{"id":"bda5e4d4-31f0-4272-9982-0c2bdea9aede","arxiv_id":"2505.14280","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Political issues on German Twitter/X are strongly aligned along a single left-right axis, and hyperactive retweeters (multipliers) exhibit the most consistent ideological sorting across issues.","lead":"A study of German Twitter/X trends finds that political debates repeatedly split into the same two ideological camps, so that separate issues line up along one left-right division. The analysis identifies a less visible user type, the multiplier, who retweets and curates others' content, as the most consistent force behind this alignment.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Causal claim that multipliers 'drive' alignment is not identified by the observational design; the paper should replace 'drive' with 'is associated with' or provide a counterfactual test.","rationale":"After reading the paper in good faith, I find the descriptive core—polarization and issue alignment in the German Twittersphere—to be carefully executed and reasonably robust. The use of degree-corrected SBMs with force-layout validation, manual checks, and the ANMI/ARI sensitivity analysis in §A.6 are genuine strengths, and the shared code and data availability are positive. My concern is not with the measurement but with the central causal claim. The abstract and title assert that influencers and multipliers 'drive' polarization and issue alignment. The evidence in §4.2–§4.3 is purely associational: multipliers (top out-degree users) have higher alignment scores than other groups. No exogenous variation in multiplier activity, no temporal precedence, and no control for confounding are provided. The authors' own discussion states they 'may only speculate about the precise logic by which they act' and calls for further investigation. This is a self-acknowledged gap. I therefore agree with the reader's CONDITIONAL verdict, but I locate the load-bearing problem in the causal identification rather than in the retweet-as-endorsement assumption. The endorsement assumption is a real measurement threat, but even if it holds, the paper has not shown that multipliers cause the alignment. A counterfactual test—removing multiplier retweets and re-estimating alignment—would provide direct evidence and is feasible with the released code. If the authors cannot run such a test, they should soften the causal language to 'associated with' or 'hypothesized to drive.'","tokens_in":21209,"tokens_out":9369,"duration_ms":89015,"concrete_test":"Re-run the full alignment pipeline (§3.2–§3.5) after deleting all retweets made by the top-1000 multiplier accounts (by total out-degree, as defined in §A.4) from each of the 1726 retweet networks, then recompute the issue alignment matrix. If the mean off-diagonal τ remains within 10% of the original, the claim that multipliers 'drive' alignment is falsified; if it drops substantially, the causal role is supported. This directly tests whether multiplier activity is necessary for the observed alignment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central to the paper's contribution is the causal statement that alignment is 'driven by' influencers and multipliers (Abstract; §5). The evidence for this is associational: the top-1000 out-degree users ('multipliers') show higher pairwise and issue alignment than influencers or random users (§4.2, §4.3, Figs. 3–5). This does not establish that multipliers cause alignment. The design has no exogenous variation in multiplier activity, no temporal precedence test, and no control for selection and reverse causality: users who are consistently on one side of many issues may retweet more precisely because they are aligned, making the 'multiplier' label an outcome rather than a cause. Confounders such as ideological extremity or algorithmic amplification are not addressed. The authors acknowledge in §5 that they 'may only speculate about the precise logic by which they act' and that 'more thorough analyses need to be conducted' to investigate coordinated behavior. Without a counterfactual or a temporal test, the central causal claim is a hypothesis, not a finding. This concern is more load-bearing than the retweet-as-endorsement assumption: even if every retweet is an endorsement, the mechanism that 'drives' the global alignment remains unidentified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the German Twittersphere between March 2021 and July 2023 using trending-topic data. A topic model assigns tweets to political issues, and retweet networks per trending topic are clustered into one or two opinion camps. The authors define user alignment across trends, extract global left- and right-leaning camps by clustering the alignment matrix of highly active users, and then compute topic-level issue alignment. They compare three user groups: top in-degree users ('influencers'), top out-degree users ('multipliers'), and a random user sample. The main empirical findings are that political discussions divide users into two broad camps, that issues align strongly across topics, and that multipliers show higher cross-topic alignment than influencers. The paper interprets this as evidence that influencers and multipliers 'drive' polarization and issue alignment.","tokens_in":21453,"tokens_out":6076,"duration_ms":64663,"significance":"If the descriptive findings hold, this is a valuable large-scale study of issue alignment in a European Twittersphere, adding to the scarce multi-topic evidence beyond US survey-based studies. The paper has notable strengths: a long observation window, a publicly shared code and data repository, explicit bot-authenticity checks, and an independent validation using ANMI and ARI in Appendix A.6. However, the central causal claim is not established by the observational design, and several methodological choices need clarification before the claimed mechanism can be accepted. The paper is likely to be of interest to the ICWSM community, but the abstract's 'driven by' wording overstates what the evidence supports.","major_comments":[{"comment":"The central causal claim that issue alignment is 'driven by' influencers and multipliers is not supported by the analysis. Sections 4.2 and 4.3 report associational differences: top out-degree users have higher pairwise alignment and higher cross-topic alignment than influencers or random users. There is no exogenous variation in multiplier activity, no temporal precedence test, and no control for reverse causation or confounders such as ideological extremity. The authors themselves state in Section 5 that they 'may only speculate about the precise logic by which they act' and that 'more thorough analyses need to be conducted.' The abstract and discussion should therefore replace 'driven by' with a weaker formulation such as 'is associated with' or should add a counterfactual or temporal test.","section":"Abstract; Section 5"},{"comment":"The silhouette threshold that decides whether a retweet network is polarized is stated with opposite directions in the main text and the appendix. Section 3.2 says: 'If the score is higher than 0.4, we keep the SBM-based cluster assignment, otherwise assume that there is only one cluster.' Appendix A.3 says: 'If the silhouette score S(K) ≤ 0.4, then we keep the clustering. Else, we discard it and assume that the network is best described by only a single cluster.' Since this threshold determines which trends enter the alignment computation, the contradiction materially affects all downstream results. The authors must correct the typo and report robustness of the main findings to the threshold choice.","section":"Section 3.2; Appendix A.3"},{"comment":"The retweet-as-endorsement assumption is load-bearing, and the paper's own migration example shows that retweets are not always endorsements. In Section 4.3, the authors describe left-leaning influencers being 'pulled' into the right-leaning cluster because right-leaning accounts retweet them. If retweets can be critical or purely amplifying, then c_k(i) is not a reliable stance signal, and the derived alignment values could be partly an artifact of this signal. The paper should explicitly discuss this limitation in the main text and ideally validate the stance interpretation on a sample of retweets with text-based stance labeling.","section":"Section 3.2; Section 4.3; Section 5"},{"comment":"The issue alignment measure in Eq. (10) uses the global left/right camp labels that were themselves obtained by clustering the user alignment matrix computed over all trends. This creates a circular dependency that could inflate the measured topic alignment. The independent ANMI/ARI checks in Appendix A.6 go some way toward validating the overall issue-alignment pattern, but they do not validate the specific influencer-vs-multiplier comparison. The authors should either provide an anchor-free validation of the multiplier findings or explicitly qualify the anchor-based results as dependent on the global camp construction.","section":"Section 3.4; Section 3.5; Appendix A.6"},{"comment":"The claim that 'multipliers align issues more strongly than influencers' is made on the basis of visual inspection of the alignment matrices. No distributions, effect sizes, confidence intervals, or statistical tests are reported for this comparison. Given that this difference is one of the main supports for the proposed mechanism, the paper should quantify the alignment difference and test whether it is larger than expected from activity levels or other observable user characteristics.","section":"Section 4.3; Figures 4 and 5"}],"minor_comments":[{"comment":"The phrase 'a part of of coordinated inauthentic behavior' contains a duplicated 'of.'","section":"Section 5"},{"comment":"The sentence 'The F1-scores for the logistic classifier and the stochastic gradient descent based classifier are reported in Tab. 1' should refer to Table A1, since the main-text Table 1 reports topics and polarization shares.","section":"Appendix A.1"},{"comment":"The phrase 'The topic Drug Legalisation , mainly related to cannabis' has an extra space before the comma.","section":"Section 4.1"},{"comment":"The caption should state whether the optimal leaf ordering is computed separately for the influencer and multiplier matrices; the two matrices appear to use different orderings, which makes direct comparison of individual entries less straightforward.","section":"Figure 5 caption"},{"comment":"The definition of J_i as the subset of influencers and multipliers who occur jointly with user i in at least one trend should clarify whether user i itself can be included in J_i even when i is not an influencer or multiplier.","section":"Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's abstract is likely to be quoted as establishing that multipliers cause issue alignment. Given the observational design and the authors' own caveats, I would condition acceptance on removing the causal wording or substantially strengthening the identification strategy. The descriptive results, with the independent ANMI/ARI validation, are promising, but the current framing overstates the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the honest bottom line: the descriptive core is solid, and the influencer/multiplier split is the most interesting thing in the paper. The authors make a real contribution by showing that German political Twitter over 2021–2023 sorts into two recurring camps, and that issue alignment is strong across topics—against what survey-based work would predict. The distinction between high in-degree (influencers) and high out-degree (multipliers) users, and the finding that multipliers show stronger cross-topic alignment than influencers, is a genuine addition to the polarization literature. They also ship code and anonymized data, and they validate their user-centric alignment measure with ANMI and ARI in the appendix. That is the right kind of robustness work, and it gives me confidence the descriptive result is not a method artifact.\n\nWhere the paper falls short is in the gap between the evidence and the causal framing. The abstract says alignment is 'driven by' influencers and multipliers. The data are associational: multipliers are more active, appear in more trends, and their alignment values are higher. Nothing in the design identifies a mechanism—there is no exogenous variation, no temporal test, and no placebo. The authors are more careful in the discussion, where they write that they 'may only speculate about the precise logic.' The abstract should match that caution. 'Associated with' is the defensible claim; 'driven by' is a hypothesis.\n\nThe circularity worry the reader raised is real but mild. The global camps are built from the same alignment matrix used later for per-topic anchoring. But the ANMI/ARI checks do not use those camps, and they reproduce the pattern, so the main result survives. The retweet-as-endorsement assumption is load-bearing, but it is standard in this literature and the authors flag it.\n\nMinor issues: several thresholds (silhouette 0.4, top-1000 cutoffs) are manual, and there is no uncertainty quantification. That is a limitation, not a fatal flaw.\n\nMy recommendation: send it to referees. This is a solid descriptive paper that deserves publication with revision. Ask the authors to soften the causal language and to make the threshold choices more transparent. I would engage with it.","headline":"Solid descriptive finding on German Twitter polarization; the influencer/multiplier distinction is worth taking seriously, but the causal framing overreaches.","tokens_in":21970,"tokens_out":2965,"would_cite":true,"duration_ms":29217,"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 German-speaking Twitter/X public sphere is split into two stable ideological camps, and a small set of hyperactive users—influencers who create content and multipliers who retweet and curate it—produces the strong alignment of…","keywords":["polarization","issue alignment","retweet networks","influencers","multipliers","topic modeling","German Twitter","stochastic block model"],"falsifier":"Take a random sample of retweets from users in each inferred camp and have annotators judge whether the retweeting user's own accompanying text or profile signals agreement with the retweeted account's stance; if a large share of retweets are hostile, mocking, or neutral amplifications, the endorsement assumption—and with it the two-camp and issue-alignment results—would fail.","tokens_in":20971,"feed_emoji":"🔁","tokens_out":6784,"duration_ms":58443,"temperature":0.7,"pith_summary":"The paper claims that the German-speaking Twitter/X public sphere is divided into two stable ideological camps, and that separate political issues—Covid, migration, climate, even greetings—are strongly aligned onto this single left-right divide. It argues that this alignment is not a natural property of the issues themselves but is driven by two small groups of hyperactive users: influencers, who produce most of the content that gets retweeted, and multipliers, who retweet heavily and bundle content from influencers into ideologically consistent packages. The multipliers are the more distinctive mechanism, acting as intermediaries between content creators and the wider audience and aligning issues more consistently than the influencers they amplify. If the claim is right, platform-level polarization is largely an outcome of curation and amplification by a few active accounts, not a simple reflection of the opinions of the general user base.","feed_headline":"Influencers and multipliers drive Twitter's single left-right divide","feed_subtitle":"A study of 1,726 German trending-topic networks finds separate issues sort users into the same two camps.","key_machinery":"The central object is the retweet network treated as an endorsement graph: a directed link is drawn from i to j when i retweets j, and densely connected groups are interpreted as opinion camps. For each of 1,726 trending-topic networks, the paper infers whether a one-block or two-block structure is most likely using a degree-corrected stochastic block model, validated with force-directed layouts and a silhouette threshold, and encodes each user's cluster membership as +1, -1, or 0. From these partitions it builds the user alignment α(i,j), the average product of two users' cluster memberships over trends in which both participate, and the issue alignment τ(T1,T2), the average product of topic-wise camp membership scores. Influencers are defined as the most-retweeted users and multipliers as the most-retweeting users; the alignment matrices for these two groups are then clustered and labeled as left- and right-leaning by manual inspection of the accounts.","core_discovery":"On the paper's own terms, the central discovery is that the German Twittersphere exhibits strong issue alignment: the opinion clusters inferred from retweet networks sort users into the same left- and right-leaning camps across a wide range of political topics, with the strongest core around Covid, journalism/media, and German politics. Comparing users with the highest in-degree (influencers) and highest out-degree (multipliers), the authors find that multipliers participate in more trends and align issues more consistently, and that deviations such as migration and Ukraine arise when users are retweeted into the opposing camp or break with their camp on a single issue. The paper reads this as evidence that the polarized public sphere is produced by two cores of strongly active users: influencers who generate ideologically charged content and multipliers who curate and amplify it.","pith_inferences":["This reading implies that the multiplier mechanism should be testable on other platforms: if hyperactive curators exist on Facebook or TikTok, the same alignment pattern should be measurable in their share or reshare networks.","A testable extension would separate passive from active sorting into a camp: the migration example suggests users can be pulled into the opposing cluster by being retweeted there, which would not necessarily mean their own opinions are aligned.","One consequence of the algorithm-amplification argument is that multipliers may shape content exposure beyond their own follower counts; an audit comparing feed exposure with and without multiplier retweets would quantify this channel.","If trending topics are disproportionately triggering, the issue-alignment result may overstate alignment in everyday political talk; sampling non-trending political tweets would show whether the alignment persists outside high-engagement content."],"forward_implications":["Separate political issues on German Twitter do not each produce their own division; they reproduce one global left-right split, so studying a single issue such as Covid already exposes the same camp structure as the rest of the political agenda.","Multipliers, not just influencers, carry the alignment effect: they are active across more trends and their topic-wise alignments are more consistent, making retweet amplification a plausible target for platform interventions.","The alignment measure based on user alignment avoids the small-overlap problem of partition-similarity scores, because it can compare users even when the overlapping node set between two networks is small.","The contrast with survey-based findings of low issue alignment is explained by the different 'surveyed' population: Twitter users are younger, more politicized, and trending topics are likely trigger points that evoke stronger engagement than neutral survey questions."],"supporting_citations":[{"why":"Supplies the core assumption that retweets are endorsements, on which the opinion-camp extraction rests.","marker":"(boyd, Golder, and Lotan 2010)"},{"why":"Established the bipolar structure of political retweet networks that this paper extends to issue alignment.","marker":"(Conover et al. 2011)"},{"why":"Prior demonstration of ideological camps in the German Twittersphere and the retweet-network method reused here.","marker":"(Gaisbauer et al. 2021)"},{"why":"Prior evidence of issue alignment in the Finnish Twittersphere that this paper's alignment result parallels.","marker":"(Chen et al. 2021)"},{"why":"Recent measurement of elite and mass polarization in social networks that motivates the influencer/multiplier distinction.","marker":"(Salloum, Chen, and Kivelä 2024)"},{"why":"The survey-based benchmark finding of weak issue alignment that the paper's strong alignment result contrasts with.","marker":"(Baldassarri and Gelman 2008)"},{"why":"Provides the concept of intermediaries and curation that defines the multiplier role.","marker":"(Friemel and Neuberger 2023)"},{"why":"Supports the claim that retweets drive algorithmic amplification, explaining why multipliers matter beyond their followers.","marker":"(Bouchaud, Chavalarias, and Panahi 2023)"}],"fun_headline_variants":["Influencers and multipliers drive Twitter's single divide","Two camps, aligned issues: influencers and multipliers on Twitter/X","Multipliers align issues across Twitter's trending topics","How influencers and multipliers shape Twitter polarization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire two-camp structure rests on the assumption that retweets are endorsements, so clusters of retweeters can be read as opinion camps; if many retweets are critical, mocking, or algorithmic amplifications, the inferred camps and alignments could be artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Influencers and multipliers drive Twitter's single divide","Two camps, aligned issues: influencers and multipliers on Twitter/X","Multipliers align issues across Twitter's trending topics","How influencers and multipliers shape Twitter polarization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000462,"raw_usage":{"total_tokens":2283,"prompt_tokens":891,"completion_tokens":1392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":1330}},"tokens_in":507,"tokens_out":1392,"duration_ms":10546,"temperature":1.0,"reasoning_tokens":1330,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:36:52.538175+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of retweets from users in each inferred camp and have annotators judge whether the retweeting user's own accompanying text or profile signals agreement with the retweeted account's stance; if a large share of retweets are hostile, mocking, or neutral amplifications, the endorsement assumption—and with it the two-camp and issue-alignment results—would fail.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Established the bipolar structure of political retweet networks that this paper extends to issue alignment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior demonstration of ideological camps in the German Twittersphere and the retweet-network method reused here."},{"cited_title":"a -Anttila , T.; and Kivel \\","cited_arxiv_id":null,"evidence_quote":"Prior evidence of issue alignment in the Finnish Twittersphere that this paper's alignment result parallels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The survey-based benchmark finding of weak issue alignment that the paper's strong alignment result contrasts with."},{"cited_title":"N.; and Neuberger, C","cited_arxiv_id":null,"evidence_quote":"Provides the concept of intermediaries and curation that defines the multiplier role."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the claim that retweets drive algorithmic amplification, explaining why multipliers matter beyond their followers."}],"review_version":1}