{"id":"b7d05010-75ff-4c38-8ed6-7237c7df2aff","arxiv_id":"2412.17342","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Across five disasters, Twitter users' activity, network structure, and influence distributions stay stable as the networks grow, while communication frequency and speed decay with distance.","lead":"This paper analyzes 13 million tweets from five US disasters to understand how crowds process risk information online. It finds that communication patterns stay stable as crisis networks grow, but information mostly spreads over short distances and a tiny number of users dominate the flow.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Memory-invariance claim rests on a within-crisis new-edge proxy that cannot detect pre-crisis relational memory; the 85% new-edge statistic is compatible with strong pre-existing ties.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: equating memory with the daily proportion of new edges and inferring that prior communication is irrelevant. I agree with that assessment, and I see it as the most central issue because the abstract explicitly promotes the memory invariance as a mechanism. The power-law fitting concerns and the spatial-decay statistical testing concerns raised by the reader are real but secondary: they affect the precision and support of descriptive claims, whereas the memory argument is an unsupported causal claim that is used to justify universality. My proposed check would directly test the memory claim with pre-crisis interaction data, which is the missing evidence. Because the reader already issued a conditional verdict that appropriately calls for tightened claims and additional statistical work, my stress-test does not shift the verdict. I would keep the verdict at CONDITIONAL, which corresponds to 'UNCHANGED' relative to the reader's output.","tokens_in":13597,"tokens_out":3689,"duration_ms":41553,"concrete_test":"For each of the five crisis cases, obtain the same users' interaction history for a 30-day pre-crisis window (via Twitter API or an archived dataset), then test whether a directed edge (retweet/reply/quote) between users i and j during the crisis is significantly more likely if i and j interacted in the pre-crisis window, controlling for user activity volume, geographic distance, and crisis-day network size. Use a matched-pair or logistic-regression design comparing pre-crisis interacting pairs to otherwise similar non-interacting pairs. If pre-crisis interaction is a significant positive predictor, the Section 3 memory conclusion collapses and the abstract's 'not affected by memory' clause must be removed or substantially qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's claim that temporally invariant patterns are 'not affected by people's memory and ties' strength' is load-bearing for the paper's mechanistic and universal framing. The only evidence given for the memory part is in Section 3 (Fig. 5, 'Proportions of new edges among all edges in each day'), where the authors find that new edges account for more than 85% of edges created each day and conclude that communication memory does not significantly affect activity or structure. This inference is a non-sequitur. The measure counts edges that first appear in the crisis-period cumulative network; it cannot see interactions that occurred before the crisis, even though the authors explicitly define memory as 'previous communications prior to the threat event.' A pair that communicated regularly before the crisis and continues during the crisis will contribute a 'new' edge on the first crisis day, and the statistic will remain high if many new users enter each day. The high new-edge proportion is therefore compatible with strong pre-existing relational memory. The observed stability of activity proportions, communication-structure ratios, and influence distributions is a descriptive finding that may hold, but the specific causal claim that memory is irrelevant is not supported by the presented analysis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper analyzes roughly 13 million tweets from five crisis events (Atlanta airport power outage, North American blizzard, Kincade wildfire, Hurricane Harvey, Hurricane Florence) to characterize collective information processing in social networks. Communication events are projected onto directed, weighted, cumulatively growing networks; the authors measure activity proportions, communication structures, PageRank-based influence, edge weights, response times, and spatial communication patterns. The central claims are that user activities and communication structures are temporally invariant as networks grow, that social influence follows a stable power-law distribution with exponent α≈1.9825, that these patterns are unaffected by people's memory and tie strength, and that communications are spatially localized with decay over distance. The paper concludes with intervention strategies for emergency communication and misinformation mitigation.","tokens_in":13737,"tokens_out":3856,"duration_ms":40289,"significance":"If the claims were fully supported, the paper would offer useful empirical regularities for modeling risk communication in disasters and for designing network interventions. The dataset is large and spans five distinct hazard types, which is a real strength for any cross-event generalization. The authors also provide a clear operationalization of several network metrics and report goodness-of-fit checks for the power-law claim. However, the manuscript's headline conclusions—especially the claims of invariance under memory and tie strength, and the universal normal/power-law descriptions—go beyond what the presented measurements actually establish. The core value of the paper is therefore as a descriptive, multi-case observational study; the causal and mechanistic statements need substantial revision and additional analysis to be justified.","major_comments":[{"comment":"The claim that activity proportions follow a normal distribution after Box-Cox transformation applies only after excluding the two extreme bins that 'account for more than 90% of the activities' (Section 3, first paragraph of Results). The abstract and conclusion state more generally that 'the activities and structures are stable' and that the normal distribution is a general pattern. As written, the analysis describes the minority of users with mixed activity types, not the full user population. The paper should either qualify the normality and invariance claims to explicitly exclude the dominant single-activity majority, or provide a complementary characterization of the full distribution that does not discard the majority of users.","section":"Section 3, Fig. 2 and Section 2.2"},{"comment":"The abstract's claim that the identified patterns are 'not affected by people's memory' rests entirely on the observation that new edges constitute over 85% of edges created each day. This is a non-sequitur. The 'new edges' statistic counts edges absent from the crisis-period cumulative network; it cannot see interactions that occurred before the crisis, which is exactly the 'previous communications prior to the threat event' that the paper defines as memory. A pair of users who communicated frequently before the crisis and continue to interact during it will appear as a 'new' edge on the first crisis day. The high proportion of new edges is therefore compatible with strong pre-existing relational memory. The presented analysis cannot support the conclusion that memory does not affect activity or structure.","section":"Section 3, memory paragraph (Fig. 5)"},{"comment":"The power-law exponent α≈1.9825 is a fitted value, not a prediction from a generative model, and its reported stability is obtained using a lower bound chosen as 'the value of the second minimum bin' rather than a principled x_min selection such as the Clauset-Shalizi-Newman procedure. With this arbitrary cutoff, the claim that the influence distribution 'follows a power law' is a summary of a histogram after normalization, not a demonstrated universal law. The paper should either (a) derive the exponent from a stated mechanism (e.g., preferential attachment) and test the prediction, or (b) report the fit with a statistically justified x_min, the uncertainty in α, and a comparison against alternative heavy-tailed distributions. Without this, the 'stable power-law' claim is not as strong as the abstract presents it.","section":"Section 2.3 and Fig. 6"},{"comment":"The conclusion that tie strength does not affect the observed patterns is inferred from the fact that median edge weights are 1 and variances are less than 3. This describes the marginal distribution of edge weights; it does not test whether repeated interactions between the same users (or the presence of a few high-weight edges) alter activity dynamics, structure ratios, or PageRank values. Because PageRank is computed on unweighted edges, the analysis is structurally unable to detect an effect of tie strength on influence. The abstract's statement that the patterns are 'not affected by people's memory and ties' strength' is therefore not established. At minimum, the authors should compare weighted and unweighted versions of the network metrics or perform a regression-style test relating tie strength to activity and structure outcomes.","section":"Section 3, edge-weight analysis (Fig. 7)"}],"minor_comments":[{"comment":"The word 'activties' in the definition of activity proportion is a typo; also the equation's denominator 'all activties of a user' should be 'all activities of a user'.","section":"Equation (1)"},{"comment":"The metric labeled 'medium response time' is aggregated by the median (the text says 'we use the medium response time'), so the label should be 'median response time' or the aggregation should be corrected.","section":"Throughout, e.g., Figs. 9–10"},{"comment":"The text states 'As shown in Fig. 7, we can find that while influential communicators only account for a very small percentage...' but Fig. 7 displays the distribution of edge weights, not the communication flows between influential and ordinary users. The correct reference appears to be Fig. 8. Please fix the cross-reference.","section":"Section 3, fourth paragraph"},{"comment":"The Box-Cox transformation and the normal-distribution claim rely on a kernel bandwidth fixed at h=0.3 and a Gaussian kernel; the sensitivity of the normality conclusion to these choices is not reported. A brief robustness check (e.g., different bandwidths) would strengthen the claim.","section":"Section 2.2"},{"comment":"The city and state distance thresholds (2 km and 100 km) are presented as 'the scale of a city' and 'the scale of a state' without a citation or justification; the conclusions about spatially bounded communication depend on these thresholds and should be justified or varied in a sensitivity analysis.","section":"Section 2.4 and Figs. 9–10"},{"comment":"The phrase 'In addition, spatially localized communication spikes and global transmission gaps in the networks' is grammatically incomplete and should be rephrased.","section":"Abstract and Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical corpus is substantial and the descriptive results (stable structure ratios, heavy-tailed influence, distance-decay of communication) are likely of interest to the risk-communication community. However, the current presentation overclaims causal invariance with respect to memory and tie strength, and the headline 'stable power-law' is built on an arbitrary lower bound. I recommend major revision with the specific requirement that the authors either retract the causal claims from the abstract or provide analyses (e.g., pre-crisis interaction data, weighted vs. unweighted comparisons, proper x_min estimation) that actually test them. I also suggest the editor ask the authors to make the supplementary figures referenced in the text available and check that the Q-Q plots and other goodness-of-fit evidence are reported for all five cases, not only Hurricane Harvey."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you want a quick sense of where crisis informatics is empirically. The genuinely useful part is the cross-case descriptive baseline: over five different threat types (airport outage, blizzard, wildfire, two hurricanes), cumulative Twitter networks show stable activity ratios, stable communication-structure ratios, and a stable heavy-tailed influence distribution once the networks reach a few days of growth. That is a real empirical pattern worth knowing, and the authors have the data (13M tweets, 5M users) and the patience to show it day by day. The spatial decay finding—same-city communication spikes and cross-region gaps—is also a reasonable descriptive contribution, even if it is presented without error bars or significance tests.\n\nThe soft spots are mostly at the level of interpretation. The stress-test is right about the memory claim: the 85% new-edges statistic cannot see pre-crisis interactions, so it cannot support the conclusion that \"communication memory ... does not significantly affect\" the patterns. That inference is a non-sequitur, and it is load-bearing for the abstract's universal framing. The normal-distribution claim for activity proportions is also qualified by the exclusion of the two extreme bins that contain over 90% of users; that is honest about what was done, but the abstract's \"stable\" wording overstates the coverage. The power-law fit (alpha ~ 1.98) is a fitted summary rather than a prediction; no alternative distributions are tested, and the lower bound is set at the second minimum bin, so the exponent is likely to be fragile. The tie-strength claim rests on a narrow edge-weight distribution, not on any direct measure of relational strength. These are fixable in a revision, but they currently make the mechanistic and universal claims in the title and abstract unsupported.\n\nWhat the paper does well is set up a transparent pipeline: cumulative network construction, PageRank, Box-Cox transform, KDE, and MLE power-law fitting, with a limitations section that acknowledges platform, country, and feed-algorithm scope. The authors clearly know the surrounding literature. The overclaims are the problem, not the analysis style.\n\nVerdict: this deserves a serious referee, but the referee should push hard on the memory/tie-strength section and demand robustness checks (alternative distributions, bootstrapped spatial decay, sensitivity to x_min and beta). It is not a desk reject; it is a revise-and-resubmit with the causal framing dialed back to descriptive.","headline":"Useful descriptive baseline for crisis Twitter dynamics, but the causal claims about memory and tie strength outrun the analysis.","tokens_in":14351,"tokens_out":2012,"would_cite":false,"duration_ms":21131,"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":"Crisis social networks display stable collective information-processing laws.","keywords":["collective information processing","risk communication","crisis social networks","power-law distribution","PageRank influence","temporal invariance","spatial communication decay","Twitter disaster data"],"falsifier":"Track each ordered pair of users across the days of the same five disasters and compare the activity mix, response times, and PageRank-influence distribution of pairs who interacted before with pairs who had no prior contact; if prior-contact pairs behave measurably differently, the paper's no-memory conclusion is contradicted.","tokens_in":13291,"feed_emoji":"📈","tokens_out":8088,"duration_ms":74931,"temperature":0.7,"pith_summary":"The paper sets out to show that collective information processing on social media during crises obeys stable statistical regularities even as the online network grows and the disaster evolves. Using about 13 million tweets from five different threats, it claims that user activity proportions, the mix of communication structures, and PageRank-based user influence all remain temporally invariant as the network grows each day. The influence scores converge to the same power-law distribution, with exponent $\\alpha \\approx 1.98$, in all five cases, and the paper argues these patterns do not depend on users' communication memory or tie strength because new edges dominate daily activity and edge weights are almost all 1. Spatially, communication frequency and medium response time decay with physical distance, so information exchange is concentrated near the threat area. If true, the stability would make crisis-time network forecasting and targeted intervention substantially easier.","feed_headline":"One power law rules crisis Twitter influence","feed_subtitle":"Across 13 million tweets from five disasters, activity, structure, and influence hold steady while networks grow.","key_machinery":"The central object is the cumulative daily human network built from Twitter retweet, reply, and quote interactions, with users as nodes and weighted directed activities as edges. Each day's network contains all nodes and edges from previous days, so temporal invariance is measured against a strictly growing graph. PageRank on this graph supplies the influence score, normalized by the second minimum bin and fitted with a maximum-likelihood power-law model $p(x) \\sim C x^{-\\alpha}$; the exponent $\\alpha \\approx 1.9825$ is the quantitative signature of stable influence. The spatial machinery is the top-100 location set with great-circle distances via the haversine formula, used to show that communication frequency and medium response time decay past the scale of a state. The causal work is inferential: because the daily share of new edges stays above 85% and edge weights cluster at 1, the invariance is attributed to the absence of memory and to weak ties.","core_discovery":"The paper's central discovery is that the risk-encoding networks formed during crises are temporally invariant in their activity, structure, and influence, with user influence following a power-law distribution across all five disasters. Activity proportions for original posts, retweets, replies, and quotes become approximately normal after a Box-Cox transformation, and this normal shape holds as new users and edges join day by day. The shares of converging, reciprocal, and self-loop communication structures also stay stable, with more than 80% of users appearing only in converging structures as information sources. Normalized PageRank influence scores from the cumulative networks collapse onto a single power-law curve with exponent $\\alpha \\approx 1.9825$ by the final day of each disaster. The paper interprets these patterns as independent of communication memory and tie strength: newly created edges account for more than 85% of daily edges, and most edge weights equal 1. Spatially, communication is bounded by distance, with localized spikes in the same or nearby locations and weak transmission to distant places.","pith_inferences":["A natural extension the paper does not pursue is testing whether the same temporal and spatial regularities appear outside crises, such as in political or entertainment news; if they do, the mechanism would reflect a general property of collective attention rather than of threat encoding alone.","The memory conclusion could be stress-tested by simulating agent-based networks with explicit memory parameters and checking whether a memory-free model reproduces the observed $\\alpha \\approx 1.98$; this would give the paper's proxy-based argument a mechanistic footing.","The observed spatial boundary is likely co-produced by platform recommendation algorithms as well as by user preference; if so, the documented gaps could shrink when algorithms favor out-of-region content, which is a testable intervention."],"forward_implications":["Emergency response agencies could identify the top roughly 2% of influential communicators early in a disaster, since the influence ranking stabilizes after the first day and persists as the network grows.","Misinformation mitigation can concentrate on these stable hubs, because most propagated information originates from them while ordinary users' original information accounts for only about 1% of communication flow.","Crisis communication models can treat activity and structure distributions as fixed, making day-by-day forecasts and simulations much simpler.","Platforms could deliberately add cross-region edges, for example recommending a popular risk message from one cluster to an influential user in another, to close the spatial transmission gaps the paper documents."],"supporting_citations":[{"why":"Supplies the maximum-likelihood power-law fitting and Kolmogorov-Smirnov goodness-of-fit test used to estimate the influence exponent.","marker":"(Clauset et al., 2009)"},{"why":"Supplies the transformation that maps activity-proportion distributions to approximately normal form.","marker":"(Box & Cox, 1964)"},{"why":"Provides the PageRank mathematics used to assign user influence scores in the networks.","marker":"(Ipsen & Wills, 2006)"},{"why":"Supplies the scale-free network framework linking a power-law influence distribution to the dominance of hubs.","marker":"(Barabási, 2009)"},{"why":"Supplies the scale pattern of collective human behavior used to set city- and state-level distance thresholds in the spatial analysis.","marker":"(González et al., 2008)"},{"why":"Supplies the method of generating synthetic distributions to compute p-values for the power-law goodness-of-fit test.","marker":"(Noulas et al., 2012)"},{"why":"Provides the social-influence experiment and theory motivating the measurement of influence in collective information processing.","marker":"(Becker et al., 2017)"},{"why":"Provides the diffusion-in-networks framework with heterogeneous activity that the paper contrasts with static-network studies.","marker":"(Akbarpour & Jackson, 2018)"}],"fun_headline_variants":["Crisis Twitter influence obeys one power law across disasters","Power-law influence holds steady in crisis Twitter networks","Five disasters, one power law for Twitter influence","Stable power-law influence despite network growth in crises","Crisis Twitter influence collapses onto a single power law"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that 'communication memory' can be measured by the daily share of newly created edges; if memory instead operates through repeated exchanges between the same users or continued sharing of the same information, the test would not detect it and the conclusion that memory does not affect the observed patterns would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Crisis Twitter influence obeys one power law across disasters","Power-law influence holds steady in crisis Twitter networks","Five disasters, one power law for Twitter influence","Stable power-law influence despite network growth in crises","Crisis Twitter influence collapses onto a single power law"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1279,"prompt_tokens":930,"completion_tokens":349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":274}},"tokens_in":546,"tokens_out":349,"duration_ms":3708,"temperature":1.0,"reasoning_tokens":274,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:34:11.780464+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track each ordered pair of users across the days of the same five disasters and compare the activity mix, response times, and PageRank-influence distribution of pairs who interacted before with pairs who had no prior contact; if prior-contact pairs behave measurably differently, the paper's no-memory conclusion is contradicted.","supporting_citations":[],"review_version":1}