{"id":"827378b2-9245-435e-89ec-e8c179023841","arxiv_id":"2501.13215","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"On a fine-grained French election retweet dataset, the multi-state voter model with zealots recovers individual users' party affiliations with 92.5% accuracy.","lead":"A simple opinion dynamics model, applied to retweets from the 2017 French election, matched the party affiliation of individual Twitter users 92.5% of the time. The result suggests that basic influence models can capture user-level opinions, not just aggregate trends.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth labels may have been annotated from the same retweet content that defines the network, making the 92.5% argmax accuracy a potential label-leakage artifact rather than evidence for the voter model.","rationale":"The central claim is that the equilibrium opinion distribution of the voter model recovers individual party affiliations. For that claim to serve as evidence of the model's validity, the ground-truth labels must be independent of the retweet data that defines the network. The manuscript does not establish this independence: it reports only that labels were assigned from 'the content of their tweets and profile description' (Section II), and retweets are tweets. The strongest evidence for the model — 92.5% argmax accuracy and 95.1% accuracy from direct zealot connections — is also the expected signature of label leakage if annotators used retweeted accounts. This concern is more fundamental than the reader's list of missing baselines and tie-breaking rules: those affect the interpretation of the number, whereas annotation leakage affects whether the number measures what the paper claims. The proposed test is feasible: the dataset is public and the annotation protocol is in the cited dataset paper; if retweets were excluded, the concern is resolved and the paper's conditional acceptance stands. I therefore keep the reader's CONDITIONAL verdict unchanged, with the annotation-independence check as the primary condition.","tokens_in":3,"tokens_out":18932,"duration_ms":397354,"concrete_test":"Obtain the annotation protocol for the Elysée2017fr dataset (refs [36,39]) and determine whether annotators saw retweets or retweeted accounts. If retweets were not excluded, re-annotate a random sample of at least 500 users using only original (non-retweet) text and profile descriptions, blind to network data; recompute the argmax accuracy of the voter model against these cleaned labels. If accuracy falls substantially (e.g., toward the majority baseline or below the reported z_i baseline), the central claim is an annotation-leakage artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II states that accounts were manually labelled on the basis of 'the content of their tweets and profile description.' On Twitter, retweets are tweets. The features feeding Eq. (1) — especially the zealot weights z_i^s and the user-to-user retweet weights w_ij — are computed from exactly this retweet stream. If annotators saw which political entities a user retweeted (or the retweeted content itself), then the ground-truth label y_i is not independent of the model's input features, even though y_i is not an input to Eq. (1). The paper's claim that 'the ground-truth opinion y_i of i does not bear involvement in the computation of x_i' addresses algorithmic circularity, not annotation leakage. The robustness result in Section IV.A — direct zealot retweet weights z_i alone already reach 95.1% argmax accuracy on the 73% of users with such connections — is exactly what one would expect if annotators used retweeted party accounts to assign labels. Without evidence that the annotation protocol excluded retweet content, the 92.5% accuracy may largely reflect the annotators' use of the same signal the model reads.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper tests the multi-state voter model with zealots at the individual-user level. Using a retweet network from the #Elysée2017fr dataset, it defines political-entity accounts as zealots with fixed party opinions and computes each user's equilibrium opinion distribution x_i by solving Eq. (1). The most likely opinion argmax(x_i) is then compared with the ground-truth party label y_i, yielding 92.5% agreement; discord probabilities ρ_ij are used to classify same-party vs. cross-party pairs with about 93% accuracy. Robustness checks compare against direct zealot connections z_i, other interaction networks (follow, mention), and unweighted/undirected variants of the retweet network. The paper argues that the parameter-free equilibrium computation provides an external, user-level validation of the voter model.","tokens_in":11815,"tokens_out":6033,"duration_ms":66151,"significance":"If the empirical claim holds, this is a valuable user-level test of an opinion dynamics model, demonstrating that a parameter-free equilibrium computation contains individual-level information, not just aggregate statistics. Strengths: no fitted parameters enter Eq. (1); ground-truth labels are not inputs to the equilibrium computation; the comparison across network definitions is informative; and the theoretical foundation is a published derivation. However, the evaluation currently lacks baselines, out-of-sample classifier evaluation, and protection against label leakage, so the magnitude of the claimed accuracy cannot yet be taken at face value.","major_comments":[{"comment":"The accuracy figures 92.5% (argmax), 93% (SVC), and 93% (logistic regression on ρ_ij) are never compared against a trivial baseline. With five parties, random guessing gives about 20%, but the paper itself notes that some parties have more supporters than others, so the majority-class baseline may be substantially higher. The comparison to z_i in Section IV.A is restricted to the 73% of users with direct zealot connections and does not substitute for a full-population baseline. Please report, on the full evaluation set, at least the majority-class accuracy and a label-shuffled null accuracy, together with the class proportions; this is necessary to interpret 92.5% as evidence for the model.","section":"Section III.A and Section III.B"},{"comment":"The SVC and logistic regression accuracies are computed on the same data used to fit the classifiers (\"fitting a basic SVC to the (x_i, y_i) pairs\", with sklearn default parameters). No train/test split or cross-validation is reported, so these numbers are in-sample and can overstate separation, especially with the default RBF kernel on about 15,000 users. Please report cross-validated or held-out accuracies for both classifiers. If the goal is prediction, the out-of-sample numbers are the relevant ones; if the goal is separation, say so and use a separation measure that does not involve fitting.","section":"Section III.A, Section III.B, Supplemental Material I"},{"comment":"The ground-truth labels were assigned by the dataset creators \"on the basis of the content of their tweets and profile description.\" Since the retweet network and z_i^s are constructed from retweets, and retweets are part of tweet content, the annotation protocol may have used the same signal that the model reads. The statement that \"the ground-truth opinion y_i of i does not bear involvement in the computation of x_i^s\" addresses algorithmic circularity but not label leakage. The robustness result in Section IV.A, where z_i alone reaches 95.1% accuracy on the 73% of users with direct zealot connections, is also consistent with annotators having used retweeted party accounts to assign labels. Please clarify the annotation protocol (for example, whether annotators saw retweet content or only profile descriptions) or validate on labels obtained independently of retweet content; without this, the 92.5% accuracy may partly reflect the annotators' use of the retweet signal rather than the voter model's dynamics.","section":"Section II (label annotation) and Section IV.A"},{"comment":"The definition of argmax accuracy counts a user as correct if the ground-truth party is among the coordinates attaining the maximum of x_i. This tie-breaking rule can inflate the reported accuracy. Please report the fraction of users with multiple argmax coordinates and recompute the headline 92.5% under a deterministic rule (for example, random tie-break averaged over seeds, or lowest-index tie-break). The same issue applies to the z_i baseline comparisons in Section IV.A. Without this, the 92.5% figure is not uniquely defined.","section":"Supplemental Material I"}],"minor_comments":[{"comment":"The word \"predict\" is used although the evaluation is retrospective classification of already observed party affiliations; consider using \"identify\" or \"retrieve\" consistently to avoid overclaiming temporal prediction.","section":"Title and Abstract"},{"comment":"The text says the SVC finds \"the best separating hyperplane,\" but sklearn's default SVC uses an RBF kernel, which does not correspond to a hyperplane in the original five-dimensional space; please specify the kernel and, if a hyperplane is intended, use a linear SVC.","section":"Section III.A"},{"comment":"It is not stated explicitly whether the reported ρ_ij values are computed with the exact formula Eq. (5) or the approximation Eq. (6); the supplemental material compares the two, but the main text should state which one is used for the logistic regression results.","section":"Section III.B"},{"comment":"Accuracy figures are reported without confidence intervals or significance tests; given the large sample sizes, even small differences may be statistically significant, so binomial confidence intervals or a paired test would help assess the network comparisons in Table I.","section":"Throughout"},{"comment":"There are minor typographical issues, including \"embedds\" in Section III.A and \"muti-state\" in the Discussion; the title also begins with \"V oter model\" with an apparent spacing artifact.","section":"Section III.A and Discussion"}],"recommendation":"major_revision","confidential_remarks":"The central empirical result depends on labels from a third-party dataset whose annotation protocol is not described in this manuscript. Before publication, I would ask the author to clarify or reproduce the annotation instructions, or to validate on an independently labelled subsample. The heavy self-citation of [11] is legitimate because it supplies the theory and algorithms, but the manuscript should be explicit about the extent to which the computational methods come from that paper; the empirical evaluation is the novel contribution. The paper fits the scope of the journal, but the evaluation standards need to be raised."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean, parameter-free user-level test of the voter model on a 15k-node retweet network, and the 92.5% argmax match is real as a computation. The paper's main strength is that it goes beyond aggregate distribution fitting and shows the equilibrium probabilities carry user-level signal, and it does the work of comparing retweet vs follow vs mention networks. But I'd be careful about the headline 'predict' claim: the evaluation has a serious potential label-leakage problem, and the paper omits baselines that would tell you how much of that 92.5% is just network assortativity.\n\nThe leakage issue is the one to focus on. The ground-truth labels were assigned by the dataset's authors 'on the basis of the content of their tweets and profile description.' Retweets are tweets. The model's features are precisely the retweet weights, including the direct zealot weights z_i^s. If the annotators saw which party accounts a user retweeted, then the label is not independent of the model's input. The paper argues that y_i is not used in computing x_i, which is true but addresses algorithmic circularity, not annotation leakage. The robustness result in Section IV.A does not dispel this: direct zealot connections alone reach 95.1% accuracy on the 73% of users with such connections. That is exactly what you'd expect if labels were inferred from retweeted party accounts. To fix this, the authors need to either produce evidence about the annotation protocol (e.g., annotators only used profile text, not retweets) or show results on a subset of users whose labels were independently verified through a different channel.\n\nBeyond that, the evaluation is under-specified. There's no chance or majority baseline; given the label distribution, predicting the most frequent party would already get you a substantial chunk. The SVC and logistic regression accuracies appear to be in-sample with default parameters—no train/test split is reported. The argmax accuracy uses a tie-breaking rule that counts a match if any of the argmax coordinates matches ground truth, which inflates accuracy. And 'predict' overstates what is a contemporaneous reconstruction from the same six-month window.\n\nThe math itself is fine: solving the linear system Eq. 1 is straightforward, the discord probability formulas are left to prior work, and the paper is honest that direct zealot connections carry much of the signal. The network comparisons are useful and support the claim that directionality and weights matter.\n\nBottom line: this deserves a serious referee, not a desk reject, but it needs major revision. I'd send it out with clear requests for an annotation-leakage section, a majority baseline, proper cross-validation, and a conservative tie-breaking rule. As it stands, I wouldn't cite the 92.5% number without a caveat.","headline":"Clean user-level test of the voter model, but the 92.5% accuracy claim is vulnerable to label leakage from retweet-based annotation and lacks basic baselines.","tokens_in":12378,"tokens_out":2719,"would_cite":false,"duration_ms":28303,"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":"The paper claims that a multi-state voter model with zealots, applied to a retweet network around the 2017 French election, predicts individual party affiliation for 92.5% of users from equilibrium opinion distributions alone.","keywords":["voter model","opinion dynamics","zealots","retweet network","political leanings","Twitter","2017 French presidential election","discord probabilities"],"falsifier":"Rewire the non-zealot retweet edges while preserving each user's total retweet weight and all zealot weights, recompute $x_i$, and compare argmax accuracy: if it stays near 92.5%, the prediction is carried by direct attachment to party accounts and the multi-step voter mechanism is not load-bearing; if it drops sharply, the user-to-user propagation is essential.","tokens_in":11391,"feed_emoji":"🗳️","tokens_out":8896,"duration_ms":81323,"temperature":0.7,"pith_summary":"This paper asks whether the voter model, a standard opinion-dynamics model, can predict the political opinion of individual users rather than only reproducing aggregate opinion distributions. Treating political party accounts as immutable zealots and retweet frequencies as influence weights, the author computes each user's equilibrium opinion distribution $x_i$ from a Twitter dataset built around the 2017 French presidential election. The model's most likely opinion matches the user's stated party affiliation for 92.5% of users, and the probability that two users disagree separates same-party from cross-party pairs with about 93% accuracy. If the result holds, user-level political information is encoded in the equilibrium of a simple, parameter-free interaction model, making interaction data alone sufficient to locate individuals in opinion space.","feed_headline":"Voter model predicts 92.5% of individual party leanings","feed_subtitle":"A parameter-free equilibrium opinion score from retweet structure matches stated party labels and identifies like-minded pairs.","key_machinery":"The central object is the individual equilibrium opinion distribution $x_i = (x_i^s)_{s \\in S}$ of the multi-state voter model with zealots on a directed weighted network, defined by the linear system $$x_i^s = \\sum_{j \\in \\mathcal{N}} w_{ij} x_j^s + z_i^s,$$ where $w_{ij}$ is the fraction of user $i$'s retweets originating from $j$ and $z_i^s$ is the aggregated weight of retweets from party-account zealots of opinion $s$. The companion object is the discord probability $\\rho_{ij}$, the equilibrium probability that users $i$ and $j$ hold different opinions. Together these quantities turn the parameter-free voter model into a user-level opinion score and a pairwise agreement score, computed without using the users' stated labels.","core_discovery":"On a directed, weighted retweet network built from six months of Twitter activity around the 2017 French presidential election, the paper solves the multi-state voter model's equilibrium equations and compares each user's equilibrium opinion distribution $x_i$ with the annotated party affiliation $y_i$. The most likely opinion under the model matches the annotated party for 92.5% of the 15,607 non-zealot users considered, with per-party accuracy ranging from 81% (PS) to 95% (FN). Within-party values of $x_i^s$ are concentrated near 1 and cross-party values near 0, supporters of the same party cluster in the opinion space spanned by the $x_i$ vectors, and discord probabilities $\\rho_{ij}$ let a logistic classifier distinguish same-party from cross-party pairs with about 93% accuracy. The model achieves this without parameter fitting and without using the users' own labels; the only fixed inputs are retweet weights and the party affiliations of political-entity accounts.","pith_inferences":["If retweets mostly signal endorsement rather than causal influence, the high accuracy shows that retweet networks are strongly sorted by party, and Eq. (1) acts as a semi-supervised label-propagation rule; the paper does not fully separate these readings.","The direct-zealot baseline $z_i$ nearly matches the full model's accuracy (95.1% versus 96.2% on the subset with direct zealot links), suggesting that most of the signal travels through one-step links to party accounts; a rewiring null model could quantify the added value of the multi-step mechanism.","The same pipeline could be extended to non-political stances or to datasets with temporally varying stated opinions, where treating some accounts as immutable zealots is less straightforward.","The discord-probability result could be used as a structure-only method to detect like-minded communities without declared labels."],"forward_implications":["The equilibrium vector $x_i$ provides a party-affiliation score for every user, including the 27% who have no direct retweet connection to a political-entity account.","Discord probabilities $\\rho_{ij}$ give a usable pairwise agreement measure: a logistic classifier on $\\rho_{ij}$ identifies same-party pairs with about 93% accuracy, and cross-party pairs are easier to identify than same-party pairs.","Directed, weighted retweet interactions are the appropriate network substrate; unweighted or undirected versions and follow or mention networks all give lower accuracy.","The parameter-free nature of the method means it can be applied directly to any interaction network with identifiable zealot nodes and edge weights.","The results extend the empirical validation of the voter model from aggregate opinion distributions to individual-level prediction in a heterogeneous online population."],"supporting_citations":[{"why":"Supplies the multi-state voter model equations for equilibrium opinion distributions and discord probabilities on directed weighted networks.","marker":"[11]"},{"why":"Establishes existence and uniqueness of the equilibrium, and the random-walk interpretation of $x_i$ as the probability of hitting a zealot of each opinion.","marker":"[12]"},{"why":"Provides the #Elysée2017fr dataset of retweets, mentions, and manual party labels used for the ground-truth comparison.","marker":"[36]"},{"why":"Documents the labeling procedure and inter-annotator agreement for the ground-truth opinions.","marker":"[39]"},{"why":"Establishes political homophily in retweet networks, motivating the robustness check that separates direct zealot connections from the full voter-model mechanism.","marker":"[41]"},{"why":"Prior empirical validation of the voter model at the aggregate level (US election vote shares), which this paper extends to individual users.","marker":"[26]"}],"fun_headline_variants":["Voter model nails 92.5% of user leanings","Retweet structure predicts party leanings 92.5% accurately","Voter model: parameter-free predictor of political opinions","92.5% accuracy: voter model maps Twitter opinions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fraction of retweets directed from $j$ to $i$ measures how much $j$'s opinion influences $i$, and that political-entity accounts can be treated as immutable zealots with known opinions.","fun_headline_variants_meta":{"raw":{"variants":["Voter model nails 92.5% of user leanings","Retweet structure predicts party leanings 92.5% accurately","Voter model: parameter-free predictor of political opinions","92.5% accuracy: voter model maps Twitter opinions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000463,"raw_usage":{"total_tokens":2276,"prompt_tokens":865,"completion_tokens":1411,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":1350}},"tokens_in":481,"tokens_out":1411,"duration_ms":10768,"temperature":1.0,"reasoning_tokens":1350,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:21:30.983958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rewire the non-zealot retweet edges while preserving each user's total retweet weight and all zealot weights, recompute $x_i$, and compare argmax accuracy: if it stays near 92.5%, the prediction is carried by direct attachment to party accounts and the multi-step voter mechanism is not load-bearing; if it drops sharply, the user-to-user propagation is essential.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the multi-state voter model equations for equilibrium opinion distributions and discord probabilities on directed weighted networks."},{"cited_title":"Pansanella, A","cited_arxiv_id":null,"evidence_quote":"Provides the #Elysée2017fr dataset of retweets, mentions, and manual party labels used for the ground-truth comparison."},{"cited_title":"Constrained Expectation-Maximisation for inference of social graphs explaining online user-user interactions","cited_arxiv_id":"2301.10646","evidence_quote":"Documents the labeling procedure and inter-annotator agreement for the ground-truth opinions."},{"cited_title":"V azquez and V","cited_arxiv_id":null,"evidence_quote":"Establishes political homophily in retweet networks, motivating the robustness check that separates direct zealot connections from the full voter-model mechanism."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior empirical validation of the voter model at the aggregate level (US election vote shares), which this paper extends to individual users."}],"review_version":1}