{"id":"562356ec-a78e-42e4-9278-691ff5668ff7","arxiv_id":"2511.06160","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LLMs solve logic-grid puzzles more accurately when the solution matches gender stereotypes than when it contradicts them, revealing implicit bias in deductive reasoning.","lead":"This paper introduces PRIME, a logic-puzzle benchmark that measures whether LLMs let gender stereotypes distort deductive reasoning. Across five models, solutions matching stereotypes are solved more accurately than anti-stereotypical ones, exposing a bias that safety guardrails and explicit-bias tests miss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"S−AS gap rests on unvalidated LLM translation of formal clues; if AS English clues are less faithful or more ambiguous, measured Δ is an artifact rather than stereotype-driven reasoning.","rationale":"Reviewing the full manuscript, the central claim is the S-versus-AS accuracy gap. The triplet design and substitution mechanism are well controlled at the formal level: Lmin_G is reused and only values are swapped to form S and AS, so formal logical complexity is matched. However, the model never sees the formal logic; it sees English sentences generated by an LLM. The pipeline verifies formal solvability but not NL faithfulness. This is the weakest point because it is a necessary condition for the measured Δ to be interpretable as reasoning bias, and it is not mentioned in Limitations. I agree with the reader's weakest_assumption. A round-trip or template control would settle it. Secondary issues—unreported t-test statistics, small typos in examples—do not change the conditional verdict; they would be fixed in revision. Therefore no change to the reader's CONDITIONAL verdict.","tokens_in":26057,"tokens_out":9524,"duration_ms":101461,"concrete_test":"Round-trip the 6,048 natural-language clue sets actually used in the evaluation: have two independent annotators (or a strong second LLM such as GPT-4) translate each English clue back into formal logic notation, then run the same python-constraint solver used in Section 4.1 to check (a) equivalence to the original Lmin_G/S/AS and (b) unique solvability. Compare failure/ambiguity rates across G, S, and AS. If AS clues are significantly more often non-equivalent or underconstrained, the reported Δ is contaminated by translation; if rates match across variants, the translation assumption is validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 verifies unique solvability on the formal logic notation Lmin, but the clues actually shown to models are English sentences generated by LLaMA-3.3-70B through few-shot prompting (Appendix D.4.2). No check is reported that these natural-language clues are truth-conditionally equivalent to the formal expressions, that they are equally unambiguous across G/S/AS, or that translation failure rates are symmetric. Because S and AS share the same formal clue structure and differ only in which values are paired with which names, any systematic asymmetry introduced by the translator—who may itself have gender-stereotyped language priors—could produce the observed Δ without the solver's deductive reasoning being biased. This is not an internal inconsistency, but it is an unvalidated link in an otherwise careful pipeline. The Limitations section does not acknowledge this gap, and Appendix D.4.2 gives only a single prompt example, not a consistency or round-trip evaluation. If AS translations are even slightly more underconstrained or awkwardly phrased, the central claim that stereotypes act as reasoning shortcuts is not yet established by the measurement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces PRIME, a framework that uses automatically generated logic grid puzzles to measure whether LLM deductive reasoning is affected by gender stereotypes when those stereotypes are logically irrelevant. Puzzle triplets (Generic, Stereotypical, Anti-stereotypical) are generated from a shared formal clue set, checked with a constraint solver, translated to English, and solved by five LLMs over four puzzle sizes. The paper reports edit-distance metrics ED_all, ED_BP, ED_G and a bias difference Δ = ED(S) − ED(AS); results show consistently lower error on S than AS puzzles, larger bias effects in bias-probing categories, a bias toward stereotypical errors, and reduced Δ under chain-of-thought prompting. The authors release dataset and code and claim that alignment/safety training does not eliminate implicit reasoning bias.","tokens_in":26318,"tokens_out":5442,"duration_ms":54929,"significance":"If the findings hold, the paper makes a valuable methodological contribution: it operationalizes implicit bias in a verifiable, structured reasoning task and supplies a reusable generator plus metrics. Strengths include the formal solver verification of the generated clue sets, the inclusion of a neutral baseline, the consistency of the S−AS ordering across model families and puzzle sizes, the error-asymmetry analysis, and a publicly released implementation. The main risk is that the central comparison rests on unvalidated LLM translation of formal clues into English; if translation ambiguity differs across G/S/AS, Δ is not a pure measure of stereotype-driven reasoning. The CoT mitigation result is suggestive but only tested on one model. With additional validation and statistical detail, the contribution would be solid.","major_comments":[{"comment":"The English clues shown to the tested models are produced by LLaMA-3.3-70B few-shot translation of the formal Lmin clues, but no round-trip or equivalence check is reported. Because the headline result is Δ = ED(S) − ED(AS), any systematic difference in translation clarity between S and AS (e.g., awkward or ambiguous AS wording) would produce the observed gap without bias in the solver. Please add a formal round-trip check (parse generated NL back to constraints and verify equivalence and unique solvability) and/or an ambiguity/clarity evaluation stratified by G/S/AS, and explicitly report the outcome in the Limitations section.","section":"4.1/4.3, Appendix D.4.2"},{"comment":"The paper states that paired t-tests were used and that Δ values are 'statistically significant in all models and sizes,' but Tables 2 and 3 report only point estimates. There are no p-values, confidence intervals, standard errors, or effect sizes, and Table 6 shows that on 4×4 many models solve zero puzzles, so the distribution of ED may be highly non-normal and the sample of 504 triplets is not informative without per-condition variance. Add an appendix reporting test statistics, CIs, and the number of usable triplets per condition, and state whether the unit of analysis is puzzle triplet or model run.","section":"4.4/5.1, Tables 2-3"},{"comment":"After substituting values into Lmin,G to produce Lmin,S and Lmin,AS, the text says only 'ensuring both remain solvable.' It is not stated whether the constraint solver re-verifies uniqueness (not just consistency) for each variant. If an S or AS variant has multiple valid solutions, its difficulty changes and ED differences no longer isolate stereotype effects. Please specify the solver check, report how many of the 504 triplets per size failed or were regenerated, and make the verification step explicit.","section":"4.3, step 3"}],"minor_comments":[{"comment":"The Limitations section says the name classification experiment is in Appendix A, but the actual experiment appears in Appendix C.","section":"Limitations"},{"comment":"Δ values are color-coded, but the color coding is lost in grayscale printing. Add explicit plus/minus signs or other symbols for accessibility.","section":"Tables 2, 3, 5"},{"comment":"The sentence 'CoT prompting is a reliable strategy for mitigating social biases' is based on experiments with LLaMA-3.1-70B only. Temper the generalization or add results from additional models.","section":"5.3"},{"comment":"Because puzzle-level accuracy on 4×4 is at or near zero for most conditions, ED_all values in that size should be interpreted and described as partial-grid performance; consider stating this explicitly when interpreting ED_all.","section":"Table 6 / Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about LLM translation validation is legitimate and is the main blocker. The paper is otherwise a solid empirical contribution with reproducible code and a neutral baseline. If the authors provide round-trip checks and statistical details, the central claim becomes much stronger; if they cannot, the conclusion should be scaled back and framed as conditional on translation equivalence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper introduces PRIME, a logic-grid-puzzle framework that generates Generic/Stereotypical/Anti-stereotypical triplets and measures bias via edit distance. The core finding—models solve stereotypical puzzles better than anti-stereotypical ones, with generic in between—is consistent across LLaMA, Qwen, Mixtral, and Gemini, across puzzle sizes, and in the error analysis. That's a real result, and the design is genuinely careful: shared puzzle structure, constraint-solver verification of unique solvability, an anonymized generic baseline, and an edit-distance metric that separates bias-probing columns from general ones. They also test mitigation (CoT and debiasing prefixes) and find CoT reduces the bias gap, plus a guardrail check showing LLaMA-Guard flags almost nothing—a nice demonstration of the gap PRIME fills.\n\nThe soft spots are two, and neither is fatal. First, the natural-language clues shown to models are generated by LLaMA-3.3-70B with few-shot prompting, but there's no verification that the English sentences are truth-conditionally equivalent to the formal logic, or equally clear across S and AS. If the translator phrases anti-stereotypical clues more awkwardly, the measured Δ could be a translation artifact rather than stereotype-driven reasoning. This is testable and fixable—round-trip translation checks or human clarity ratings would do—but as it stands the central claim rests on an unvalidated link. Second, the paper says it uses paired t-tests but never reports test statistics, p-values, or confidence intervals; the table just has point estimates. That's incomplete reporting for a benchmark paper.\n\nMinor quibbles: the CoT result is described as 'consistent' but Appendix F.3.1 shows cases where CoT fails or flips to anti-stereotypical bias, so the headline is a bit strong. And the name-to-gender assumption is acknowledged and checked, which is fine.\n\nThe paper deserves a serious referee. It's a new, reproducible evaluation artifact—they ship code and data—and the finding is meaningful for fairness evaluation. The translation validation and statistical reporting should be addressed in revision, but the core direction looks solid.","headline":"A genuinely new probe for implicit bias in LLM reasoning, with a solid core finding and one unvalidated link in the pipeline that needs fixing before the benchmark is adopted.","tokens_in":26760,"tokens_out":3147,"would_cite":true,"duration_ms":31224,"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":"This paper claims that LLMs solve logic puzzles more accurately when the solution matches gender stereotypes than when it contradicts them, even though stereotypes are logically irrelevant to the puzzle.","keywords":["implicit bias","large language models","logic grid puzzles","deductive reasoning","gender stereotypes","chain-of-thought","bias evaluation","edit distance"],"falsifier":"A concrete check: have several human raters or a second LLM judge the ambiguity of every translated clue while blind to variant, and analyze only puzzle triplets whose clue sets are matched for surface-level difficulty; if the S−AS performance gap disappears on the matched subset, the gap is a translation artifact rather than stereotype-driven reasoning. Alternatively, solving the same puzzles from formal-logic clue notation directly (bypassing the LLM translation) would either confirm or falsify the shortcut claim.","tokens_in":25962,"feed_emoji":"🧩","tokens_out":5208,"duration_ms":47016,"temperature":0.7,"pith_summary":"PRIME, a new evaluation framework, generates logic grid puzzles in three versions—stereotypical, generic, and anti-stereotypical—that share the same logical structure and differ only in whether demographic associations confirm or contradict common gender stereotypes. Testing several LLM families across puzzle sizes, the paper finds a consistent ordering: models solve stereotypical puzzles best and anti-stereotypical puzzles worst, with generic puzzles in between. The authors argue this is evidence that stereotypes act as reasoning shortcuts: models lean on socially familiar associations instead of pure deduction, even when those associations carry no logical information. The paper also shows that chain-of-thought prompting narrows the gap substantially, while a static debiasing prefix does not reliably work. If correct, the result implies that safety-aligned models can still systematically favor stereotypical conclusions in reasoning-heavy tasks, with direct implications for fairness in automated decision-making.","feed_headline":"LLMs solve puzzles better when answers confirm stereotypes","feed_subtitle":"A logic-puzzle benchmark shows safety-aligned models favor stereotypical solutions even when stereotypes are irrelevant.","key_machinery":"The load-bearing object is the puzzle triplet: a generic puzzle plus stereotypical and anti-stereotypical variants generated from one shared solution grid and one minimal clue set, so the three versions are logically equivalent and differ only in how names map to stereotype-bearing values. Two measures carry the analysis: edit distance (how many element swaps separate the model's solved grid from the ground truth) computed overall, on the bias-probing category, and on neutral categories; and the bias difference Δ = ED(S) − ED(AS), where negative values indicate stereotypical bias. This machinery lets the paper separate reasoning accuracy from bias and localize where bias enters the inference","core_discovery":"The central discovery is that implicit gender stereotypes measurably distort deductive reasoning in LLMs. Using logic grid puzzles—formal constraint-satisfaction tasks with a unique solution that requires no outside knowledge—the paper creates structurally identical puzzles that either align a stereotype (e.g., a woman's name paired with 'nurse'), contradict it (the same woman paired with 'doctor'), or remove demographic cues entirely. Across all models and puzzle sizes, performance is highest on stereotypical puzzles, intermediate on generic ones, and lowest on anti-stereotypical ones, with statistically significant differences. The bias is concentrated in the stereotype-probing category (t","pith_inferences":["The same triplet mechanism should transfer to other identity dimensions such as race, nationality, or religion; a testable prediction is that the size of the bias gap will track how strongly names cue the dimension in the model's training data.","PRIME could be turned into a continuous bias-auditing benchmark: release new puzzle triplets over time to detect when stereotype biases shift after model updates, much like regression testing.","A natural extension is to vary clue type (e.g., 'Either Or' vs 'Neither Nor') to identify which logical forms are most susceptible to stereotype leakage; the current paper reports aggregate behavior, not per-clue-type breakdowns.","If the translation-confounding concern is genuine, an obvious control would be to render clues in a semi-formal notation or to use multiple translators; the paper's claim would be strengthened if the S−AS gap persisted under such controls."],"forward_implications":["If the central claim holds, safety-aligned LLMs that refuse explicitly biased prompts still exhibit systematic stereotype-driven shortcuts in deductive reasoning, so refusal rates are not a reliable signal of fairness.","Standard reasoning benchmarks that score only overall accuracy will miss this bias; evaluating patterns across stereotype-aligned and stereotype-contradicting variants is needed to surface it.","Chain-of-thought prompting reduces the measured bias gap by about 59% on average, suggesting structured step-by-step reasoning is a viable mitigation, though it does not eliminate the gap and weakens on larger puzzles.","Because bias is localized in stereotype-probing categories and largely absent from neutral categories, targeted auditing of identity-relevant decisions is feasible without auditing every reasoning output.","Scaling model size improves reasoning accuracy but does not remove the stereotype gap, so larger models are not automatically fairer."],"fun_headline_variants":["Logic puzzles expose LLMs' hidden gender bias","LLMs reason better when answers match stereotypes","New benchmark reveals LLMs' implicit bias in deduction","Puzzle test shows LLMs lean on stereotypes to solve logic"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The natural-language clues shown to the models are faithful, equally clear translations of the formal logic across the generic, stereotypical, and anti-stereotypical variants, so that the only meaningful difference between variants is the demographic association.","fun_headline_variants_meta":{"raw":{"variants":["Logic puzzles expose LLMs' hidden gender bias","LLMs reason better when answers match stereotypes","New benchmark reveals LLMs' implicit bias in deduction","Puzzle test shows LLMs lean on stereotypes to solve logic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000124,"raw_usage":{"total_tokens":911,"prompt_tokens":686,"completion_tokens":225,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":163}},"tokens_in":430,"tokens_out":225,"duration_ms":3006,"temperature":1.0,"reasoning_tokens":163,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T23:19:45.351290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: have several human raters or a second LLM judge the ambiguity of every translated clue while blind to variant, and analyze only puzzle triplets whose clue sets are matched for surface-level difficulty; if the S−AS performance gap disappears on the matched subset, the gap is a translation artifact rather than stereotype-driven reasoning. Alternatively, solving the same puzzles from formal-logic clue notation directly (bypassing the LLM translation) would either confirm or falsify the shortcut claim.","supporting_citations":[],"review_version":1}