{"id":"08628883-aa5a-464e-b54f-db05f617093e","arxiv_id":"2502.03937","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Models sharing training data, algorithms, or foundation weights produce strongly correlated errors, a risk the paper quantifies across three deployment scenarios.","lead":"This paper measures how errors of different machine learning models are correlated when the models share a dataset, share features, or come from the same foundation model. It finds that correlated errors are common, and argues that this raises risk for safety-critical systems that deploy multiple models together.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scenario 3 measures correlation across foundation models, not the same-foundation correlation defined in Definition 2, so the homogenization claim is unsupported by the reported experiment.","rationale":"The reader's weakest assumption correctly identifies that the Scenario 3 experiments do not instantiate Definition 2. I agree this is the most load-bearing concern because it directly undermines the paper's most novel contribution: quantifying homogenization from shared foundation models. The central claim in the abstract explicitly highlights foundational models as a source of substantial aggregated risk, and the only experiment intended to support that portion of the claim estimates a different estimand. The issue is not merely missing error bars or small sample size; it is a mismatch between the defined correlation and the computed correlation. If Definition 2 were properly instantiated, the authors would need multiple fine-tuned models per foundation model, which the current protocol does not provide. I considered whether the lack of a baseline for test-set difficulty in Scenario 1 is equally serious, but that concern affects the interpretation of absolute correlation magnitudes rather than invalidating the reported estimand; the Scenario 3 mismatch makes the foundational-model claim untestable from the reported numbers. The paper does provide a clear formal framework, reproducible code, and honest discussion of limitations, so the appropriate response is to require additional experiments that directly match Definition 2 rather than to reject the work outright. The verdict remains CONDITIONAL, as the reader recommended, pending this reanalysis.","tokens_in":10954,"tokens_out":5869,"duration_ms":56992,"concrete_test":"For each image foundation model F in Table III (e.g., ResNet50), fine-tune R=30 models on CIFAR-10 and R=30 on EUROSAT using independent random seeds and identical fine-tuning protocols. Compute the 30 pairs of average error frequencies (E(f1,F,Z1), E(f2,F,Z2)) and their Pearson correlation within F. Repeat for all 8 image foundation models and, with computational budget, for 2-3 LLMs on financial_phrasebank and emotion-balanced. Report the distribution of within-F correlations alongside the current across-F correlations. If the within-F correlations are not significantly positive (or are no larger than correlations computed across unrelated foundation models), the Scenario 3 evidence in Figures 6-7 does not support the homogenization claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-C3 presents Scenario 3 as evidence for homogenization, but the experiment does not estimate the quantity in Definition 2. Definition 2 is the correlation between the performance metrics E(f1(F),Z1) and E(f2(F),Z2) of two fine-tuned models f1 and f2 derived from the same foundation model F. Figures 6b and 7b instead compute, for a set of different foundation models F (8 image models, 8 LLMs), the Pearson correlation across F of the error frequencies on dataset Zi vs dataset Zj. That is the cross-foundation-model correlation Corr_F(E(f_F,Z1), E(f_F,Z2)), not the within-F correlation between two fine-tuned models. Since each foundation model is fine-tuned only once per dataset, there is no replication from which a within-F correlation could be estimated. Consequently, the reported matrices cannot support the claim that defects in a shared foundation model propagate to correlated errors in downstream models; they only show that different foundation models have similar relative performance across datasets. The homogenization conclusion, a key novel contribution, rests on this mismatch.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines three scenarios in which errors of multiple ML models may be correlated—models trained on the same dataset with different architectures (Scenario 1), models trained on overlapping feature sets (Scenario 2), and fine-tuned models sharing a foundation model (Scenario 3)—and provides formal definitions of error correlation (Definition 1) and performance correlation for fine-tuned models (Definition 2). It then reports empirical correlation matrices for tabular regression (California Housing), image classification (CIFAR-10, EUROSAT, MNIST, Fashion-MNIST), and text classification (financial_phrasebank, twitter-financial-news-sentiment, emotion-balanced, ag_news), using Pearson and ϕK coefficients. The central claim is that aggregate risks are substantial when models share algorithms, training data, or foundation models.","tokens_in":11115,"tokens_out":7924,"duration_ms":80527,"significance":"The paper addresses an important and understudied problem: correlated failures across concurrently deployed ML models. The decomposition into three practical scenarios and the attempt to measure correlations directly are useful contributions, and the availability of code supports reproducibility. If the empirical claims were fully supported, the work would be valuable for AI risk management and portfolio-level analysis of model failures. However, the Scenario 3 experiment does not estimate the quantity in Definition 2, and the single-split point estimates provide thin support for the quantitative conclusions. The framework is a good starting point, but the current evidence does not yet establish the advertised homogenization claim.","major_comments":[{"comment":"The experiment reported in Section IV-C3 does not estimate the correlation defined in Definition 2. Eq. (4) is the correlation, for a fixed foundation model F, between the performance metrics E(f1(F), Z1) and E(f2(F), Z2) of two fine-tuned models derived from that same F. Figures 6b and 7b instead compute, for each pair of datasets, the Pearson correlation across the set of foundation models F of the error frequencies E(f_F, Zi) and E(f_F, Zj). Because each foundation model is fine-tuned only once per dataset, there is no replication from which a within-F correlation between two fine-tuned models could be estimated. The cross-foundation correlation Corr_F(E(f_F,Zi), E(f_F,Zj)) can be large even when downstream errors are independent given F, simply because some datasets are easier than others; it therefore cannot support the abstract's claim that risks are substantial 'when models share ... foundational models' or the homogenization conclusion in Section V. To support this claim, the authors need to fine-tune each foundation model multiple times per dataset (varying seeds or data order) and estimate the within-F correlation of Eq. (4); the text should also specify the probability distribution over f1(F) and f2(F), since Eq. (4) is degenerate if fine-tuning is deterministic.","section":"Section III-A / Section IV"},{"comment":"Definition 1 states that the randomness in Eq. (3) comes from the training and test datasets, but the experiments in Sections IV-C1 and IV-C2 use one fixed random split and one training run per model, then correlate the error vectors over the test set. This estimates a conditional correlation between two realized fitted models, not the marginal correlation over training randomness specified in Definition 1. This distinction matters for the paper's message: two algorithms that are similar on average may still show a different empirical correlation for a particular pair of trained models. The authors should either perform multiple training runs and test splits (with bootstrap or standard-error reporting) and average the resulting correlations, or explicitly redefine the estimand as the correlation of errors of fixed models. The absence of any uncertainty quantification for Figures 1-4, 6b, and 7b is a further concern, since the conclusions are quantitative statements about the magnitude of risk.","section":"Section IV-C2"}],"minor_comments":[{"comment":"The code is described as 'available online at this repository' but no URL is given; please provide a working link.","section":"Section IV"},{"comment":"The subsection heading contains a typo: 'fintuning' should be 'fine-tuning'.","section":"Section IV-C3"},{"comment":"The symbol 'M∫' in the paragraph after Figure 4 should be 'M_f' (or 'Mf') for consistency with the earlier definition.","section":"Section IV-C2"},{"comment":"The first contribution bullet contains a grammar error: 'We introduces' should be 'We introduce'.","section":"Section I"},{"comment":"The reference for the twitter-financial-news-sentiment dataset [36] points to a Hugging Face model URL rather than a dataset citation; please cite the dataset properly.","section":"References"},{"comment":"The notation f1(F) in Definition 2 is awkward because F is a model, not a family; consider writing f_{1,F} or 'fine-tuned from F' to avoid confusion.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The Scenario 3 mismatch identified in the stress-test note is real and is the primary barrier to acceptance. The paper is a reasonable empirical study but currently falls short of its advertised contribution. If revised to estimate Definition 2 with repeated fine-tuning runs and to provide uncertainty quantification, I would be willing to reconsider."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a worthwhile empirical start on an important question, but the headline result about foundation models overreaches because the experiment measures something different from the definition.\n\nWhat's new and good: The paper defines error correlation cleanly and applies it to two concrete scenarios. Scenario 1 (different architectures, same dataset) and Scenario 2 (overlapping features) are straightforward, and the results are sensible: similar models, and models that share predictive features, produce positively correlated errors. The feature-importance check in Scenario 2 is a nice touch. The authors are upfront that the study is empirical and that theoretical bounds would be desirable.\n\nWhere it falls down: Scenario 3. Definition 2 is the correlation between the performance of two fine-tuned models drawn from the same foundation model F. To estimate that, you need multiple fine-tuning runs from the same F, or at least some source of within-F variation. The experiment instead fine-tunes each foundation model once per dataset and then correlates error frequencies across the eight different foundation models (Figures 6b and 7b). That is Corr_F(E(f_F,Z1), E(f_F,Z2)) across F, not Corr(E(f1(F),Z1), E(f2(F),Z2)) for a fixed F. So the matrices do not support the homogenization claim; they show that foundation models with generally higher error on one dataset also tend to have higher error on another. This mismatch sits in the paper's main claimed novelty, and the conclusion's wording about 'same foundational architecture' does not correct it.\n\nOther soft spots are smaller: single random splits, no error bars or repeated runs; only eight foundation models; no significance tests. The code is mentioned but the actual URL is missing. These are fixable.\n\nBottom line: Scenarios 1 and 2 are a decent contribution on their own, and the conceptual framing is useful for risk management. The foundation-model result needs rethinking—either reformulate Definition 2 to match what is actually measured, or run the experiment with repeated fine-tuning from the same checkpoint. I would send this to peer review, not desk-reject, because the topic matters and the fix is clear. A careful referee should focus on Section IV-C3 and ask for error bars and a corrected experiment.","headline":"Useful measurement for same-data and overlapping-feature scenarios, but the foundation-model experiment does not estimate the correlation defined in the paper.","tokens_in":11616,"tokens_out":3149,"would_cite":false,"duration_ms":28085,"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":"Errors of machine learning models are not independent: shared data, algorithms, and foundation models create strong correlations that raise the risk of simultaneous failure.","keywords":["model error correlation","aggregated risk","foundation model homogenization","ensemble diversity","correlated failures","machine learning risk","phi_K coefficient","model portfolio"],"falsifier":"Fine-tune two models from the same foundation model on the same training set and compute the per-input error correlation (Definition 1) on a held-out test set; if that per-input correlation is close to zero while only the aggregated error-frequency correlation (Definition 2) stays high, the homogenization conclusion would not follow. The claim predicts both should be substantial.","tokens_in":10752,"feed_emoji":"📉","tokens_out":4055,"duration_ms":39286,"temperature":0.7,"pith_summary":"This paper argues that errors of different machine learning models are not independent: models that share a training dataset, similar algorithms, overlapping predictive features, or a common foundation model make errors that are strongly correlated. If true, portfolios of AI models deployed together can fail simultaneously more often than individual failure rates suggest, which matters for safety-critical uses such as medical dosing prediction or connected vehicles. The authors define correlation coefficients for model errors, then measure them empirically across tabular, image, and text tasks, finding pervasive positive correlations in all three scenarios they construct. The intended contribution is a framework for quantifying the systemic risk that arises when many models share common building blocks, rather than treating each model's error in isolation.","feed_headline":"Model errors correlate strongly when models share data or base models","feed_subtitle":"Shared training data, similar algorithms, or one foundation model can make many AI models fail at once.","key_machinery":"The central objects are the two correlation coefficients: the per-input error correlation rho between two models (Definition 1) and the aggregate performance correlation Pi between two fine-tuned models built from the same foundation model (Definition 2). The mechanism is that error randomness comes from training and test data, so models that share data, architecture, features, or pretrained weights produce dependent error terms. For regression, Pearson's coefficient is used; for classification, the phi_K coefficient, which captures nonlinear dependence among categorical variables, is used. These coefficients are estimated empirically on real datasets across the three scenarios.","core_discovery":"The central claim is that error correlations between machine learning models are substantial and systematic. Definition 1 formalizes the correlation of per-input error terms between two models, with randomness coming from training and test data, and Definition 2 formalizes the correlation of aggregate performance between two fine-tuned models that share a foundation model. Empirically, on the California Housing dataset, random forest, XGBoost, and generalized additive models show strong positive error correlations, and two neural networks correlate strongly with each other and with linear regression. On CIFAR-10, similar architectures (the two convolutional networks, the two small neural networks) have high correlated errors, while on a financial text task the eight tested large language models, all transformer-based, show widespread correlated errors. When models share highly predictive overlapping features, their errors correlate strongly, and when different models are fine-tuned from the same foundation model on different datasets, their error frequencies across datasets are positively correlated, particularly for text tasks that are semantically related. The authors interpret this as empirical evidence for homogenization: downstream models inherit the foundation model's error patterns, making simultaneous failures more likely.","pith_inferences":["If error correlations persist at scale, model risk management should treat a set of foundation-model-derived systems as a single correlated exposure, analogous to a loan portfolio with a common factor; diversifying across several independent foundation models could become a concrete mitigation lever.","The framework could be extended to tail dependence, measuring the correlation of rare catastrophic errors, which would capture joint failures better than Pearson correlation when error distributions are heavy-tailed.","A testable extension is to predict cross-dataset performance correlation from a quantitative measure of task similarity between the fine-tuning datasets; the paper's ag_news outlier suggests such a relationship exists but is not formalized."],"forward_implications":["A portfolio of models trained on the same public dataset will have a higher-than-independent probability of simultaneous failure, so risk assessment must account for error correlation, not just individual accuracy.","Widespread use of transformer-based foundation models concentrates error correlations: different applications fine-tuned from similar base models can fail in unison.","Organizations that build models from the same highly predictive public features will have correlated errors even if they use different model algorithms.","Fine-tuning many downstream models from one foundation model propagates that model's defect patterns; the resulting homogenization is measurable through correlation of error frequencies across tasks.","Mitigation strategies should actively diversify data sources, model architectures, and foundation models to reduce systemic risk from correlated errors."],"supporting_citations":[{"why":"Supplies the bagging ensemble baseline that motivates the value of error independence, which the paper contrasts with its correlation framework.","marker":"[8]"},{"why":"Provides the diversity measures for classifier ensembles that the paper positions its correlation coefficients against.","marker":"[9]"},{"why":"Defines the phi_K correlation coefficient used for all classification error correlations in the experiments.","marker":"[29]"},{"why":"Introduces the concept of foundation model homogenization that Scenario-3 and Definition 2 build upon.","marker":"[6]"},{"why":"Discusses the algorithmic leviathan and homogenization risks that motivate the foundation-model correlation analysis.","marker":"[7]"},{"why":"ImageNet is one of the public datasets cited as a source of shared-data error correlation in Scenario-1 for vision models.","marker":"[23]"},{"why":"Microsoft COCO is the second public vision dataset cited for the same shared-data correlation mechanism.","marker":"[24]"},{"why":"ResNet architectures serve as the vision foundation models whose fine-tuned variants are compared in Scenario-3.","marker":"[51]"},{"why":"Supplies the interpretation threshold (coefficient above 0.7 is strong) used to describe the empirical correlation magnitudes.","marker":"[64]"}],"fun_headline_variants":["Shared data or base models make AI errors correlate","AI models fail together when trained on same data","Error correlations in ML models tied to sharing","Models with common roots show correlated errors","When models share, their errors align"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Scenario-3 experiments treat the correlation of error frequencies across different datasets and across different foundation models as a valid estimate of Definition 2's correlation between two fine-tuned models that share one foundation model, but the experiments never directly compute the per-input error correlation between two models sharing the same foundation model on the same test set.","fun_headline_variants_meta":{"raw":{"variants":["Shared data or base models make AI errors correlate","AI models fail together when trained on same data","Error correlations in ML models tied to sharing","Models with common roots show correlated errors","When models share, their errors align"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000139,"raw_usage":{"total_tokens":1121,"prompt_tokens":875,"completion_tokens":246,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":180}},"tokens_in":491,"tokens_out":246,"duration_ms":2927,"temperature":1.0,"reasoning_tokens":180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:09:49.701330+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fine-tune two models from the same foundation model on the same training set and compute the per-input error correlation (Definition 1) on a held-out test set; if that per-input correlation is close to zero while only the aggregated error-frequency correlation (Definition 2) stays high, the homogenization conclusion would not follow. The claim predicts both should be substantial.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the bagging ensemble baseline that motivates the value of error independence, which the paper contrasts with its correlation framework."},{"cited_title":"Kuncheva and Christopher J","cited_arxiv_id":null,"evidence_quote":"Provides the diversity measures for classifier ensembles that the paper positions its correlation coefficients against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the phi_K correlation coefficient used for all classification error correlations in the experiments."},{"cited_title":"The algorithmic leviathan: Arbitrariness, fairness, and opportunity in algorithmic decision-making systems","cited_arxiv_id":null,"evidence_quote":"Discusses the algorithmic leviathan and homogenization risks that motivate the foundation-model correlation analysis."},{"cited_title":"Lawrence Zitnick","cited_arxiv_id":null,"evidence_quote":"Microsoft COCO is the second public vision dataset cited for the same shared-data correlation mechanism."},{"cited_title":"Correlation coefficients: Appropriate use and interpretation","cited_arxiv_id":null,"evidence_quote":"Supplies the interpretation threshold (coefficient above 0.7 is strong) used to describe the empirical correlation magnitudes."}],"review_version":1}