{"id":"b4dcbc37-eed7-498d-ae83-ef2f11dfc835","arxiv_id":"2411.12949","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EIN augments graph rumor detectors with an eUSD-style state encoder trained against LLM-generated stance labels, achieving 0.6 to 1.6 percentage point accuracy gains over RAGCL and better depth-robustness on DRWeibo, Weibo, and Pheme.","lead":"An epidemiology-inspired encoder supervised by LLM-generated stance labels is bolted onto graph-based rumor detectors, producing small accuracy gains on three public datasets and reportedly smoother performance across propagation-tree depths. The paper is a representative case of first-principles-inspired ML where the actual lift likely comes from pseudo-labeled auxiliary supervision rather than the epidemiological model itself.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LLM pseudo-label supervision in L_p is the load-bearing link: its accuracy is unmeasured and the EIN w/o ERL ablation does not isolate it from encoder capacity, so the epidemiology mechanism is confounded.","rationale":"Reader's weakest assumption already identified the LLM-label issue, and I agree that it is the primary stress point. The abstract's claim is not just that EIN wins, but that epidemiology-informed state modeling is the reason; Eq. (7) combines the data-driven embedding with an encoder whose only task-specific supervision comes from L_p. Because the encoder's input is just tree depth and node count (Eqs. 2-5), any text-derived knowledge from the LLM can enter only through the target distributions in L_p. If those targets are biased, the depth-robustness result could be an artifact of a depth/size prior learned from LLM labels rather than of the eUSD dynamics. The absence of a human-label accuracy check and the absence of a same-capacity ablation with random label supervision leave this confound open. Additional concerns about missing error bars in Figure 4 and per-dataset hyperparameter tuning are real but secondary. The core empirical regularity is plausible and the code is released, so the conditional verdict stands without change.","tokens_in":16850,"tokens_out":10166,"duration_ms":130304,"concrete_test":"On all three datasets, run a controlled 5-seed comparison: (1) full EIN; (2) EIN with the encoder kept but L_p removed; (3) EIN with L_p computed from randomly shuffled LLM stance labels; (4) EIN with L_p computed from human-annotated stance labels on a sample (e.g., 200 threads per dataset). Report overall ACC/AUC/F1 and depth-stratified accuracy for each variant. If (2) or (3) matches full EIN, the specific LLM label content is not the driver; if (4) differs materially from full EIN, pseudo-label bias propagates into the learned encoder. Also report the agreement rate between Gemma 2-9B and human labels on the same sample.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the epidemiology-informed encoder, guided by LLM-generated stance labels, improves accuracy and depth robustness—rests on the auxiliary loss L_p in Eq. (8). L_p assumes Gemma 2-9B's pairwise stance labels are good enough to define target state distributions p_{u,t}, p_{s,t}, p_{d,t}. The paper reports no agreement between these pseudo-labels and human annotations, and no experiment that holds the encoder architecture and all other training signals fixed while varying only the LLM supervision. The reported ablation 'EIN w/o ERL' (Section 4.4) removes the epidemiology-informed representation learning component, so it conflates two effects: (i) the extra capacity and inductive bias of the encoder and (ii) the information carried by the LLM pseudo-labels. If Gemma's stance judgments correlate with the rumor label (e.g., rumor threads are denial-heavy), then L_p is effectively distilling a rumor prior into a small depth/size-conditioned network, and the observed gains would not demonstrate the epidemiological state mechanism. This does not make the method internally inconsistent, but it makes the central claim under-identified: the contribution of the claimed first-principles component is not separable from the LLM supervision.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Epidemiology-informed Network (EIN), a plug-in module for graph-based rumor detectors that augments the data-driven tree embedding with a state embedding produced by an 'environmental Unknown-Support-Denial' (eUSD) encoder. To train this encoder without human stance annotations, the authors use Gemma 2-9B to generate pairwise stance labels between a post and its parent, derive node-level state labels via an XOR rule, and add a KL-divergence loss L_p between the predicted and pseudo-label state distributions. Experiments on DRWeibo, Weibo, and Pheme report accuracy, AUC, and F1 improvements over ten baselines, including variants built on BiGCN and ResGCN backbones, plus a depth-stratified analysis intended to show robustness across shallow and deep propagation trees. The code is released.","tokens_in":16993,"tokens_out":6878,"duration_ms":68582,"significance":"If the central claim holds, the work offers a simple, backbone-agnostic way to improve robustness in rumor detection without inference-time LLM calls, and it proposes an interesting use of LLM pseudo-labels for stance supervision. The manuscript ships code, covers three datasets, and evaluates with multiple baselines, which are real strengths. However, the claimed mechanism is not empirically isolated from the LLM supervision, the gains on Pheme overlap within standard deviations, and the robustness analysis lacks error bars and significance tests. The 'epidemiology-informed' framing is also stronger than what the equations justify, since all key parameters are learned and no test verifies the learned dynamics against the eUSD model. These issues leave the central attribution of the gains under-evidenced, but they are addressable in revision.","major_comments":[{"comment":"The definition of the target distribution is time-inconsistent. The text defines p_{u,t} = |C_u|/(n-1) with n the number of nodes in the tree, which does not depend on t. If the intent is to use depth-specific counts at stage t, the construction is not provided; if the target is indeed the stationary global distribution, the time-indexed notation and the comparison with time-indexed encoder outputs \\hat{p}_{u,t} are misleading. Either way, the objective as written does not clearly enforce the claimed temporal dynamics, and this is load-bearing for the representation learning module.","section":"Section 3.3, Eq. (8)"},{"comment":"The ablation 'EIN w/o ERL' removes the epidemiology-informed representation learning component, but the description does not clarify whether the epidemiology-informed encoder itself remains in the network. As reported, the ablation conflates two effects: the added encoder capacity and the information carried by the LLM-generated pseudo-labels. The paper does not provide a control that keeps the encoder and removes only the auxiliary loss, nor a control with random or constant stance labels, nor an evaluation of the agreement between Gemma's stance labels and human annotations. Without these, the improvement attributed to 'epidemiology-informed representation learning' could be due to the pseudo-labels encoding a rumor prior or to plain regularization.","section":"Section 4.4, Figure 5"},{"comment":"Several headline gains are within one standard deviation of the baseline. On Pheme in Table 2, EIN's ACC 82.74±0.77 vs BiGCN 82.00±1.28, AUC 81.78±1.30 vs 79.76±3.38, and F1 76.45±1.66 vs 73.84±4.64 all overlap; in Table 3, EIN+ResGCN improves ResGCN on Pheme by only 0.46 ACC, 0.19 AUC, and 0.24 F1. The abstract's claim that EIN 'outperforms state-of-the-art methods' and the Section 4.2 claim that it 'consistently enhances' both backbones are not supported without paired significance tests or confidence intervals.","section":"Tables 2 and 3"},{"comment":"The robustness claim is based on bars without error bars or statistical tests, and the depth-stratified evaluation is severely imbalanced: Table 4 shows only 4.65% of DRWeibo test samples have depth>5 and only 10.08% of Weibo test samples have depth=1. The visible differences in those sparse bins are not interpretable as evidence of robustness. The paper should report the number of test samples per bin, add significance tests or confidence intervals, and restrict the claim to bins with sufficient support.","section":"Section 4.3, Figure 4"},{"comment":"The eUSD model is described as first-principles domain knowledge, but the implemented encoder is a learned linear residual network: Eq. (3) is U_{t+1} = (1-α-β) U_t with learnable α and β, and Eq. (4) comprises learned linear maps W_s, W_d, and W_u. No experiment checks whether the learned states actually follow eUSD dynamics, and no comparison is made against a same-capacity MLP with no epidemiology structure beyond the EIN w/o ERL variant. The 'epidemiology-informed' attribution is therefore not demonstrated; the eUSD form is a hypothesis, not a constraint.","section":"Sections 2.2 and 3.1"}],"minor_comments":[{"comment":"The depth categories are labeled '5+' in the figures and 'depth≥5' in Table 4; the text also uses 'depth≥5' and 'depth ≥ 5' on the same page. Please standardize the notation.","section":"Figures 1 and 4 and Table 4"},{"comment":"The model name 'LeRuD' is spelled 'LeRUD' in the second paragraph of Section 4.2; use one consistent name throughout.","section":"Section 4.2"},{"comment":"The definition b_u = W_{u0} - αW_{u0} - βW_{u0} = (1-α-β)W_{u0} can flip sign when α + β > 1, which is inconsistent with the 'number of nodes in a tree' interpretation of the initial Unknown state. Please clarify or constrain the initialization.","section":"Eq. (2)"},{"comment":"The XOR rule for propagating state labels assumes a two-valued stance and logical transitivity across the tree. This assumption should be stated explicitly and ideally checked on a small human-annotated sample, especially since the LLM prompt itself is only asked for binary labels.","section":"Algorithm 1"},{"comment":"The ablation results are presented without standard deviations, so it is impossible to tell whether the differences between EIN, EIN-reg, and EIN w/o ERL are meaningful; please add error bars or a numerical table.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical contribution is promising, but the 'first-principles' framing is substantially stronger than what the model actually implements; the mechanism may be better described as LLM-supervised learned state dynamics. The self-citation [15] is not abusive, though the paragraph in Section 5.2 connects it to the present work only loosely. The manuscript is within scope for WWW, and the release of code is a positive signal. My main concern for the editor is whether the authors can obtain human-validated stance labels or a meaningful pseudo-label accuracy measure; without that, the central attribution claim may remain under-identified even after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid applied-ML paper with a real but modest empirical result, and the central 'epidemiology-informed' claim is softer than the abstract suggests. The new bit is combining an eUSD-style state encoder with LLM-generated stance pseudo-labels as an auxiliary training signal, plugged into existing graph-based rumor detectors. That combination is not in the cited prior work, and the paper backs it with experiments on three datasets, released code, and a depth-stratified analysis. Credit where due: the writing is clear, the baselines are reasonable, and the plug-and-play framing means the empirical claim is easy to reuse.\n\nThe soft spots are in proportion to how soft they actually are. First, the encoder in Eqs. 3-4 reduces to a learned linear residual update—alpha, beta and all W's are trainable, with no conservation or structural constraint from the ODE—so the 'first-principles' content is mostly an initialization and architectural template. That is not disqualifying, but the eUSD-vs-USD ablation in Section 4.4 compares two learned variants, and it does not validate the epidemiological mechanism. Second, the LLM pseudo-label loss L_p is never checked against human annotations, and the 'EIN w/o ERL' ablation removes the encoder and the loss together, so it cannot separate added capacity from the LLM supervision. The stress-test note is right that if Gemma's stance labels correlate with the rumor label, L_p is effectively distilling a rumor prior into a depth-conditioned network. Third, the headline gains are 0.6-1.6 pp over RAGCL, often within one standard deviation on individual metrics; the depth-robustness figure has no error bars, and alpha/beta and lambda are tuned per dataset. These are addressable with a validation-based protocol, significance tests, and an ablation that keeps the encoder but drops L_p.\n\nOverall: the performance claim is not circular—it is measured on held-out test data—and the paper is honest about its hyperparameter sensitivity. The work is a competent applied contribution, not a breakthrough. I would send it to serious peer review and ask for those three fixes.","headline":"Solid, reproducible applied-ML paper with modest gains; the epidemiology framing and LLM-supervision contribution are both less clean than claimed, but the work deserves proper refereeing.","tokens_in":17676,"tokens_out":2302,"would_cite":true,"duration_ms":25580,"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":"This paper claims that an epidemiology-inspired state encoder, supervised by LLM-generated stance pseudo-labels, makes graph-based rumor detectors more accurate and more robust across shallow and deep propagation trees.","keywords":["rumor detection","graph neural networks","epidemiology-informed model","propagation tree","stance labels","large language models","robustness","KL divergence"],"falsifier":"Train EIN on the same three datasets with random or flipped LLM stance labels while keeping the encoder and KL loss; if accuracy does not drop, the stance-supervision mechanism is not carrying the gain. Separately, human-annotate a random sample of reply posts and compare Gemma 2-9B's labels to the human labels; near-chance agreement on ambiguous replies would undercut the claimed mechanism.","tokens_in":16518,"feed_emoji":"🦠","tokens_out":8117,"duration_ms":68156,"temperature":0.7,"pith_summary":"The paper argues that graph-based rumor detectors fail on both ends of the propagation-tree spectrum: shallow trees give too few interactions to learn from, while deep trees drown the signal in noisy, mixed-stance replies. It proposes EIN, a plug-in that models each rumor tree's spread with an environmental Susceptible–Infectious-style dynamics (Unknown → Support/Denial) and uses a large language model to label each reply's stance during training, creating an auxiliary supervision signal. The central claim is that this epidemiology-informed representation, added to any graph backbone, improves accuracy on the DRWeibo, Weibo, and Pheme datasets and flattens the performance drop across tree depths, all without any LLM calls at inference time.","feed_headline":"Epidemic math improves rumor detection at all depths","feed_subtitle":"An epidemic-state encoder plus LLM stance labels beats graph baselines on three real-world datasets.","key_machinery":"The eUSD model is a three-compartment environmental transmission system: $dU/dt = -\\alpha U e - \\beta U e$, $dS/dt = \\alpha U e$, $dD/dt = \\beta U e$, with environment influence $e$ set to 1. It is discretized by a forward difference quotient into recurrent updates of three learnable state embeddings (Unknown, Support, Denial) over the tree's depth steps. State labels are produced by an LLM that classifies each reply's stance toward its parent, combined by XOR along the path to derive the reply's state relative to the root. A KL-divergence loss matches the encoder's softmax-normalized state distributions to the LLM-derived distributions at every depth and is added to the rumor-classification cross-entropy loss with weight $\\lambda$; this auxiliary loss is what steers the data-driven GNN toward epidemiology-informed dynamics.","core_discovery":"The central discovery is that the eUSD (environmental Unknown–Support–Denial) transmission model, discretized into learnable state embeddings, supplies a robust prior for rumor-spread dynamics that data-driven GNNs lack. When the encoder's predicted state distributions are trained to match LLM-generated state labels via a KL-divergence auxiliary loss, the resulting tree-level representation consistently improves accuracy, AUC, and F1 for both BiGCN and ResGCN across DRWeibo, Weibo, and Pheme, with the gains most visible where the baselines degrade: depth-1 trees and trees deeper than five levels.","pith_inferences":["The paper never measures how accurate Gemma 2-9B's stance labels are against human annotations; if those pseudo-labels are biased on sarcastic or ambiguous replies, the KL loss would imprint that bias into the learned dynamics.","The 'EIN w/o ERL' ablation removes the encoder and the loss together, so it cannot isolate the contribution of the LLM pseudo-labels; a variant that keeps the encoder but randomizes the KL targets would settle it.","Because the encoder is a small recurrent state machine, the same plug-in pattern could transfer to other tree-structured social tasks, such as early fake-news alerting or stance detection on low-depth threads.","The depth-robustness finding suggests a general recipe: injecting a principle-based prior into a GNN can compensate for both data sparsity and noisy structure, which may extend beyond rumor detection to other graph classification problems."],"forward_implications":["EIN is a plug-in: adding it to BiGCN and ResGCN improves all three metrics on all three datasets in the paper's reported results.","The LLM is used only during training, so the deployed detector has no extra inference-time cost over its graph backbone.","The robustness results cover depth-1 trees (few interactions) and depth>5 trees (noisy cascades), the early-detection and high-controversy cases that matter in practice.","The eUSD formulation outperforms the classic USD model in the paper's ablation, indicating that root-post-driven (environmental) influence fits rumor propagation better than population-driven influence.","The sensitivity analysis shows the method needs per-dataset tuning of the transition-rate initialization and the auxiliary-loss weight $\\lambda$ to reach its best numbers."],"supporting_citations":[{"why":"Supplies BiGCN, the backbone model EIN is plugged into for propagation and dispersion modeling.","marker":"[2]"},{"why":"Supplies ResGCN, the other backbone used for EIN evaluation and the preliminary depth-robustness experiment.","marker":"[36]"},{"why":"Supplies the RAGCL baseline, the DRWeibo dataset, and the Word2Vec node-feature and hyperparameter settings adopted by EIN.","marker":"[7]"},{"why":"Supplies the epidemic model of rumor diffusion that motivates the epidemiological analogy.","marker":"[6]"},{"why":"Supplies the environmental transmission model on which the eUSD formulation is built.","marker":"[4]"},{"why":"Supplies the stochastic environmental-transmission model used to justify eUSD's environment factor $e$.","marker":"[34]"},{"why":"Supplies Gemma 2-9B, the LLM that generates the stance pseudo-labels for training.","marker":"[32]"},{"why":"Supplies FSNet, the stance-supervised rumor verification model that motivates using stance labels and contrasts with LLM-generated labels.","marker":"[19]"},{"why":"Supplies the Weibo dataset used in evaluation.","marker":"[21]"},{"why":"Supplies the Pheme dataset used in evaluation.","marker":"[38]"}],"fun_headline_variants":["Epidemic model outshines GNNs in rumor detection at all depths","LLM-labeled epidemic states make rumor nets robust to tree depth","Epidemic-informed net handles shallow and deep rumor cascades","Epidemic model plus LLM stance labels fixes rumor depth bias","Rumor detection immune to tree depth via epidemic encoder"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The auxiliary loss assumes the LLM's stance labels are accurate enough to supervise the state encoder; the paper never validates Gemma 2-9B's labels against human annotations, and if they are systematically wrong on sarcastic or ambiguous replies, that error gets baked into the learned state dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Epidemic model outshines GNNs in rumor detection at all depths","LLM-labeled epidemic states make rumor nets robust to tree depth","Epidemic-informed net handles shallow and deep rumor cascades","Epidemic model plus LLM stance labels fixes rumor depth bias","Rumor detection immune to tree depth via epidemic encoder"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00088,"raw_usage":{"total_tokens":3790,"prompt_tokens":916,"completion_tokens":2874,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":2785}},"tokens_in":532,"tokens_out":2874,"duration_ms":19752,"temperature":1.0,"reasoning_tokens":2785,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:00:48.556145+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train EIN on the same three datasets with random or flipped LLM stance labels while keeping the encoder and KL loss; if accuracy does not drop, the stance-supervision mechanism is not carrying the gain. Separately, human-annotate a random sample of reply posts and compare Gemma 2-9B's labels to the human labels; near-chance agreement on ambiguous replies would undercut the claimed mechanism.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Weibo dataset used in evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Pheme dataset used in evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies BiGCN, the backbone model EIN is plugged into for propagation and dispersion modeling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RAGCL baseline, the DRWeibo dataset, and the Word2Vec node-feature and hyperparameter settings adopted by EIN."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the epidemic model of rumor diffusion that motivates the epidemiological analogy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the environmental transmission model on which the eUSD formulation is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the stochastic environmental-transmission model used to justify eUSD's environment factor $e$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies FSNet, the stance-supervised rumor verification model that motivates using stance labels and contrasts with LLM-generated labels."}],"review_version":1}