{"id":"2b3b5a07-381d-4e63-97e5-881088f3305a","arxiv_id":"2505.01482","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"On the GPQA benchmark, GPT-4o's highest accuracy came from self-consistency prompting, about 53 percent correct, but its explanations were least similar to the reference solutions, while direct answer and chain-of-thought prompts gave the closest explanations.","lead":"This paper tested seven ways of prompting GPT-4o on 448 graduate-level science questions, comparing both the answers and the explanations the model gave. It is a useful calibration for anyone choosing prompts for LLM evaluation, since the prompt that scored best on accuracy produced the least faithful-looking explanations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Explanation-quality metric is confounded by missing explanations and prompt-induced style, so the accuracy–explanation decoupling claim is not established.","rationale":"The reader's weakest_assumption identifies cosine similarity as a valid measure of reasoning quality and notes the unresolved handling of missing self-consistency explanations. My concern is the same load-bearing issue, sharpened to a concrete confound: the paper never states how absent explanations enter the similarity statistics, and the prompt-specific structure (self-consistency's missing explanations, multipath's forced justifications) can create the observed ranking without reflecting reasoning quality. The reader also flagged mechanical inconsistencies between the abstract and tables, which I agree are real but treat as secondary to the missing-explanation confound. My recommended verdict remains CONDITIONAL, as the empirical accuracy data and per-subdomain counts are useful, but the central interpretive claim requires either a validated explanation-quality metric or a robustness analysis of the missing-data handling. Since the reader already assigned CONDITIONAL, I mark this as UNCHANGED rather than escalating to REJECT: the paper may still be salvageable with a reanalysis. I agree fully with the reader's weakest_assumption, as our concerns converge on the same metric-validity problem.","tokens_in":12880,"tokens_out":2614,"duration_ms":26889,"concrete_test":"Re-run the explanation analysis on the logged responses: first count missing explanations per prompt, then recompute Table 2 medians on the subset of questions where every prompt produced a non-empty explanation. If self-consistency is no longer the lowest-median technique on that common subset, the claim that explanation quality and accuracy decouple collapses to a data-handling artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central conclusion in Section VI, that 'reasoning quality does not always correlate with accuracy,' depends entirely on comparing explanation quality across prompts via Table 2. Section V states that 'For some questions, self-consistency only provided the answers with no explanation,' but the paper never reports how many explanations were missing per prompt or how these cases entered the cosine similarity computation. If missing explanations were excluded, self-consistency's median of 0.645 is computed on a non-random subset; if coded as zero or low similarity, the gap is an artifact of the coding rule. Additionally, the similarity metric likely tracks prompt-induced style: multipath prompting forced justifications for all four choices, and self-consistency aggregated multiple paths, so the lower cosine similarities may reflect stylistic divergence rather than worse scientific reasoning. The authors acknowledge they are non-experts who did not grade explanations, so there is no external validation that MPNet cosine similarity tracks reasoning quality. The abstract's arithmetic errors (52.99% vs. 52.90%; 'second worst' vs. worst in Table 2) further undermine confidence, but the missing-explanation handling is the load-bearing flaw because it directly determines the decoupling claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates seven prompting techniques (direct answer, chain-of-thought, zero-shot chain-of-thought, self-ask, self-consistency, decomposition, and multipath) with GPT-4o on the 448-question GPQA-main dataset. It measures answer accuracy and, for each prompt, the cosine similarity between the model's elicited explanation and the GPQA ground-truth explanation using an MPNet sentence transformer. The paper reports that self-consistency achieves the highest accuracy (237/448, 52.90% in Table 1) but the lowest median explanation similarity, while direct answer, chain-of-thought, and zero-shot chain-of-thought produce explanations most similar to the ground truth. On this basis the paper concludes that reasoning quality does not always correlate with accuracy, and proposes a research agenda for improving interpretability and robustness.","tokens_in":12921,"tokens_out":5394,"duration_ms":51921,"significance":"If the central claim held, the paper would provide a useful empirical baseline: on a challenging scientific benchmark, prompt choice can decouple answer accuracy from explanation quality, with self-consistency improving accuracy while lowering explanation similarity. The paper has clear strengths: it uses all 448 GPQA-main questions, reports domain-level breakdowns, gives transparent accuracy tables, and checks distributional assumptions with Shapiro-Wilk tests. However, the headline conclusion currently rests on an unvalidated embedding-similarity measure and an unreported rule for handling missing explanations, so the significance of the empirical finding is conditional on those issues being resolved. The paper does not provide code or human expert validation, but the tabulated measurements are detailed enough to reanalyze.","major_comments":[{"comment":"Section V states that 'For some questions, self-consistency only provided the answers with no explanation,' but Table 2 reports cosine-similarity counts that sum to 448 for every prompt, including self-consistency. The manuscript never states how many explanations were missing or how missing entries entered the cosine similarity computation (excluded, empty string, zero vector, or imputed). Because the central conclusion that reasoning quality does not correlate with accuracy depends on self-consistency's median of 0.645 being lower than the other prompts, the missing-explanation rule is load-bearing. Please report the number of missing explanations per prompt and provide a sensitivity analysis (for example, excluding those questions, or using worst-case and best-case imputation) to show that the ranking is stable.","section":"Section V, Table 2"},{"comment":"The paper equates MPNet cosine similarity between the prompt-induced explanation and the GPQA ground-truth explanation with 'scientific reasoning quality,' but provides no validation that this metric tracks reasoning quality. The authors state that they are non-experts and did not grade explanations, and the metric is plausibly confounded by prompt-induced style: multipath prompting explicitly forces justifications for all four choices, and self-consistency aggregates multiple reasoning paths, so lower cosine similarities may reflect stylistic divergence rather than worse scientific reasoning. Please add human expert ratings on a subset of explanations or a control analysis that compares cosine similarity with surface-form variables (for example, explanation length, number of equations, or template phrasing) to support the construct validity of the measure. Without such evidence, the accuracy-explanation decoupling claim is not established.","section":"Section III.C, Table 2"},{"comment":"Section III claims that setting the knowledge cutoff to December 2023 and the model date to April 1, 2024 'ensure[s] no data leakage,' but GPQA was published in November 2023 and the evaluated snapshot is GPT-4o-2024-08-06. These API parameters do not alter the training data, so they cannot guarantee that GPQA was not seen during training. The interpretation that accuracy reflects reasoning rather than memorization is therefore unsupported. Please add a discussion of contamination risk and, if possible, a quantitative bound (for example, evaluation on a post-cutoff held-out set or a citation to a contamination analysis of the model snapshot).","section":"Section III, leakage control"},{"comment":"The paper interprets R-squared values above 0.99 for the linear fits to cumulative correctness as evidence that 'the model's accuracy does not fluctuate wildly.' For a cumulative sum of per-question Bernoulli outcomes regressed on question index, a high R-squared is expected from the accumulation process itself and is not informative about consistency; the slope being close to the mean accuracy is also definitional. This claim should be removed or reframed, or replaced with a direct measure of variability such as per-subdomain accuracy with confidence intervals or a runs test.","section":"Section V, Fig. 1, Table V"}],"minor_comments":[{"comment":"The abstract reports self-consistency accuracy as 52.99%, but Table 1 gives 237/448 = 52.90%; the same inconsistency appears in Section IV.A. Please correct the abstract and body to match the table.","section":"Abstract, Section IV.A, Table 1"},{"comment":"The abstract says self-consistency performed 'second worst' in explaining answers, while Section IV.B ranks by median and places self-consistency last, and Table 2 shows multipath as lower only by mean. Please state the ranking convention explicitly and reconcile the abstract, the text, and Table 2.","section":"Abstract, Section IV.B, Table 2"},{"comment":"The text says the descending order by median is 'direct answer, CoT, zero-shot CoT,' but the table values are CoT 0.736, direct answer 0.733, and zero-shot CoT 0.730. Reorder the text or note the rounding that makes the order differ.","section":"Section IV.B, Table 2"},{"comment":"Decomposition 'Option 1' is said to have been removed because of hallucination, but the paper does not report how many questions were affected or what the Option 1 results were beyond that statement. A brief quantitative description would help readers assess the removal.","section":"Section IV.A, Section V"}],"recommendation":"major_revision","confidential_remarks":"The empirical measurements appear to be tabulated carefully, but the headline conclusion is not yet robust to the missing-explanation handling or to the unvalidated similarity metric. I would recommend requiring the sensitivity analysis and construct-validity discussion in revision before acceptance. There is also a mismatch between the abstract's numbers and Table 1 that should be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about prompt-engineering baselines: they run seven prompt variants on all 448 GPQA-main questions with GPT-4o, measure both accuracy and MPNet cosine similarity between the model's explanations and GPQA's ground-truth explanations. The specific combination — accuracy plus explanation similarity across these prompts — is genuinely not in the cited GPQA literature, and the multipath prompt is a new variant. The accuracy arithmetic checks out, the full dataset is used, and the paper is honest that the authors are non-experts who did not grade explanations. As a descriptive baseline, it's useful.\n\nThe problem is the load-bearing conclusion. The claim that \"reasoning quality does not always correlate with accuracy\" rests entirely on the similarity rankings in Table 2, and that metric is confounded. Section V says that for some questions self-consistency returned no explanation at all, but the paper never states how those cases entered the cosine computation. Excluding them biases the median to a non-random subset; coding them as zero creates an artifact. Either way the gap between self-consistency (0.645) and the simple prompts (~0.73) is suspect. Also, MPNet similarity probably tracks prompt-induced style rather than reasoning quality: multipath forced justifications for all four choices, and self-consistency aggregates multiple paths, so those outputs are stylistically different by construction. Without expert grading or some other validation, the similarity numbers do not measure \"scientific reasoning.\"\n\nThe abstract has mechanical errors — self-consistency is 52.99% there vs. 52.90% in Table 1, and \"second worst\" vs. worst in Table 2 — and Section V says CoT had the highest similarity while Section IV.B says direct answer did. The R-squared \"stability\" analysis is tautological (cumulative correctness slope is the mean accuracy), and no significance tests separate differences that are often just a few questions. The leakage claim is also shaky: setting a knowledge cutoff in the API does not guarantee the model hasn't seen GPQA, which was public well before the snapshot date.\n\nMy take: the empirical grid is a decent baseline for prompt accuracy on GPT-4o/GPQA, and the accuracy rankings are likely reproducible. But the decoupling headline is not established. If you referee this, ask for the missing-explanation handling, a confound check (e.g., length/style controls), and error bars. It deserves peer review — the question is legitimate and the data are real — but major revision is needed before the reasoning-quality claim can stand.\n\nMy recommendation: send it out, but make clear the explanation-quality analysis needs a redo.","headline":"A useful prompt-accuracy baseline on GPQA, but the explanation-quality metric is too confounded to support the headline decoupling claim.","tokens_in":13659,"tokens_out":2033,"would_cite":true,"duration_ms":21156,"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":"GPT-4o's answer accuracy and explanation quality come apart depending on the prompting technique used, with the most accurate prompt producing the least ground-truth-aligned explanations.","keywords":["large language models","scientific reasoning","prompt engineering","GPQA","GPT-4o","chain-of-thought","self-consistency","explanation similarity"],"falsifier":"Regress the Table 2 cosine similarities on explanation length or on verbatim term overlap with the ground truth: if either predictor alone reproduces the technique rankings, the explanation-quality measure tracks style or coverage rather than reasoning. Alternatively, rephrase GPQA questions without changing content and rerun the prompts: if accuracy on rephrased questions drops well below 52.9%, the 'pattern recognition' interpretation is confounded by memorized benchmark answers.","tokens_in":12504,"feed_emoji":"🧪","tokens_out":6870,"duration_ms":60440,"temperature":0.7,"pith_summary":"This paper tries to establish how GPT-4o's scientific reasoning behaves on the 448-question GPQA-main benchmark under seven prompting techniques, measuring not only how often the model answers correctly but also how closely its explanations resemble expert-written ground-truth explanations. The central finding is a decoupling: self-consistency produces the highest accuracy (52.9%), while the simplest prompts—direct answer, chain-of-thought, and zero-shot chain-of-thought—produce the explanations most similar to the ground truth (median cosine similarity around 0.73). The paper takes this split as evidence that LLMs can appear to reason well on a task while the process behind correct answers is not the same as the process that produces a good scientific explanation. The result matters because it offers a repeatable way to treat accuracy and reasoning quality as separate dimensions when building LLM systems for science, medicine, and law.","feed_headline":"Best accuracy and best explanations come from different prompts","feed_subtitle":"On GPQA science questions, self-consistency wins accuracy while simple prompts best match expert explanations.","key_machinery":"The machinery is a paired measurement: each answer is scored for correctness and, separately, for the similarity of its explanation to the GPQA ground-truth explanation. Similarity is computed as cosine similarity between MPNet sentence embeddings of GPT-4o's prompted explanation and the dataset's expert explanation, grouped into weak, moderate, and strong bands. The paper also fits linear regressions to the cumulative correctness curves and uses R-squared values above 0.99 to argue that the slope of each curve is a stable per-prompt summary. The load-bearing identity is the comparison of the two resulting rankings: the accuracy ranking places self-consistency first, while the explanation-similarity ranking places the three simplest prompts first, and this inversion is the evidence for the paper's claim that reasoning quality and accuracy can part ways.","core_discovery":"The paper's central claim is that on GPQA-main, GPT-4o's answer accuracy and the quality of its scientific reasoning, as measured by the cosine similarity between the model's elicited explanations and GPQA's ground-truth explanations, do not correlate across prompting techniques. In its accuracy ranking, self-consistency leads with 237 out of 448 correct (52.9%), followed by direct answer (52.2%), zero-shot CoT (50%), multipath (48.4%), decomposition (47.8%), self-ask (46.9%), and chain-of-thought (43.8%). In the explanation-similarity ranking, the order is different at the top: direct answer, CoT, and zero-shot CoT reach median similarity values around 0.73 and are classified as strong, while self-consistency has the lowest median similarity (0.645) and is classified as moderate. The paper interprets this separation as evidence that the model often relies on pattern recognition and prompt-induced response habits rather than consistently applying logical inference, and it proposes a research agenda for structured reasoning frameworks, hybrid approaches, and human-in-the-loop verification.","pith_inferences":["A length-controlled re-run of the same seven prompts would test whether the cosine-similarity ranking reflects logical content or simply the verbosity of each prompt's explanations.","The decoupling implies a practical trade-off not explicitly recommended by the paper: deployments that prioritize auditable reasoning might choose direct-answer or CoT prompts, while deployments that prioritize raw accuracy might choose self-consistency.","Question 257, where most prompts refuse to answer because no listed choice matches the model's computed value, offers a natural probe for whether the model's 'pattern recognition' behavior masks a consistent latent computation.","Running the same accuracy-plus-similarity protocol on other frontier models would show whether the decoupling is a GPT-4o-specific artifact or a general property of LLM prompting."],"forward_implications":["Self-consistency achieves the highest accuracy (237/448, 52.9%) yet ranks last in explanation similarity, so maximizing accuracy alone can select the prompt whose explanations least resemble expert reasoning.","Direct answer, CoT, and zero-shot CoT explanations all reach median cosine similarity around 0.73, placing simple single-path prompts closest to the GPQA ground-truth explanations.","Each prompt's cumulative correctness curve is nearly perfectly linear (R-squared above 0.99), so the slope of the curve can serve as a stable summary of a prompt's per-question accuracy.","GPT-4o refuses to answer certain questions under most prompts, notably question 257 where its computed answer is not among the choices, which the paper reads as evidence of internal computation that can override the required choice format.","Accuracy and explanation quality should be reported as separate dimensions when evaluating scientific reasoning, since the two rankings differ."],"supporting_citations":[{"why":"Supplies the 448 GPQA-main questions and the expert-written ground-truth explanations that the paper compares every model explanation against.","marker":"[5]"},{"why":"Introduces chain-of-thought prompting, one of the techniques the paper applies and compares.","marker":"[14]"},{"why":"Introduces self-consistency, the prompt that reaches the highest accuracy in the paper.","marker":"[15]"},{"why":"Provides the MPNet sentence-transformer model whose embeddings produce the cosine-similarity measure of explanation quality.","marker":"[32]"},{"why":"Documents known biases of cosine embedding similarity that the paper must argue do not affect the specialized GPQA explanations.","marker":"[31]"}],"fun_headline_variants":["Accuracy and reasoning quality don't align across LLM prompts","Simple prompts yield better reasoning than complex ones in LLMs","LLM answer accuracy and explanation quality diverge by prompt","Prompting trade-off: right answers vs clear reasoning in GPT-4o"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the cosine similarity between MPNet embeddings of GPT-4o's explanation and GPQA's ground-truth explanation actually measures scientific reasoning quality, together with the assumption that the model's December 2023 knowledge cutoff prevents it from having memorized the GPQA questions and answers.","fun_headline_variants_meta":{"raw":{"variants":["Accuracy and reasoning quality don't align across LLM prompts","Simple prompts yield better reasoning than complex ones in LLMs","LLM answer accuracy and explanation quality diverge by prompt","Prompting trade-off: right answers vs clear reasoning in GPT-4o"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000696,"raw_usage":{"total_tokens":3224,"prompt_tokens":1102,"completion_tokens":2122,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":2063}},"tokens_in":718,"tokens_out":2122,"duration_ms":16158,"temperature":1.0,"reasoning_tokens":2063,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:21:17.641516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Regress the Table 2 cosine similarities on explanation length or on verbatim term overlap with the ground truth: if either predictor alone reproduces the technique rankings, the explanation-quality measure tracks style or coverage rather than reasoning. Alternatively, rephrase GPQA questions without changing content and rerun the prompts: if accuracy on rephrased questions drops well below 52.9%, the 'pattern recognition' interpretation is confounded by memorized benchmark answers.","supporting_citations":[{"cited_title":"Explaining and Improving Contrastive Decoding by Extrapolating the Probabilities of a Huge and Hypothetical LM","cited_arxiv_id":"2411.01610","evidence_quote":"Supplies the 448 GPQA-main questions and the expert-written ground-truth explanations that the paper compares every model explanation against."},{"cited_title":"Chain-of-thought prompting elicits reasoning in large language models,","cited_arxiv_id":null,"evidence_quote":"Introduces self-consistency, the prompt that reaches the highest accuracy in the paper."}],"review_version":1}