta":{"raw":{"variants":["Self-reflection exposes LLMs' hidden implicit bias","LLMs pass explicit bias tests, fail implicit ones","Masked tests reveal LLMs' real stereotypes","Scaling cuts explicit bias, fuels implicit bias","LLMs claim fairness, yet show implicit bias"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1253,"prompt_tokens":868,"completion_tokens":385,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":313}},"tokens_in":484,"tokens_out":385,"duration_ms":4040,"temperature":1.0,"reasoning_tokens":313,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:13:57.814955+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a probe that measures association strength without requiring a conscious choice — for example, comparing the model's token probabilities for stereotype-consistent versus stereotype-inconsistent completions — fails to reproduce the strong implicit scores, or if a model can be induced to fill the blanks stereotype-consistently while showing no corresponding probability difference, the claimed implicit–explicit distinction would be an artifact of the completion prompt.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the theoretical distinction between explicit and implicit bias that the whole evaluation framework is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the social-psychology claim that implicit bias forms early and persists, used to explain the observed human-like inconsistency."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the account that social norms suppress explicit expression, which the paper analogizes to alignment training."}],"review_version":1}