{"id":"2302e368-fea8-42f8-809d-bc24188dc4d0","arxiv_id":"2507.05939","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A continual multimodal misinformation detector that uses Dirichlet process-based expert expansion to curb forgetting and a neural-ODE dynamics model to anticipate evolving fake-news distributions.","lead":"This paper presents DAEDCMD, a method for detecting misinformation in streaming social media posts that aims to forget less and adapt to emerging events. It combines a mixture-of-experts design, which grows new experts only when needed, with a dynamics model that tries to predict how fake content changes over time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (12) appears to train the dynamics model on future event distributions, which would make the reported 'anticipate future' gains a leakage artifact rather than held-out prediction.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing issue: Eq. (12) is written as an integral over all K time periods with ground-truth distributions for every event, which conflicts with the sequential training loop in Algorithm 1 and creates an unresolved leakage possibility for the 'anticipate future' component. I agree that this is the most serious concern because the dynamics model is the central new contribution and the ablation results indicate it contributes the largest performance gain on the hardest dataset. If the implementation is chronological, Eq. (12) should be rewritten with a causal integral; if it is not, the empirical gains are inflated by future information. The other concerns noted by the reader—missing significance tests and underspecified evaluation splits—are secondary: they affect the strength of the comparison but do not directly threaten the validity of the anticipation mechanism. A concrete re-run with a causal L_DM would settle whether the headline claim is genuine. I therefore keep the reader's CONDITIONAL verdict without adjustment.","tokens_in":19176,"tokens_out":3427,"duration_ms":42098,"concrete_test":"Re-run the GAMED row of Table 2 on GossipCop and Twitter with a strictly chronological dynamics objective: while training event k, compute L_DM using only μ_τ and σ_τ for τ ≤ k (or τ < k for a stricter held-out version), keeping all other hyperparameters and the MoE module identical. Compare the resulting Macro F1 and accuracy with Table 2. If the improvements over GAMED+EWC and GAMED+LoRAMoE shrink to within one point or vanish, the headline gains should be attributed to future-distribution leakage; if the gains persist, the concern is resolved. The paper should also explicitly state whether Eq. (12)'s integral is over all K events or only observed events.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that DAEDCMD outperforms six MMD baselines and three continual learning methods. The paper attributes much of this to the environmental dynamics module: the ablation in Table 4 shows 'w/o Feature e' is the largest drop on Twitter. That module is supervised by Eq. (12), which defines L_DM = (1/K) ∫_0^K [||μ̂_τ − μ_τ||² + ||σ̂_τ − σ_τ||²] dτ, where μ_τ and σ_τ are computed as the ground-truth mean and covariance of all data whose temporal label is τ. Under Algorithm 1, training proceeds sequentially over events k = 1..K; while processing event k, the distributions for τ > k should not yet be observable. The paper never states whether the integral is restricted to τ ≤ k or whether future events' μ_τ, σ_τ are nonetheless used in the training loss. If the latter, the dynamics model is trained on the future distributions it is then asked to 'predict,' so the gains attributed to anticipating the future are not from held-out prediction. If the former, Eq. (12) as written is not the actual objective and the experiment cannot be reproduced from the description. Either way, the load-bearing component of DAEDCMD—the continuous-time anticipation of future environments—is not supported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DAEDCMD, a multimodal misinformation detector designed for continual data streams. It combines a mixture-of-experts architecture with a Dirichlet-process-style dynamic expansion of event-specific experts to mitigate catastrophic forgetting, and a continuous-time dynamics model based on neural ODEs to predict the evolution of a Gaussian 'environmental distribution' of fake-article features. The model is evaluated on three continual MMD datasets (GossipCop, Weibo, Twitter), applied on top of six MMD baselines and compared to three continual-learning baselines, reporting consistent accuracy/F1 gains. The paper also includes ablations, a forgetting analysis, a parameter sensitivity study, and a case study.","tokens_in":19412,"tokens_out":8675,"duration_ms":87201,"significance":"Continual misinformation detection is a practical and timely problem, and the paper's combination of parameter isolation with an explicit model of distribution dynamics is interesting. The empirical evaluation is broad: six MMD backbones, three datasets, and three continual-learning baselines, with a public code link and ablations that isolate the contribution of each module. If the future-anticipation mechanism is shown to be causally valid—that is, if the dynamics model is trained without access to future events' statistics—the results would be a solid step toward deployment in streaming settings. At present, that validity is not established because the training objective in Eq. (12) is ambiguous about whether future-event statistics are used.","major_comments":[{"comment":"The dynamics-model objective L_DM in Eq. (12) is defined as (1/K)∫_0^K [‖μ̂_τ−μ_τ‖₂²+‖σ̂_τ−σ_τ‖₂²] dt, where μ_τ and σ_τ are the ground-truth mean and covariance of all data whose temporal label is τ. Under the sequential training loop in Algorithm 1, when event k is being processed, data for events τ>k have not yet been observed, so the integral over 0..K is not computable without using future-event statistics. The paper does not state whether the integral is evaluated over all τ (which would leak future information into the dynamics model) or only over τ≤k (which would make Eq. (12) as written not the actual objective). This ambiguity is load-bearing: the ablation in Table 4 shows that removing the environmental feature causes the largest drop (e.g., -2.53 on Twitter), so a substantial part of the reported gain is attributed to this module. Please specify the exact computation, and if future statistics are used, rerun the experiments with a causally valid objective; if only past/current statistics are used, rewrite Eq. (12) and the surrounding text accordingly.","section":"§3.3, Eq. (12); Algorithm 1"},{"comment":"The Dirichlet process mixture is allegedly implemented by the responsibility formula, but the notation Σρ_{<i,m} is undefined: it is not stated whether this is a sum over previous samples' responsibilities for expert m, how it is accumulated across batches, or how it is updated when a new expert is created. In addition, the 'variational generation loss' L_VG is implemented as the reconstruction loss ‖z_i E_m^G − z_i‖₂², which is not a variational lower bound, and the responsibilities are not normalized. The text calls this a Dirichlet process-based approach, but the connection to a Dirichlet process is not established. Please either provide a formal derivation with the exact update rules, or revise the terminology to describe the heuristic it actually is. This matters because the 'remember past' contribution is supported by this module.","section":"§3.2, Eqs. (4)-(6)"},{"comment":"The caption states that the bold results are 'statistically significant than the baseline models,' but no significance test is described anywhere in the paper or supplementary. Several comparisons have overlapping standard-deviation intervals (e.g., GossipCop F1_fake for CAFE+DaedCmd is 61.05±0.8 vs 59.21±1.5, with intervals that overlap), so the claim of significance as used in the abstract needs a formal test. Please report the test procedure (e.g., paired bootstrap or Wilcoxon signed-rank test over the five seeds) and the resulting p-values or confidence intervals.","section":"§4.2, Table 2"},{"comment":"The text states that 'the Twitter dataset, being the smallest in scale, achieves the largest improvement,' but Table 3 shows Twitter has 13,924 articles, which is larger than GossipCop (12,840) and Weibo (9,528). If 'scale' refers to the number of unique images (Twitter has 514), that should be stated explicitly; as written, the sentence is internally inconsistent and the explanation of the ordering of improvements needs to be corrected.","section":"§4.2, Table 3 and text"}],"minor_comments":[{"comment":"Step 13 says 'Construct the distribution P(ê|y=1) with Eq.(10)', but Eq. (10) defines the ODE derivatives; the distribution parameters are computed in Eq. (9). Please correct the reference.","section":"Algorithm 1, step 13"},{"comment":"The integral in Eq. (12) is written with a 'dt' but the integration variable should be τ (the temporal index). The notation is inconsistent and should be unified.","section":"§3.3, Eq. (12)"},{"comment":"The routing weight r_i is written as z_i W_R, but the dimension of W_R is not specified; please clarify that it outputs a scalar, or describe the output dimension if it is a vector.","section":"§3.2, Eq. (7)"},{"comment":"The method name is rendered inconsistently as 'DaedCmd' and 'DAEDCMD' across the abstract and body; please unify the spelling.","section":"General"},{"comment":"The column 'Avg.↓' is not defined; please state how the average decline is computed and whether it is averaged over the five metrics listed.","section":"Table 4"},{"comment":"The GitHub URL in the introduction contains spaces ('https:// github.com/ wangbing1416/ DAEDCMD'); please format it as a proper link.","section":"Section 1, footnote"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the ambiguity in Eq. (12). Given that the GitHub code is publicly linked, I suggest asking the authors to point to the exact lines where L_DM is computed and to confirm whether the statistics μ_τ, σ_τ for all τ are precomputed before training. If future statistics are used in training the dynamics model, the 'anticipate future' results would need to be re-derived under a strictly causal protocol; if not, the paper can be repaired by rewriting the objective. The paper is otherwise within the scope of the venue, and the empirical breadth is a strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's main claim—that anticipating future environmental distributions helps continual MMD—rests on a training objective that is either miswritten or leaky. Eq. (12) defines L_DM as an integral over all time periods τ from 0 to K, with μ_τ, σ_τ computed from every event's data. Algorithm 1 trains event-by-event; at step k, events k+1..K are not yet available. The paper never says whether the integral is truncated to seen events or whether future-event distributions are used anyway. If the latter, the dynamics model is trained on the future it is then asked to predict, and the ablation gains for the feature ê (the largest single drop on Twitter) are a leakage artifact. If the former, the equation isn't the actual objective and the experiment can't be reproduced. That ambiguity is load-bearing, and it has to be fixed.\n\nThe rest of the paper has real value. Combining Dirichlet-process expert expansion with a neural ODE over Gaussian parameters of fake-news features is a new combination for this task. The empirical setup is broad: six MMD backbones, three continual-learning baselines, three datasets, five seeds, code and data released, plus ablations, a case study, and a time-complexity comparison. Improvements are consistent, and the largest gains on Twitter align with the paper's story about event heterogeneity.\n\nThe other soft spots are smaller but worth naming. The Table 2 caption claims 'statistically significant' results, but no significance test is reported. The test-set construction is underspecified—are these held-out subsets of each event, or a fixed future event? The Dirichlet process is a heuristic: the responsibility formula in Eqs. (4)-(5) has an undefined -log Σρ term that should be spelled out. And the 'environmental distribution' is internal, computed from the model's own features; that's acceptable as a latent variable, but predicting it isn't predicting an external ground truth.\n\nMy bottom line: the method is plausible, the empirical pattern is worth seeing, and the leakage ambiguity is resolvable with a clear sequential-restricted objective and a proper future-only evaluation. Send it to peer review; a competent referee can push on those points. I wouldn't want to see it desk-rejected.","headline":"Load-bearing ambiguity in the dynamics training objective (Eq. 12) makes the 'anticipate future' gains unverifiable as written, but the empirical breadth and method deserve a serious referee.","tokens_in":19987,"tokens_out":4440,"would_cite":true,"duration_ms":48346,"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 claims that a continual multimodal misinformation detector combining a Dirichlet-process-expanded mixture of experts with a neural-ODE model of evolving fake-article distributions consistently and significantly outperforms six…","keywords":["continual learning","multimodal misinformation detection","catastrophic forgetting","mixture of experts","Dirichlet process","neural ordinary differential equations","environmental distribution dynamics","social media misinformation"],"falsifier":"Train DaedCMD under a strict streaming protocol in which the integral in Eq. (12) is evaluated only over already-seen events and then test on the next future event; if the accuracy gains over LoRAMoE shrink to within noise, the reported advantage depends on access to future events' distributions during training.","tokens_in":18878,"feed_emoji":"🕵️","tokens_out":5918,"duration_ms":62821,"temperature":0.7,"pith_summary":"The paper tackles a practical gap: misinformation detectors are trained offline but deployed on a stream of new events, where performance on old events drops and future events have shifted distributions. It proposes DaedCMD, a continual detector with two modules: a Dirichlet-process-expanded mixture of experts that isolates event-specific parameters to reduce forgetting, and a continuous-time dynamics model that predicts the fake-article feature distribution for future events. The paper's central claim is that this combination consistently outperforms six multimodal misinformation baselines and three continual learning methods on GossipCop, Weibo, and Twitter, with the largest gains on the smallest, most imbalanced dataset. If true, it would give practitioners a drop-in way to keep misinformation detectors current on live social media streams.","feed_headline":"Streaming fake-news detector gains up to 5.3 points","feed_subtitle":"New continual-learning method beats six multimodal baselines and three rivals across three social datasets.","key_machinery":"The central object carrying the future-anticipation claim is the environmental distribution $P(\\hat{e}|y=1)=\\mathcal{N}(\\mu,\\sigma^2)$, a Gaussian over the multimodal features of fake articles, whose mean and covariance evolve in continuous time through the neural ordinary differential equations $\\frac{d\\mu}{dt}=G_{\\phi_\\mu}(\\mu,t)$ and $\\frac{d\\sigma}{dt}=G_{\\phi_\\sigma}(\\sigma,t)$, integrated with the dopri5 solver. This lets the model sample an environmental feature $\\hat{e}$ from a predicted future distribution and concatenate it with the forgetting-resistant feature for veracity prediction. The forgetting-prevention machinery is the dynamically adapted mixture of experts, where a Dirichlet-process responsibility in Eq. (4) decides when to spawn a new event-specific expert and an exponential-moving-average update protects the event-shared expert from interference.","core_discovery":"DaedCMD claims to solve continual multimodal misinformation detection by combining two mechanisms. To remember past knowledge, it uses a mixture-of-experts structure with event-shared experts and event-specific experts, where a Dirichlet-process responsibility score decides when a genuinely new event deserves its own expert rather than allocating one per event. To anticipate future knowledge, it models the environmental distribution of fake articles as a Gaussian distribution over multimodal features and learns a continuous-time dynamics model, trained with neural ordinary differential equations, that predicts the mean and covariance of that distribution at future time points. The paper reports that adding DaedCMD to six existing multimodal misinformation detectors and three continual learning baselines consistently improves accuracy and F1 across three datasets, with performance gains ordered GossipCop less than Weibo less than Twitter.","pith_inferences":["Beyond the paper: a strict online evaluation would restrict the dynamics loss in Eq. (12) to already-seen events; the paper does not state which protocol was used, so the future-prediction gains may depend on access to later events' distributions.","Beyond the paper: because the dynamics model tracks only the fake class, extending it to model the real-class distribution could improve calibration and provide a testable variant.","Beyond the paper: replacing the clustering-based split with a pure timestamp split would test whether the gains persist on genuinely time-ordered streams rather than semantically clustered events."],"forward_implications":["Deploying DaedCMD on a live misinformation feed should slow the accuracy decline on older events that continuing training normally causes.","Because DaedCMD wraps the base detector, the same continual-learning modules can be reused to upgrade existing multimodal misinformation detectors rather than retraining them from scratch.","Because experts are created only when a new event is sufficiently different, the parameter count grows with the number of genuinely novel events, not with every batch.","On small and imbalanced datasets, the environmental-dynamics feature supplies information that sparse training data cannot, so the expected practical gain is largest where detection is hardest."],"supporting_citations":[{"why":"Supplies the neural ordinary differential equation framework and dopri5 solver used by the environmental dynamics model.","marker":"[2]"},{"why":"Supplies the Dirichlet process mixture method used to decide when to add a new event-specific expert.","marker":"[23]"},{"why":"LoRAMoE, a multi-expert LoRA continual learning baseline and the architecture DaedCmd builds on.","marker":"[9]"},{"why":"EWC, the parameter-importance continual learning baseline and a formulation of catastrophic forgetting.","marker":"[21]"},{"why":"Replay, the rehearsal-based continual learning baseline that stores past data for joint training.","marker":"[22]"},{"why":"CAFE, an MMD baseline and the source of the contrastive alignment loss used for modality fusion.","marker":"[3]"},{"why":"BMR, a mixture-of-experts MMD baseline that DaedCmd is applied to and compared with.","marker":"[42]"},{"why":"GAMED, the state-of-the-art MMD baseline used for the main ablations and case studies.","marker":"[30]"},{"why":"EANN, the source of the single-pass clustering procedure that splits the datasets into continual subsets.","marker":"[37]"}],"fun_headline_variants":["Dirichlet-process experts and neural ODEs improve continual fake-news detection","New AI remembers past and anticipates future to spot fake news better","Continual multimodal fake-news detector gains up to 5.3 points","Mixture-of-experts with Dirichlet process and ODEs for evolving fake news","Fake-news detector with memory and foresight outperforms in streaming settings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The future-prediction objective trains the dynamics model against the true distribution of every event, including events that appear later in the stream, so the 'prediction' is not a held-out prediction unless the loss is restricted to past events.","fun_headline_variants_meta":{"raw":{"variants":["Dirichlet-process experts and neural ODEs improve continual fake-news detection","New AI remembers past and anticipates future to spot fake news better","Continual multimodal fake-news detector gains up to 5.3 points","Mixture-of-experts with Dirichlet process and ODEs for evolving fake news","Fake-news detector with memory and foresight outperforms in streaming settings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000486,"raw_usage":{"total_tokens":2382,"prompt_tokens":916,"completion_tokens":1466,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":1366}},"tokens_in":532,"tokens_out":1466,"duration_ms":15560,"temperature":1.0,"reasoning_tokens":1366,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:14:57.206075+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DaedCMD under a strict streaming protocol in which the integral in Eq. (12) is evaluated only over already-seen events and then test on the next future event; if the accuracy gains over LoRAMoE shrink to within noise, the reported advantage depends on access to future events' distributions during training.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the neural ordinary differential equation framework and dopri5 solver used by the environmental dynamics model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Dirichlet process mixture method used to decide when to add a new event-specific expert."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LoRAMoE, a multi-expert LoRA continual learning baseline and the architecture DaedCmd builds on."},{"cited_title":"Dynamically Addressing Unseen Rumor via Continual Learning","cited_arxiv_id":"2104.08775","evidence_quote":"Replay, the rehearsal-based continual learning baseline that stores past data for joint training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CAFE, an MMD baseline and the source of the contrastive alignment loss used for modality fusion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BMR, a mixture-of-experts MMD baseline that DaedCmd is applied to and compared with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GAMED, the state-of-the-art MMD baseline used for the main ablations and case studies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"EANN, the source of the single-pass clustering procedure that splits the datasets into continual subsets."}],"review_version":1}