{"id":"cbc9ca70-c4b4-41a4-a41b-5f3dadef1fcb","arxiv_id":"2602.21947","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Frontier LLMs are systematically miscalibrated at predicting causal-discovery algorithm performance, achieving only 15.9% calibrated coverage with most models below a random baseline.","lead":"Large language models asked to predict how well causal-discovery algorithms perform miss the true result in most cases—15.9% coverage across eight models. Seven of eight models do worse than uniformly random interval guessing, and the paper's evidence that the best model merely memorized benchmarks is not conclusive.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Memorization attribution for Claude is confounded: synthetic data are linear Gaussian, which favors NOTEARS and penalizes LiNGAM, so the algorithm-specific degradation also matches assumption-aware reasoning.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the paper's memorization inference is confounded by the linear-Gaussian generation of the synthetic datasets. The observed NOTEARS stability and LiNGAM collapse are predicted by assumption-aware reasoning, so they cannot by themselves distinguish memorization from reasoning. I agree with the reader's assessment that the central quantitative finding (low calibrated coverage) is well supported, but the stronger claim about Claude's performance being 'attributable to benchmark memorization' is not established by the current experimental design. The paper's own limitations section explicitly concedes this. The proposed concrete test—varying the noise distribution while holding the benchmark/synthetic distinction—would directly settle whether the algorithm-specific degradation is driven by data-generation/algorithm-assumption match or by memorization. Because the reader already reached CONDITIONAL on essentially this basis, I recommend no change to the verdict; the paper should be accepted only with the memorization claim either softened or re-tested on non-confounded synthetic data, and with code/data release for reproducibility.","tokens_in":17077,"tokens_out":3693,"duration_ms":37096,"concrete_test":"Generate a matched set of synthetic datasets with the same DAGs and sample sizes but non-Gaussian noise (e.g., exponential or uniform), and rerun the identical protocol for Claude and ideally all models. If LiNGAM's synthetic coverage collapse disappears or reverses under non-Gaussian data—while NOTEARS's boost weakens—the algorithm-specific pattern tracks data-generation/algorithm-assumption match, not benchmark memorization. Under the paper's memorization account, LiNGAM should still collapse and NOTEARS should still boost regardless of noise distribution. Add a second condition with nonlinear generating mechanisms (e.g., sine or logistic links) where NOTEARS's linear objective is misspecified; reasoning predicts NOTEARS degradation on such data, whereas memorization predicts no change. Compare Claude's benchmark-to-synthetic coverage differences across these conditions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's second central claim—that Claude's 39.4% coverage is 'attributable to benchmark memorization rather than principled reasoning' (Abstract; §4.1)—rests on the algorithm-specific synthetic degradation pattern: LiNGAM collapses (−23.2% across all models; −16.0% for Claude) while NOTEARS does not (+1.7% across all models; +24.3% for Claude) (§4.1 Table 3; §4.4 Table 5). The authors interpret this dissociation as 'the hallmark of pattern matching against benchmark statistics' (§4.1). However, §3.1 states that all four synthetic datasets are generated from random DAGs with 'linear Gaussian data generating processes.' Under that generative model, NOTEARS's optimization objective (Eq. 8) is exactly the least-squares/linear-Gaussian likelihood, whereas LiNGAM's identifiability argument explicitly requires non-Gaussian noise (Eqs. 5–6; Appendix A). A model that genuinely reasoned from algorithm assumptions would therefore predict NOTEARS to improve and LiNGAM to degrade on synthetic data. The observed dissociation is equally consistent with principled reasoning about algorithm-data assumptions and with benchmark memorization. Section 6 concedes that the memorization inference 'relies on indirect behavioral signals ... rather than direct training data attribution, so we cannot rule out alternative explanations.' Since the headline includes the memorization attribution, the strongest version of the claim is not supported by the reported experiments. The core negative finding—15.9% calibrated coverage and seven models below the random baseline—is not threatened by this confound, but the memorization attribution is.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates whether eight frontier LLMs can provide calibrated interval predictions for the performance of four causal-discovery algorithms on thirteen datasets. Ground-truth means are obtained from 5,200 bootstrap algorithm runs, and LLM ranges are elicited with three prompt formulations per condition, yielding 1,664 coverage comparisons. The primary empirical finding is a mean calibrated coverage of 15.9%, with seven of eight models below a uniform-random baseline of 36.5%; the best model, Claude, reaches 39.4% but the authors attribute this edge to benchmark memorization rather than reasoning. The paper proposes the term 'algorithmic blindness' for the observed failure.","tokens_in":17410,"tokens_out":7429,"duration_ms":74799,"significance":"If the primary coverage finding is accepted, this is a useful and timely negative result: it suggests that current LLMs cannot serve as calibrated zero-shot performance predictors for causal-discovery algorithm selection. The evaluation is unusually explicit about protocol: multiple prompt formulations, external ground truth from repeated runs, and uninformed baselines. No fitted parameters are recycled into the evaluation, and the aggregate coverage result is internally consistent. However, the paper's strongest interpretive claim—that Claude's above-random coverage is memorization—is not supported by the reported experiments, because the synthetic-data design is confounded with the algorithms' own identifiability assumptions. The primary contribution can stand after reanalysis or softening of the memorization claim, but the current abstract overstates what the data establish.","major_comments":[{"comment":"The synthetic datasets are generated from linear Gaussian DAGs (Section 3.1). Appendix A states that NOTEARS's optimization objective in Eq. (8) is exactly the linear-Gaussian least-squares problem, while LiNGAM's identifiability requires non-Gaussian noise (Eqs. 5–6). The observed dissociation—LiNGAM collapses on synthetic data (−23.2% across all models; −16.0% for Claude) while NOTEARS does not (+1.7%; +24.3% for Claude)—is therefore exactly what assumption-aware reasoning would predict. It cannot be used as 'the hallmark of pattern matching against benchmark statistics' (§4.1) nor as support for the Abstract's claim that Claude's 39.4% is 'attributable to benchmark memorization.' Section 6 concedes that alternative explanations cannot be ruled out. This load-bearing inference should either be removed from the headline or tested with non-Gaussian synthetic data and/or direct training-d","section":"§4.1, §4.4, Appendix A"},{"comment":"The three memorization probes are confounded with graph size and algorithmic difficulty. The benchmark-versus-synthetic comparison mixes 5–70-node benchmarks with 12–60-node synthetic graphs, so the observed 2.26× range-width compression and the 2.6× cross-model disagreement gap partly reflect that larger graphs have more possible structures and a wider SHD range. The monotonic 15× collapse in cross-model agreement with synthetic network size is not decisive against principled reasoning: any reasonable estimator would show larger uncertainty and wider disagreement on larger, more complex instances. The authors should supply size-matched controls, or normalize by graph order/edge density and SHD scale, before interpreting these patterns as memorization signals.","section":"§4.5, Tables 7–8, Figures 5–6"},{"comment":"The random baseline is not a no-reasoning floor: a predictor that always returns the full valid range for every metric would trivially achieve 100% coverage while carrying no information. The claim that seven models are 'worse than random guessing' is only meaningful under a scoring rule that trades coverage against interval width. The 'Mean Score' columns in Tables 1 and 4 are never defined, nor is any proper interval score (e.g., Winkler score or CRPS) reported. Please define the mean score and either report width-matched baseline comparisons or restrict the random-baseline claim to the specific uniform-interval construction described.","section":"§3.4, Tables 1 and 4"}],"minor_comments":[{"comment":"The aggregation step averages lower and upper bounds across the three prompt formulations, which can create an interval that no single prompt produced. Reporting per-formulation coverage in addition to the aggregated coverage would clarify how much of the result depends on this arithmetic averaging.","section":"§3.3"},{"comment":"Several reference entries contain formatting errors (e.g., 'V owels' in the D'ya like DAGs entry; 'Bühlman' should be 'Bühlmann'). The 'and 1 others' placeholders should be resolved with full author lists.","section":"References"},{"comment":"No code or data release is mentioned. Given the study is empirical and prompt-sensitive, releasing the exact prompts, raw model outputs, and ground-truth tables would substantially strengthen the paper's contribution.","section":"Reproducibility"},{"comment":"The statement that models with true understanding 'would provide consistent estimates regardless of minor phrasing variations' is stronger than warranted. Prompt sensitivity alone is not a direct measure of understanding; it is compatible with genuine competence plus sensitivity to surface form. The evidence of instability is interesting, but the interpretive language should be more cautious.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is valid and lands squarely: the memorization attribution in the abstract and §4.1 is confounded by the linear-Gaussian synthetic data, which favor NOTEARS and violate LiNGAM. The aggregate negative coverage result is a reasonable empirical contribution and can likely be published after weakening the memorization claim or adding non-Gaussian controls. The undefined 'Mean Score' and the non-floor random baseline also need attention before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core empirical finding is solid and worth taking seriously: across 1,664 comparisons, eight frontier LLMs produce intervals that contain the true algorithmic mean only 15.9% of the time, and seven of eight do worse than a random baseline. That is a genuinely new and important negative result for anyone thinking about LLM-assisted algorithm selection. Second, the paper's secondary claim—that the best model's marginal above-random performance is “attributable to benchmark memorization rather than principled reasoning”—is not established by the experiments as designed. The synthetic datasets are linear Gaussian, which satisfies NOTEARS's objective and violates LiNGAM's identifiability condition. So the observed dissociation, with LiNGAM collapsing and NOTEARS improving, is exactly what assumption-aware reasoning would predict. The authors even note this in Appendix A, and then in Section 6 concede that their memorization inference relies on indirect behavioral signals. The claim may be true, but the paper's current evidence does not rule out the more parsimonious explanation that some models reason from algorithm-data assumptions.\n\nWhat the paper does well: the methodology is transparent, the baselines are sensible, the prompt-robustness analysis is a good practice, and the aggregate tables are consistent with the headline coverage figure. The distinction between declarative knowledge and calibrated procedural prediction is a useful framing, and the “algorithmic blindness” label, though a bit grandiose, points at a real phenomenon. The paper is also honest about limitations—it flags the indirectness of the memorization signal and the correlation among metrics.\n\nWhere it's softer: the memorization attribution is the main weakness, and it is load-bearing for the abstract's strongest sentence. A simple fix would be to drop or weaken that claim and instead report the algorithm-specific dissociation as an open question. It would also help to report a significance test for Claude versus the random baseline, since the 2.9-point gap is currently described as not meaningful but no test is given. Minor issues: the random baseline's 36.5% is high, which is explained but still worth a sensitivity check; and the paper treats metrics as independent without much discussion of how that affects coverage interpretation.\n\nThis paper deserves a serious referee. The central negative result is important, the experimental design is reproducible in principle, and the weak spot is a well-defined interpretational flaw, not a structural one. I would send it out, with the expectation that the memorization claim be revised or re-tested on non-Gaussian synthetic data. For a reading group, it would generate a good methodological discussion.","headline":"Strong negative result on LLM coverage, but the paper's strongest claim—that Claude's edge is memorization—is not supported by the design.","tokens_in":17874,"tokens_out":1266,"would_cite":true,"duration_ms":13966,"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":"Frontier LLMs cannot produce calibrated predictions of causal discovery algorithm performance, scoring 15.9% calibrated coverage with most models below random guessing.","keywords":["algorithmic blindness","calibrated coverage","large language models","performance prediction","causal discovery","algorithm selection","memorization","uncertainty calibration"],"falsifier":"Generate held-out synthetic datasets with a linear non-Gaussian data-generating process (satisfying LiNGAM's identifiability condition) and with a nonlinear Gaussian process (challenging NOTEARS's linear objective), rerun the 52-condition calibration protocol, and check whether LiNGAM's synthetic coverage collapse persists and NOTEARS's boost disappears. If coverage tracks assumption fit rather than benchmark fame, the memorization inference is falsified.","tokens_in":16961,"feed_emoji":"📉","tokens_out":5121,"duration_ms":46158,"temperature":0.7,"pith_summary":"This paper asks whether large language models can predict how well a causal discovery algorithm will perform on a given dataset, in the form of calibrated ranges. Across 1,664 comparisons with eight frontier LLMs, the answer is no: the models' predicted intervals contain the true algorithmic mean only 15.9% of the time, and seven of eight models do worse than a random-range baseline. The single model that edges past random (Claude at 39.4%) appears to do so by recalling benchmark-associated statistics rather than by reasoning about algorithm-data interactions. The authors name this failure 'algorithmic blindness' and argue it reflects a structural gap between declarative knowledge about algorithms and calibrated procedural prediction. If correct, the result rules out zero-shot LLMs as trustworthy uncertainty-aware algorithm selectors and motivates empirical evaluation or learned performance predictors instead.","feed_headline":"LLMs miss true algorithm performance 84% of the time","feed_subtitle":"Seven of eight frontier models score worse than random guessing; the best model's edge looks like benchmark recall.","key_machinery":"The central object is 'calibrated coverage' (Eq. 3): for each model, dataset, algorithm, and metric, whether the LLM's aggregated predicted range [l, u] contains the empirical algorithmic mean from 100 bootstrap runs of the algorithm. The paper builds its argument around this interval-coverage metric plus three behavioral memorization probes: range-width compression (tighter predictions on familiar benchmarks than on synthetic data), cross-model agreement collapse (models agree on benchmarks but diverge on synthetic data), and algorithm-specific degradation (LiNGAM uniquely collapses on synthetic data). Together these are intended to separate genuine reasoning from retrieval of benchmark-ass","core_discovery":"The central claim is that frontier LLMs are algorithmically blind: they cannot translate their verbal knowledge of algorithmic assumptions into numerical expectations about empirical performance. Using causal discovery as a testbed, the paper measures calibrated coverage, the fraction of cases where a model's stated range contains the true mean from 100 bootstrap algorithm runs, and finds 15.9% mean coverage across 1,664 model-dataset-algorithm-metric comparisons. Seven of eight models fall below a uniform random baseline of 36.5%, and the best model's 39.4% coverage is marginally above random and attributed to memorization of benchmark statistics rather than principled reasoning. The wide-b","pith_inferences":["The synthetic-data evidence for memorization is confounded: the held-out synthetic graphs are linear Gaussian, which satisfies NOTEARS's optimization objective and violates LiNGAM's non-Gaussian identifiability condition. LiNGAM's synthetic collapse and NOTEARS's stability are exactly what assumption-aware reasoning would predict, so the dissociation may not demonstrate memorization.","A cleaner test of the memorization hypothesis would hold algorithmic assumptions fixed: generate synthetic data that respects each algorithm's identifiability conditions (non-Gaussian for LiNGAM, linear Gaussian for NOTEARS) and check whether collapse persists.","The 36.5% random baseline is inflated by the wide valid metric domains; a more decision-relevant baseline would be a heuristic that always guesses the unconditional mean or median range from training data, which the paper shows scores near Claude.","The paper's protocol (bootstrap ground truth, three prompt formulations, aggregated coverage) is directly transferable to other algorithm-selection tasks; running the same 1,664-comparison calibration test on SAT solvers or graph algorithms would show whether algorithmic blindness is a general LLM property or specific to causal discovery's statistical nuance."],"forward_implications":["Practitioners should not use zero-shot LLM intervals to choose among causal discovery algorithms; a uniform random range is better calibrated than most frontier models.","Simple dataset-level heuristics outperform seven of eight models, so cheap statistical baselines should be the reference floor for any LLM-based algorithm selector.","The coverage failure is not a confidence-calibration artifact: intervals are far too wide and still miss the true mean, implying absence of a performance model rather than miscalibrated uncertainty.","The memorization account predicts that above-random performance should concentrate on algorithms and metrics with dense benchmark literature; new or obscure algorithms should show the largest degradation.","If algorithmic blindness is general, LLM-assisted algorithm selection in other domains (optimization, SAT solving) should be treated as unvalidated until the same calibration test is run."],"fun_headline_variants":["Algorithmic blindness: LLMs fail 84% on performance prediction","Frontier LLMs worse than random at algorithm calibration","LLMs can't gauge algorithm performance: 84% miss rate","Causal discovery exposes LLM algorithmic blindness","Seven of eight LLMs underperform random on algorithm metrics"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's claim that the best model's edge is memorization, not reasoning, rests on the assumption that genuine algorithmic reasoning would show similar synthetic-data performance across all four algorithms; but the synthetic data are linear Gaussian, which happens to match one algorithm's objective and violate another's, so the observed pattern does not distinguish reasoning from memorization.","fun_headline_variants_meta":{"raw":{"variants":["Algorithmic blindness: LLMs fail 84% on performance prediction","Frontier LLMs worse than random at algorithm calibration","LLMs can't gauge algorithm performance: 84% miss rate","Causal discovery exposes LLM algorithmic blindness","Seven of eight LLMs underperform random on algorithm metrics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000293,"raw_usage":{"total_tokens":1495,"prompt_tokens":646,"completion_tokens":849,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":390,"completion_tokens_details":{"reasoning_tokens":778}},"tokens_in":390,"tokens_out":849,"duration_ms":8230,"temperature":1.0,"reasoning_tokens":778,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:50:51.475484+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate held-out synthetic datasets with a linear non-Gaussian data-generating process (satisfying LiNGAM's identifiability condition) and with a nonlinear Gaussian process (challenging NOTEARS's linear objective), rerun the 52-condition calibration protocol, and check whether LiNGAM's synthetic coverage collapse persists and NOTEARS's boost disappears. If coverage tracks assumption fit rather than benchmark fame, the memorization inference is falsified.","supporting_citations":[],"review_version":2}