{"id":"2b3c173b-dfd6-4745-9e65-a822d677e5f7","arxiv_id":"2608.10968","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"In a multiplex opinion-dynamics model with repulsive influence, neutral exposure maximizes polarization, controversy-seeking algorithms nearly match it, and algorithmic homophily shields users, because the cross-bloc attention fraction controls radicalization rate.","lead":"A mathematical model of social media shows that when people repel strongly opposed opinions, an uncurated feed is the most radicalizing design, while a feed that filters out disagreement protects against extremism. The paper turns this into a rate theory: one number, the fraction of attention a platform directs across opposing camps, sets how fast opinions race to the extremes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The rate-limited claim is not resolved by the onset data: for the radicalizing platforms Table I predicts λc≈0.008–0.016, below the smallest simulated λ=0.05, so the simulations demonstrate saturation, not the absence of a threshold.","rationale":"Read in good faith, the paper is a careful modeling study with an internally consistent deterministic reduction: Eq. (9) follows from a symmetric two-bloc state with stationary slot fractions, and the robustness controls (homogeneous strengths, κ=4, reflecting boundaries, k and ρ sweeps) support the qualitative inversion. The reason I do not simply accept the paper is that its most distinctive quantitative output, the parameter-free λc(T), is not actually resolved by the reported onset simulations. The smallest nonzero attention share is 0.05, while the predicted crossovers are 0.008–0.016; the data can only show 'already saturated at 0.05,' which is equally compatible with a small but nonzero threshold. The y0 input is also taken from the same simulation family rather than derived, so the 'no parameters fitted' claim is weaker than stated. The reader's weakest assumption (fast rewiring and symmetric blocs) is related but less decisive, because the ρ=1/20 controls and the homogeneity control show the inversion survives partial slot memory and unequal strengths, whereas the low-λ regime is simply unmeasured. A focused low-λ, long-horizon scan with an independent y0 measurement would settle the matter; until then a conditional verdict is appropriate.","tokens_in":16386,"tokens_out":11789,"duration_ms":120272,"concrete_test":"Run the Level-4 agent-based model for the neutral and controversy platforms at λ = 0.005, 0.01, 0.02, 0.03, 0.04 with reference parameters, extending the horizon to T = 400 (and 800 for λ=0.005), with at least 12 seeds; record ⟨|x|⟩, Var(x), and the early-time cluster count. The rate-limited prediction is that ⟨|x|⟩ grows on the time scale trad(λ) of Eq. (11) for every λ>0, eventually reaching the boundary, while a threshold-limited model shows no separation growth below a critical λ no matter how long T is. Separately, measure y0 in the assimilative Level-1/Level-2 runs at the same D and σx as the Level-4 runs and recompute λc; if the implied onset disagrees with the measured onset at λ<0.05, the parameter-free crossover needs revision.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central theoretical result is Eq. (9) and the finite-horizon crossover Eq. (12), which the abstract and Sec. III B present as the parameter-free engine of the inversion. Figure 9 samples the Var(x)–λ onset only at λ ≥ 0.05, while Table I gives λc ≈ 0.016 (neutral) and 0.008 (controversy) at T=80. The comparison is thus made entirely in the saturated regime: any model with a nonzero onset threshold below 0.05 would reproduce the figure equally well. The paper rules out the uniform-state thresholds λ* ≈ 0.34/0.55, but the headline claim that 'any λ>0 with nonzero cross-bloc exposure produces outward drift' is not tested. In addition, the numerical λc depends on y0=0.6, read off from the assimilative runs rather than derived; if the pre-radicalization fragmented state at low λ contains more than two blocs or unequal blocs, the effective p(y) of Eq. (8) differs and Eq. (12) shifts. Because the rate-limited versus threshold-limited distinction is the paper's central message, this unresolved low-λ regime is the most load-bearing weakness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a continuous-time multiplex model in which agents diffuse in physical space and interact through an adaptively rewired digital attention graph under a conserved attention budget. Influence is assimilative, indifferent, or repulsive depending on opinion distance, and three platform designs are compared: similarity-driven, neutral, and controversy-seeking engagement kernels. The central result is an inversion: when repulsion is active, the neutral platform is the most radicalizing, the controversy-driven platform is close behind, and strong algorithmic homophily is the least radicalizing. A two-bloc reduction yields a rate equation dy/dt = 2 alpha_tot lambda eta p(y) y with p(y) = E(2y)/(E(0)+E(2y)), leading to a finite-horizon crossover lambda_c(T) in the digital attention share. The well-mixed limit reproduces the neutral Var(x)-versus-lambda curve nearly quantitatively. Additional results include a mobility-attention phase diagram, a null result for geographic opinion structure under opinion-independent mobility, and a Peclet-number threshold for the onset of spatial opinion domains under homophilic mobility.","tokens_in":16752,"tokens_out":5823,"duration_ms":57740,"significance":"If the central claim holds, the paper makes a conceptual contribution by framing platform-induced radicalization as rate-limited rather than threshold-limited and by mapping any curation kernel to a state-dependent cross-bloc exposure profile whose quadrature sets the radicalization time. The paper is unusually transparent: simulation code and data are openly available, the analytic reduction is clearly stated, and the platform ordering is tested across the (eta, epsilon2) grid, system sizes up to N=1600, and five control variants. The robustness controls (homogeneous strengths, finite-variance strengths, reflecting boundaries, attention-window size, and rewiring rate) strengthen confidence in the ordering itself. However, the headline claim that 'any lambda > 0 with nonzero cross-bloc exposure produces outward drift' is not directly resolved by the presented simulations, because the predicted crossovers for the neutral and controversy platforms lie below the smallest sampled attention share. The empirical support for rate-limited, rather than threshold-limited, radicalization is therefore incomplete.","major_comments":[{"comment":"The central rate-limited claim is tested only in the saturated regime. Table I predicts lambda_c = 0.016 (neutral) and 0.008 (controversy) at T=80, while Fig. 9 samples lambda only at 0.05 and above. The simulations therefore rule out thresholds at or above 0.05 (including the uniform-state values 0.34 and 0.55), but they are fully consistent with any nonzero threshold below 0.05. Since the distinction between 'rate-limited' and 'threshold-limited' is the paper's headline, the comparison in Fig. 9 is not sufficient. I recommend adding simulations at lambda = 0.005, 0.01, 0.02, 0.03, and 0.04 for the neutral and controversy platforms, and/or measuring the early-time drift d<|x|>/dt to test the predicted linear scaling in lambda. Without such a test, the statement in the abstract that radicalization is rate-limited, not threshold-limited, goes beyond what the data establish.","section":"Sec. III B / IV E, Eq. (12), Fig. 9"},{"comment":"The claim that the crossover prediction involves 'no parameters fitted' is overstated because y0 = 0.6 is read off from simulation output (Figs. 1 and 3) rather than derived from the model parameters. The crossover scales as ln(1/y0); changing y0 to 0.5 raises lambda_c by about 35%, and y0 below 0.45 would postpone the onset until the repulsion threshold 2y >= epsilon2 is reached. The prediction is parameter-free with respect to the onset data, but not parameter-free in the strict sense stated in the abstract. I ask the authors to either derive y0 from the bounded-confidence fragmentation condition or report how y0 varies across initial conditions, noise levels, and box sizes, and to state explicitly that the tabulated lambda_c values inherit this measured input.","section":"Sec. III B, Table I"},{"comment":"The two-bloc reduction assumes symmetric, equal-strength blocs and fast rewiring relative to bloc drift. The authors note that the fast-rewiring condition is only marginally met for the neutral kernel at the reference parameters, and for heavy-tailed influence strengths the realized cross-bloc exposure fluctuates around Eq. (8) with convergence that is slow in N. While Table II shows the platform ordering is robust at lambda = 0.6, the quantitative values of lambda_c in Table I inherit these assumptions. A sensitivity analysis of Eq. (12) with respect to asymmetric bloc sizes and finite slot memory would clarify the regime of validity of the reported crossover values and would strengthen the rate-limited conclusion.","section":"Sec. III A, Eq. (8)"}],"minor_comments":[{"comment":"The appearance of s_j^2 in the well-mixed equation is explained in the text, but a one-line derivation showing that slot sampling proportional to E s_j combined with s_j-weighted averaging produces the squared strength would help the reader verify this step.","section":"Eq. (13)"},{"comment":"The dotted vertical lines marking lambda_c are described as part of the left panel, but the centre and right panel captions also mention lambda_c(T); please clarify which panels contain the crossover markers.","section":"Fig. 9 caption"},{"comment":"The phrase 'the finest nonzero attention share sampled' would be more informative if the exact lambda grid were stated in the text or caption, since the low-lambda resolution is central to the threshold question.","section":"Sec. IV E"},{"comment":"The word 'polarization' is used both for Var(x) and for the two-cluster categorical state. The definitions in Sec. II E are clear, but the double use may confuse readers; a footnote or a different name for one of the two quantities would improve readability.","section":"Sec. II E"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Name],\n\nThe short version: this paper is worth your time. It delivers a clean analytical mechanism for a counterintuitive inversion—under repulsive influence, neutral uncurated platforms radicalize fastest and homophilic curation shields—and it backs that mechanism with a transparent two-bloc reduction and extensive robustness checks. The code and data are out there. The authors are also honest about the contested repulsive branch and the heuristic status of the reduction.\n\nThe new thing is Eq. (9): any engagement kernel maps to a cross-bloc exposure profile p(y), and the radicalization rate is set by p(y) times the digital attention share λ. That framing turns the platform-design debate into a kinetic question, and the finite-horizon crossover λc(T) from Eq. (12) orders the designs cleanly. The well-mixed limit reproduces the neutral curve nearly quantitatively with no fitted parameters, which is a real plus.\n\nNow the soft spot, and it is the load-bearing one. The paper's headline is 'radicalization is rate-limited, not threshold-limited,' and the abstract states any λ>0 produces outward drift. But the simulated onset scan samples λ at 0 and then 0.05, while Table I predicts λc ≈ 0.016 (neutral) and 0.008 (controversy). So all the simulation shows is that saturation is reached at the smallest sampled λ; it cannot distinguish a genuinely threshold-free mechanism from one with any threshold below 0.05. The authors rule out the uniform-state thresholds of 0.34 and 0.55, but that is not the same as ruling out small thresholds. To secure the central claim they should either simulate at λ = 0.01–0.03 or prove the reduction holds for arbitrary λ without extra assumptions.\n\nSecond, the 'no parameters fitted' framing leans on y0 = 0.6 read off from the bounded-confidence cluster positions. y0 enters λc directly. It is not a fitted parameter in the usual sense, but calling the crossover parameter-free is too generous; a sensitivity analysis on y0 or a derivation from the fragmentation dynamics would settle it. This is a minor issue, but the abstract overstates.\n\nThird, the two-bloc reduction assumes symmetric blocs and fast rewiring, and the authors note the fast-rewiring condition is only marginally met for the neutral kernel. That weakens the quantitative force of Eq. (9) though the ordering seems robust.\n\nOverall: a serious modeling paper, clearly thought through, with appropriate caveats about the contested empirical basis for repulsion. It deserves refereeing. The referee should push on the low-λ regime and the y0 dependence before publication.\n\nBest,\n[Name]","headline":"Clean kinetic mechanism for the exposure-polarization inversion, but the headline no-threshold claim is under-tested because the sampled λ grid starts at 0.05, above the predicted crossover.","tokens_in":17261,"tokens_out":3024,"would_cite":true,"duration_ms":28191,"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":"The paper argues that under repulsive social influence, an uncurated platform is the most radicalizing design because the stationary cross-bloc attention fraction, not any threshold, sets the radicalization rate.","keywords":["opinion dynamics","radicalization","polarization","bounded confidence","algorithmic curation","attention budget","multiplex networks","agent-based model"],"falsifier":"Measure the influence function $F(x_i, x_j)$ directly in a controlled opinion-change experiment: if the response to strongly opposed views has no negative (repulsive) branch and merely flattens to indifference, then the outward drift in Eq. (9) is zero and the predicted inversion cannot occur; if the negative branch exists, the measured $\\epsilon_2$ and $\\eta$ make the finite-horizon crossover $\\lambda_c(T)$ quantitatively testable.","tokens_in":16141,"feed_emoji":"📱","tokens_out":11014,"duration_ms":99324,"temperature":0.7,"pith_summary":"What would happen if social-media platforms radicalize people by exposing them to disagreement rather than by hiding it? This paper builds a stochastic model in which agents move through space and receive algorithmically curated content under a fixed attention budget, and it claims that once influence includes a repulsive response to strongly opposed views, the ranking of platform designs inverts: an opinion-blind uncurated feed is the most radicalizing, a controversy-seeking feed nearly as much, and strong algorithmic homophily is the least radicalizing because it starves the repulsive channel. The central relation is a derived rate law: the engagement kernel sets the stationary fraction $p(y)$ of attention slots that point to the opposite opinion bloc, and that fraction, not any critical attention share, controls the outward drift of the blocs. The model yields a finite-horizon crossover $\\lambda_c(T)$ in the digital attention share that orders the three designs with no parameters fitted to the onset data, and it reports that uncurated exposure sets the ceiling that curated platforms approach as they stop withholding repulsive content. If the claim is right, it supplies a mechanism for field observations in which cross-cutting exposure increased polarization and implies that exposure-diversity interventions can have either sign depending on how prevalent negative influence is.","feed_headline":"Uncurated feeds radicalize most, hiding disagreement protects","feed_subtitle":"A stochastic model shows that cross-bloc attention, not a threshold, sets the radicalization rate--which homophily starves.","key_machinery":"The load-bearing object is the cross-bloc attention fraction $p(y)=E(2y)/(E(0)+E(2y))$: under fast rewiring, the stationary probability that one of an agent's $k$ attention slots points into the opposite bloc when the blocs sit at $\\pm y$. It maps any curation kernel $E$ to a state-dependent repulsive-exposure profile and enters the two-bloc rate equation $dy/dt=2\\alpha_{\\mathrm{tot}}\\lambda\\eta\\,p(y)\\,y$, valid once the bloc separation exceeds the repulsion threshold. Because that equation is separable, the radicalization time is an exact quadrature over $1/(y\\,p(y))$, and the constant-exposure approximation gives the closed-form crossover $\\lambda_c(T)$. This is the object that turns platform design into a number: the three kernels' profiles are pointwise ordered, and their cross-bloc fractions at the observed starting separation span two orders of magnitude.","core_discovery":"The paper's central claim is an inversion of the standard platform-ranking story. In a model with Brownian mobility, bounded-confidence physical interaction, and an algorithmically rewired digital layer under conserved attention, purely assimilative influence makes an opinion-blind platform heal fragmentation and homophilic curation build echo chambers. Once the influence function gains a repulsive branch for opinion distances beyond $\\epsilon_2$, the order reverses: the neutral platform drives opinions to the boundary and pins them there, the controversy-seeking platform arrests just short of the boundary, and the similarity-driven platform only slowly creeps outward. The author reduces this to a two-bloc statement: when two blocs sit at $\\pm y$, the stationary cross-bloc attention fraction is $p(y)=E(2y)/(E(0)+E(2y))$, and the bloc separation obeys $dy/dt = 2\\alpha_{\\mathrm{tot}}\\lambda\\eta\\,p(y)\\,y$. Radicalization is therefore rate-limited, not threshold-limited: any $\\lambda>0$ with nonzero cross-bloc exposure produces outward drift, and the finite-horizon crossover $\\lambda_c(T)=\\ln(y_f/y_0)/(2\\alpha_{\\mathrm{tot}}\\eta\\,p_0\\,T)$, with no parameters fitted to the onset data, places the neutral and controversy onsets near $\\lambda\\approx0.01$–$0.02$ and the similarity onset near $\\lambda\\approx1$, matching the simulations.","pith_inferences":["If the rate-limited picture transfers to real platforms, an exposure-diversity intervention should change polarization continuously with dose and with the population's repulsion prevalence; the mixed field record could then reflect different populations sitting at different points on the $\\lambda_c(T)$ curve rather than contradictory evidence.","The $p(y)$ mapping suggests a practical diagnostic: from logged feed exposure and opinion estimates, one could measure the realized cross-bloc fraction and forecast whether an exposure policy will help or backfire without fitting a complete opinion-dynamics model.","The same two-bloc machinery should extend to any curation objective whose exposure profile is pointwise ordered over the traversed opinion range—such as chronological, novelty-maximizing, or advertiser-optimized feeds—so the three kernels studied here are boundary examples of a broader design space."],"forward_implications":["With repulsive influence active, the neutral platform pins mean extremism at about 1, the controversy platform arrests near 0.95, and the similarity platform reaches only about 0.79 on the simulation horizon—so the platform ordering is quantitative, not just qualitative.","Radicalization has no positive critical attention share within the deterministic reduction; apparent thresholds in finite experiments are horizon effects captured by $\\lambda_c(T)$.","Uncurated exposure sets the saturation ceiling: the similarity platform approaches it as $\\gamma\\to1$, and the controversy platform reaches it once its engagement peak $\\delta\\geq\\epsilon_2$, at which point it radicalizes twice as fast as the neutral platform.","Once the platform owns most of an agent's attention, physical mobility becomes irrelevant to the outcome, so geography determines the social state only at low digital attention share.","Under opinion-independent mobility there is no geographic opinion structure, and reinstating it requires homophilic drift above a Péclet threshold $\\chi\\ell/D\\sim1$."],"supporting_citations":[{"why":"supplies the algorithmic-bias amplification mechanism that motivates the similarity-driven curation baseline.","marker":"[1]"},{"why":"provides the link-recommendation polarization dynamics that the adaptive digital layer extends.","marker":"[2]"},{"why":"is the field experiment showing cross-cutting exposure increased polarization, the empirical target of the inversion.","marker":"[4]"},{"why":"supplies the assimilation–indifference–repulsion influence function that drives the radicalization channel.","marker":"[7]"},{"why":"gives the longitudinal empirical support for negative influence that justifies activating the repulsive branch.","marker":"[9]"},{"why":"presents the post-distribution curation model whose attractive/repulsive outcomes the paper contrasts with its kinetic account.","marker":"[13]"},{"why":"derives stationary polarization phases for collaborative filtering, the stationarity-oriented comparison point.","marker":"[17]"}],"fun_headline_variants":["Uncurated feeds radicalize most; homophily protects","Radicalization is rate-limited, not threshold-limited","Cross-bloc attention sets radicalization speed, not threshold","Neutral platforms pin opinions; similarity curation shields","Finite-horizon crossover predicts radicalization onset"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire rate theory assumes that attention slots renew so quickly that the platform's exposure mix reaches its stationary value before the two opinion blocs drift noticeably, and that the two blocs are symmetric in strength—an assumption the paper itself flags as only marginally met for the neutral kernel at its reference parameters.","fun_headline_variants_meta":{"raw":{"variants":["Uncurated feeds radicalize most; homophily protects","Radicalization is rate-limited, not threshold-limited","Cross-bloc attention sets radicalization speed, not threshold","Neutral platforms pin opinions; similarity curation shields","Finite-horizon crossover predicts radicalization onset"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000429,"raw_usage":{"total_tokens":2306,"prompt_tokens":1171,"completion_tokens":1135,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":787,"completion_tokens_details":{"reasoning_tokens":1057}},"tokens_in":787,"tokens_out":1135,"duration_ms":11844,"temperature":1.0,"reasoning_tokens":1057,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:09:13.029239+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the influence function $F(x_i, x_j)$ directly in a controlled opinion-change experiment: if the response to strongly opposed views has no negative (repulsive) branch and merely flattens to indifference, then the outward drift in Eq. (9) is zero and the predicted inversion cannot occur; if the negative branch exists, the measured $\\epsilon_2$ and $\\eta$ make the finite-horizon crossover $\\lambda_c(T)$ quantitatively testable.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the link-recommendation polarization dynamics that the adaptive digital layer extends."},{"cited_title":"S ˆ ırbu, D","cited_arxiv_id":null,"evidence_quote":"supplies the algorithmic-bias amplification mechanism that motivates the similarity-driven curation baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the field experiment showing cross-cutting exposure increased polarization, the empirical target of the inversion."},{"cited_title":"Jager and F","cited_arxiv_id":null,"evidence_quote":"supplies the assimilation–indifference–repulsion influence function that drives the radicalization channel."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives the longitudinal empirical support for negative influence that justifies activating the repulsive branch."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"presents the post-distribution curation model whose attractive/repulsive outcomes the paper contrasts with its kinetic account."},{"cited_title":"Bellina, C","cited_arxiv_id":null,"evidence_quote":"derives stationary polarization phases for collaborative filtering, the stationarity-oriented comparison point."}],"review_version":1}