{"id":"b46e61c5-c587-406b-888f-eea09e96517c","arxiv_id":"2509.04013","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Paraphrasing benchmark questions keeps LLM rankings stable but reduces their accuracy, suggesting static benchmarks overestimate model robustness.","lead":"The paper reworded questions from six standard AI benchmarks and tested 34 language models on the rewritten versions. It found that model rankings stay nearly the same but accuracy drops substantially, suggesting benchmark scores can overstate real-world robustness.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paraphrase semantic fidelity is unvalidated; the accuracy-drop result may reflect altered difficulty rather than model brittleness.","rationale":"The reader's CONDITIONAL verdict is appropriate. The most serious threat is unvalidated paraphrase semantic fidelity. The paper provides indirect evidence (ordered lines, reverse-order stability) that paraphrases are not chaotic, but this does not rule out systematic difficulty shifts. The training/test data inconsistency (§3.1 vs §4.2) and lack of inferential statistics for the drop are additional weaknesses, but they are secondary: the semantic-equivalence assumption, if false, would invalidate the entire measurement. A human or independent evaluation would settle the question. Given the paper's transparent data/code release and the plausibility of the finding, CONDITIONAL remains the right verdict.","tokens_in":14203,"tokens_out":3679,"duration_ms":35383,"concrete_test":"Select 1,000 question–paraphrase pairs stratified across all six benchmarks and have three independent annotators (or a validated high-quality LLM judge with human adjudication) rate whether each paraphrase preserves the original semantic content and keeps the same correct answer. Recompute the mean accuracy drop and Kendall's tau between original and paraphrased rankings using only the pairs unanimously rated as semantically equivalent. If the drop shrinks to a few percentage points or tau falls below the reported 0.9 threshold, the central claim fails; if the pattern persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that absolute performance drops significantly while rankings remain stable—rests on the assumption that the automatically generated paraphrases preserve both meaning and difficulty. The authors explicitly acknowledge in Section 5 that 'Further validation of the generated paraphrases is needed,' and the only validation offered (Section 4.1.3) is indirect: it shows that adding paraphrases increases answer diversity and that reversing order yields similar curves. This does not establish semantic equivalence; it only shows that the paraphrases are not trivially identical and carry no systematic order effect. If GPT-4o-mini's paraphrases systematically alter question difficulty (e.g., by adding or removing clues, or by changing the implicit reasoning required), the observed accuracy drop is an artifact of the paraphrase generation process, not evidence of LLM brittleness. Since the paper's headline result and its 'benchmarks overestimate robustness' conclusion depend on this assumption, it is the most load-bearing point. The lack of human or independent validation makes the empirical contribution conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper systematically examines whether LLM benchmark scores are robust to paraphrased question inputs. Using GPT-4o mini, the authors generate five paraphrases for each question in six multiple-choice benchmarks (ARC-C, HellaSwag, MMLU, OpenBookQA, RACE, SciQ), yielding roughly 265k paraphrased instances. They evaluate 34 instruction-tuned LLMs in a zero-shot, top-1 decoding setting. The central finding is that while model rankings remain stable across paraphrased inputs (Kendall's tau > 0.9 for all benchmarks), absolute accuracy drops substantially: in Figure 6, most models are categorized as 'Over,' meaning their original-question accuracy exceeds their paraphrase distribution. The paper interprets this as evidence that LLMs are brittle to linguistic variation and that current benchmarks overestimate real-world robustness. It also reports consistency analyses, showing that a substantial fraction of questions elicit different answers across paraphrases, and a size-dependent correlation between accuracy and consistency.","tokens_in":14415,"tokens_out":4955,"duration_ms":49356,"significance":"If the central claim holds, the paper makes a valuable empirical contribution by scaling paraphrase-robustness evaluation to 34 models, six benchmarks, and over 260k paraphrased questions. The public release of data, code, predictions, and plots is a strength, as is the breadth of the evaluation. The finding that rankings are stable while absolute scores drop is consistent with a growing body of work on prompt sensitivity and has concrete implications for how benchmark scores should be reported and interpreted. However, the significance depends critically on the unvalidated assumption that the automatically generated paraphrases preserve both meaning and difficulty. The paper itself acknowledges this in Section 5 ('Further validation of the generated paraphrases is needed'), and the internal checks in Section 4.1.3 do not establish semantic equivalence. The study is therefore best viewed as a large-scale, reproducible probe that currently supports a conditional conclusion.","major_comments":[{"comment":"The central claim that absolute accuracy drops because of LLM brittleness rests on the assumption that the GPT-4o mini generated paraphrases preserve semantic content and question difficulty. The only validation offered (Section 4.1.3) is indirect: adding paraphrases increases answer diversity, reverse order gives similar curves, and first/last paraphrase accuracies are equal on average. None of these checks establishes semantic equivalence; a systematic difficulty shift (e.g., adding or removing clues, changing implicit reasoning requirements) would produce exactly the same observable patterns. The paper's own Section 5 states that 'Further validation of the generated paraphrases is needed.' This is a load-bearing gap because if the paraphrases are not semantically equivalent, the accuracy drop is an artifact of the generation process, not evidence of LLM brittleness.","section":"Section 3.3 and Section 5"},{"comment":"The methodology combines test and training instances: 'we present results over both test and sampled training data, treating them uniformly in our evaluation.' This is problematic for a claim about generalization. Models may have seen training instances during pretraining or instruction tuning, especially for older benchmarks that the paper itself suspects contamination (Section 4.2). Figure 6, however, states that sampling was performed 'only on the test set and on the validation set,' creating an inconsistency across the paper. The central analyses should be run on held-out test/validation data only, and the paper should clarify which figures and numbers use the combined set.","section":"Section 3.1"},{"comment":"The abstract and conclusions use the word 'significantly' for the absolute performance drop, but no statistical significance test is reported. The Over/In/Under categories are defined by whether the original accuracy falls outside the interquartile range of the 1,000-sample paraphrase distribution, which is not a hypothesis test. To support 'significant' decline, the authors should provide paired tests (e.g., bootstrap confidence intervals for per-model differences, or a Wilcoxon signed-rank test across models) and report effect sizes for each benchmark.","section":"Section 4.2 and Figure 6"},{"comment":"The paraphrases are generated by GPT-4o mini, which is also one of the 34 evaluated models (model 5 in Table 2). This does not make the experiment circular, but it is a potential confound: the paraphrase distribution may reflect GPT-4o mini's stylistic preferences, and GPT-4o mini itself may have a systematic advantage or disadvantage on these paraphrases. The paper should discuss this and ideally demonstrate robustness by generating paraphrases with a second, independent model (e.g., an open-source model) and repeating a subset of the experiments.","section":"Section 3.3 and Table 2"}],"minor_comments":[{"comment":"The sentence 'This means that paraphrasing indeed introduces increasing confusion for LLMs' is an interpretive leap. The observed increase in answer diversity is consistent with meaningful linguistic variation, but it is also consistent with paraphrase-induced task alteration. Please rephrase to separate observation from interpretation.","section":"Section 4.1.3"},{"comment":"The claim that 'we observed no changes in model responses when modifying the phrasing of the prompt' is not backed by a systematic experiment. Either provide details of this check or soften the claim.","section":"Section 3.4"},{"comment":"Models 30-34 show identical or near-identical accuracies (0.22-0.25) across all benchmarks, which likely indicates degenerate output behavior (e.g., always selecting the same option). The paper excludes them from Figure 6 but includes them in earlier analyses. This should be discussed explicitly, as these models can affect the accuracy-consistency correlation in Section 4.1.2.","section":"Table 2 / Figure 6"},{"comment":"The correlations with benchmark release date are all non-significant (p = 0.53, 0.60, 0.33), yet the conclusion states 'We also uncover some evidence of potential data contamination.' The text should clearly mark this as a speculative conjecture, not a finding, and avoid the word 'evidence' without stronger support.","section":"Figure 7"},{"comment":"The text reports average accuracy of 0.54 for both first and last paraphrase. This is used to support semantic equivalence, but equality of average accuracy does not imply that individual questions have the same difficulty. Please clarify the logical scope of this observation.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's main empirical contribution is potentially valuable and the data release is commendable. However, the load-bearing issue of paraphrase validity is explicitly acknowledged as unresolved, and the training/test conflation further weakens the generalization claim. A major revision is appropriate: the authors should add a human or independent validation of a sample of paraphrases, re-run the central analyses on held-out data, and provide statistical significance tests for the accuracy drop. If the paraphrase validation fails, the headline result may need substantial reinterpretation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a useful paper and the headline result is probably right: across six benchmarks and 34 models, paraphrasing drops absolute accuracy while Kendall tau between original and paraphrased rankings stays above 0.9. That's a new measurement, not just a restatement of prompt-sensitivity work. The scale is real—265k paraphrases, data and code posted—and the consistency-vs-accuracy analysis (small models negatively correlated, large positively) is a nice nuance.\n\nThe main soft spot is exactly where the stress-test lands: the paraphrases are generated by GPT-4o-mini and never validated by humans or an independent generator. Section 5 says 'further validation of the generated paraphrases is needed,' which is honest but means the central claim rests on an assumption. If the paraphrases systematically altered difficulty, the accuracy drop would be an artifact. The paper's indirect checks—reverse-order curves, first-vs-last paraphrase parity—rule out a simple order effect, and the drop is consistent across all six benchmarks, so I don't think the finding is manufactured. But it is conditional. A reviewer should ask for a human-annotated sample or at minimum a second paraphrase generator.\n\nTwo smaller issues. First, the data-split story is inconsistent: Section 3.1 says results cover test plus sampled training data, treated uniformly; Section 4.2 says Figure 6 sampling used only test/validation. This needs clarification because it affects what the boxplots mean. Second, the Over/In/Under counts are descriptive; no test on whether the drop is statistically significant across models. Minor, but easy to add.\n\nI disagree with the reader's weakest-assumption framing only in degree. The paraphrase-fidelity gap is real, but the paper's own checks plus prior work make the direction of the effect credible. I'd send this to a serious referee. The fix is not expensive—human-validate a few hundred paraphrases, clarify splits, add a significance test. The paper is for anyone building or interpreting LLM leaderboards.","headline":"Large, reproducible study showing LLM rankings survive paraphrasing while absolute scores drop; main caveat is unvalidated paraphrase fidelity, which the authors themselves flag.","tokens_in":14899,"tokens_out":2884,"would_cite":true,"duration_ms":27450,"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":"Paraphrasing benchmark questions drops LLM accuracy significantly while preserving model rankings.","keywords":["LLM evaluation","benchmark reliability","paraphrase robustness","linguistic variability","multiple-choice question answering","ranking stability","consistency","data contamination"],"falsifier":"Take a random sample of 500 original questions per benchmark and their five paraphrases, and have human judges who do not know the model answers answer both versions. If humans find the paraphrases equally easy, or if the paraphrases are judged to change the intended answer, then the observed model accuracy drop would be an artifact of the paraphrase generation rather than evidence about model robustness.","tokens_in":14107,"feed_emoji":"📉","tokens_out":6250,"duration_ms":59717,"temperature":0.7,"pith_summary":"Large language models are usually scored on benchmarks with fixed, standardized wording. This paper asks whether those scores hold up when the same questions are rephrased. Across six multiple-choice benchmarks and 34 models, it finds that relative rankings barely move (Kendall's tau above 0.9) while absolute accuracy falls: most models score worse on paraphrased inputs than on the original wording. The authors read this as evidence that current benchmarks provide a valid comparative ordering but overestimate real-world robustness to linguistic variability. They also find that consistency across paraphrases is only a meaningful sign of correctness for larger models; smaller models can be consistently wrong.","feed_headline":"LLM accuracy drops sharply when benchmark questions are reworded","feed_subtitle":"Rankings hold up, but absolute scores overestimate how models handle real-world phrasing.","key_machinery":"The load-bearing mechanism is the paraphrase perturbation protocol. For each original question, an automatic paraphrasing model generates five rewrites that preserve meaning, avoid negations, and keep question form and answer order fixed; every LLM then answers all six versions under zero-shot, top-1 decoding. Accuracy is compared through 1,000 sampled paraphrase sets, producing a boxplot of what the model would score under realistic wording variability, and ranking stability is measured with Kendall's tau between original and median paraphrased accuracy.","core_discovery":"The paper's central claim is that the standard practice of reporting one accuracy number on fixed benchmark wording produces a misleading picture. When each of 52,966 questions across six multiple-choice benchmarks is automatically reworded into five paraphrases and answered by 34 instruction-tuned large language models, the relative ordering of model accuracies is preserved (Kendall's tau above 0.9 on every benchmark), but the absolute accuracy distribution shifts downward: on MMLU, 28 of 34 models score worse under paraphrased samples than on the original wording, 5 overlap, and 1 improves. The authors interpret this decoupling as evidence that current benchmarks are reliable for comparing","pith_inferences":["If paraphrase sensitivity were added to leaderboards as a second axis, the Pareto frontier would change: some models with lower original accuracy but high consistency could be preferable for deployment.","A direct contamination test is available: compare 'Over' counts for models whose pretraining cutoff predates a benchmark's release against models trained after it; the paper's release-date correlation predicts a clear difference.","The protocol can be extended to a continuous robustness metric, such as accuracy variance across paraphrases, and to human-authored paraphrases, which may produce larger drops than automatic ones.","For open-ended generation tasks, the same decoupling might not hold, because correctness judgments themselves are noisy; the multiple-choice setting is the cleanest place to observe the effect."],"forward_implications":["Reported benchmark numbers should be read as upper bounds on performance under fixed wording, not as expected accuracy in applications where users rephrase questions.","Existing leaderboards remain useful for relative comparison; a model ranked higher on original wording will almost always rank higher under paraphrases.","Robustness-aware evaluation should report accuracy distributions or paraphrase sensitivity alongside single scores.","Consistency-under-paraphrase should not be used as a standalone quality metric, especially for small models, where it can indicate stubbornness rather than understanding.","The negative correlation between benchmark release date and the number of models in the 'Over' category is consistent with data contamination in older benchmarks; fresh paraphrased test sets can expose memorization."],"supporting_citations":[{"why":"Supplies MMLU, one of the six benchmark datasets whose questions are paraphrased.","marker":"[16]"},{"why":"Supplies ARC-C, one of the six benchmark datasets.","marker":"[11]"},{"why":"Supplies HellaSwag, one of the six benchmark datasets.","marker":"[46]"},{"why":"Supplies RACE, one of the six benchmark datasets.","marker":"[20]"},{"why":"Supplies OpenBookQA, one of the six benchmark datasets.","marker":"[24]"},{"why":"Supplies SciQ, one of the six benchmark datasets.","marker":"[44]"},{"why":"Provides the notion of semantic brittleness used to interpret accuracy drops under rephrasing.","marker":"[21]"},{"why":"Documents sensitivity of LLMs to prompt formatting, motivating the robustness question.","marker":"[35]"},{"why":"Supports the interpretation that benchmarks 'expire' as models overfit, used for the contamination conjecture.","marker":"[29]"},{"why":"Supports the argument that single average scores misreport evaluation results.","marker":"[5]"}],"fun_headline_variants":["Benchmark rankings hold, but LLM accuracy falls on reworded questions","Reworded benchmarks drop LLM accuracy while rankings stay stable","LLM scores decline on paraphrased benchmarks, but relative order survives","Paraphrased questions reveal LLM benchmarks overstate robustness","Stable rankings, shrinking scores: LLM benchmarks fail rewording test"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The generated paraphrases preserve the original questions' meaning and difficulty, so the accuracy drops measure model fragility to wording rather than changes in the task itself.","fun_headline_variants_meta":{"raw":{"variants":["Benchmark rankings hold, but LLM accuracy falls on reworded questions","Reworded benchmarks drop LLM accuracy while rankings stay stable","LLM scores decline on paraphrased benchmarks, but relative order survives","Paraphrased questions reveal LLM benchmarks overstate robustness","Stable rankings, shrinking scores: LLM benchmarks fail rewording test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2443,"prompt_tokens":738,"completion_tokens":1705,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":1613}},"tokens_in":482,"tokens_out":1705,"duration_ms":12304,"temperature":1.0,"reasoning_tokens":1613,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:27:01.587574+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of 500 original questions per benchmark and their five paraphrases, and have human judges who do not know the model answers answer both versions. If humans find the paraphrases equally easy, or if the paraphrases are judged to change the intended answer, then the observed model accuracy drop would be an artifact of the paraphrase generation rather than evidence about model robustness.","supporting_citations":[{"cited_title":"Hendrycks, C","cited_arxiv_id":null,"evidence_quote":"Supplies MMLU, one of the six benchmark datasets whose questions are paraphrased."},{"cited_title":"Zellers, A","cited_arxiv_id":null,"evidence_quote":"Supplies HellaSwag, one of the six benchmark datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies RACE, one of the six benchmark datasets."},{"cited_title":"Mihaylov, P","cited_arxiv_id":null,"evidence_quote":"Supplies OpenBookQA, one of the six benchmark datasets."},{"cited_title":"Welbl, N","cited_arxiv_id":null,"evidence_quote":"Supplies SciQ, one of the six benchmark datasets."},{"cited_title":"Sclar, Y","cited_arxiv_id":null,"evidence_quote":"Documents sensitivity of LLMs to prompt formatting, motivating the robustness question."},{"cited_title":"Variations in Relevance Judgments and the Shelf Life of Test Collections","cited_arxiv_id":"2502.20937","evidence_quote":"Supports the interpretation that benchmarks 'expire' as models overfit, used for the contamination conjecture."},{"cited_title":"Burnell et al","cited_arxiv_id":null,"evidence_quote":"Supports the argument that single average scores misreport evaluation results."}],"review_version":1}