{"id":"5a0200eb-718c-4659-a9ef-607bf46672f0","arxiv_id":"2501.06764","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Targeted Pareto gradient integration across fusion levels improves multimodal fake news detection accuracy on FakeSV and FVC.","lead":"This paper proposes MTPareto, a training method that treats each layer of a multimodal fusion network as a separate task and combines their gradients using Pareto optimization with conflict thresholds. The authors report accuracy gains of 2.40% and 1.89% over their own hierarchical fusion model on two fake news video datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FVC evaluation may leak near-duplicate videos across folds: the paper's own Section IV-A states only 200/180 unique videos, yet five-fold CV is run without deduplication; the 1.89% FVC gain may be an artifact.","rationale":"The reader's weakest_assumption (test-set threshold selection on FakeSV) is real and I agree it prevents the 2.40% number from being fully verified. However, I think a more load-bearing problem sits in the FVC evaluation, because it is an internal inconsistency: Section IV-A states the corpus has 6,415 rows but only 200/180 unique videos, yet Section IV-A2 describes 'five-fold cross-validation' without any deduplication or grouped split. If rows are split randomly, near-duplicates leak between train and test and the 94.35% FVC accuracy and 1.89% TPareto gain cannot be interpreted. This is checkable by a grouped split; if the gain persists, the concern is resolved. I also note the secondary issues raised by the reader (no error bars, no MMPareto comparison, no code), but the two evaluation-set artifacts are the load-bearing ones. The verdict stays CONDITIONAL, not because the method is implausible, but because both headline improvements need a clean re-evaluation before being treated as established.","tokens_in":8207,"tokens_out":9645,"duration_ms":95302,"concrete_test":"Re-run the FVC experiments in Table I with a five-fold split computed on unique video IDs: all near-duplicate rows of one unique video stay in the same fold (equivalently, evaluate on the 200 fake + 180 real unique videos) and recompute TPareto vs. w/o-TPareto accuracy. If the 1.89% FVC gain disappears or falls inside seed noise, the FVC half of the central claim is an artifact of duplicate leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Both headline improvements are vulnerable to evaluation-set artifacts. On FVC, Section IV-A reports 3,957 fake + 2,458 real videos but only 200 unique fake and 180 unique real videos; Section IV-A2 then says \"we use five-fold cross-validation\" with no mention of grouping by unique video ID or deduplication. A random split over 6,415 rows puts near-duplicate uploads of the same original video in both training and test, letting the model memorize video identity. The reported 94.35% FVC accuracy and the 1.89% TPareto gain in Table I would then not measure fake-news detection; this is an internal inconsistency, not a missing baseline. On FakeSV, the TPareto hyperparameters gamma=0.25 and k=1 are chosen in Section IV-C (Tables III-IV) by accuracy on the same test set used for the headline in Section IV-B, without a validation split or repeated seeds; with 16 grid cells and 0.5-1% differences, best-cell selection can manufacture the 2.40% gain. Both issues must be fixed before the central claim is supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MTPareto, a framework for multimodal fake news detection from short videos. A hierarchical fusion network combines text, audio, image, and extra features through three fusion modules, and a TPareto optimization algorithm computes gradient weights via a quadratic program (Eqs. (4)-(5)) for the shared parameters of each fusion module, with all-modal gradients dominating and heuristic truncation for weight and angle conflicts. The reported experiments on FakeSV and FVC claim that TPareto improves accuracy over the same architecture without it by 2.40% and 1.89%, respectively, and that the full framework outperforms the compared baselines.","tokens_in":8426,"tokens_out":4118,"duration_ms":39801,"significance":"The methodological idea of treating hierarchical fusion levels as multi-task objectives and using Pareto gradient integration with a designated all-modal objective is interesting and plausibly useful. The paper's ablation results are encouraging: across Fusion Levels I-III in Table II, TPareto consistently improves over the same architecture without TPareto, and the Pareto weighting is defined by a gradient-space optimization rather than directly from final accuracies. These strengths make the central claim worth taking seriously. However, the evaluation protocol currently has two load-bearing weaknesses: potential leakage through near-duplicate videos in FVC and hyperparameter selection on the same test set used for the headline FakeSV number, so the empirical support for the claimed gains is not yet conclusive.","major_comments":[{"comment":"Section IV-A1 reports that FVC contains only 200 unique fake and 180 unique real videos, while Section IV-A2 says five-fold cross-validation is used without mentioning any grouping or deduplication. With random splits over the 6,415 rows, near-duplicate uploads of the same original video appear in both training and test, so the FVC results in Table I (94.35% accuracy; 1.89% TPareto gain) may partly measure memorization of video identity rather than generalization to unseen content. The authors should run grouped cross-validation by unique video ID or otherwise demonstrate that duplicates do not cross folds, and report those results.","section":"Section IV-A and IV-A2"},{"comment":"Section IV-C selects the angle cosine threshold (0.25) and non-all-modal weight threshold (1) by inspecting accuracy on FakeSV, and the same FakeSV test set is then used for the headline numbers in Table I. Tables III and IV show 0.5-1.0% differences across 16 grid cells, so picking the best cell can manufacture part of the claimed 2.40% improvement. The authors should select thresholds on a validation split or nested cross-validation and report test-set results for the chosen configuration, ideally with multiple seeds.","section":"Section IV-C, Tables III-IV"},{"comment":"All headline numbers in Tables I and II are single runs with no standard deviations, confidence intervals, or significance tests. Given the small reported margins in the exploration study and the variance typical in this setting, the improvements should be accompanied by multiple-seed statistics to support the claim of a significant improvement in Section V.","section":"Tables I and II"}],"minor_comments":[{"comment":"The phrase 'through concentration' should be 'through concatenation' to correctly describe the feature fusion operation.","section":"Section II-B.1"},{"comment":"The text contains typos: 'donated as' should be 'denoted as', and 'minimal site effect' should be 'minimal side effect'.","section":"Section III-B"},{"comment":"The clipping notation introduced in Eq. (3) is cramped and unclear; please define the range explicitly, for example by writing the clip operation as a projection onto [0,1].","section":"Equation (3)"},{"comment":"There is a formatting issue in the TikTec row for FVC f1: '77. 02' contains a stray space and should be '77.02'.","section":"Table I"},{"comment":"The caption phrase 'Compared with the normal additive method leading to an unexpected deviation' is hard to parse; please rephrase to state clearly what deviation the TPareto integration prevents.","section":"Figure 1 caption"},{"comment":"The paper does not mention code or data release; for reproducibility of the threshold choices and the data splits, the authors should provide code and exact split definitions.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The name MTPareto is very close to the cited MMPareto [26]; the authors should make the distinct contribution explicit in the revision. Also, if the FVC deduplication issue cannot be fixed with the available data, the paper's scope may need to be narrowed to the FakeSV results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is a legitimate extension of Pareto MTL: treat each intermediate fusion level as a task, then do a gradient integration that prioritizes the all-modal objective while using angle and weight thresholds to control interference from lower-level losses. The formulation is clean, the ablation across fusion levels is consistent with the story, and the paper is honest that its base architecture is slightly worse than MMAD without TPareto. That part earns credit.\n\nThe soft spots are mostly in the evaluation, and they are serious. On FVC, the paper itself reports only 200 unique fake and 180 unique real videos out of 6,415 total, then says five-fold cross-validation is used without any mention of grouping by unique video. A random split will put near-duplicate uploads of the same original video in both training and test, letting the model memorize video identity. That alone could explain the 94.35% accuracy and the 1.89% TPareto gain. This is an internal inconsistency, not a missing baseline, and it has to be fixed before the FVC result means anything.\n\nOn FakeSV, the two TPareto hyperparameters (cosine threshold 0.25, weight threshold 1) are chosen in Section IV-C by inspecting accuracy on the same test set used for the headline 2.40% gain. With a 4×4 grid, best-cell selection can manufacture a couple of points even when the true effect is noise. The authors need a held-out validation split or repeated-seed means with error bars. There are also no significance tests anywhere, no comparison against the closest cited method MMPareto, and no code or data released.\n\nNone of this proves the method is wrong. The within-architecture ablation showing gains at each fusion level is a good sign, and the Pareto optimization machinery is standard. But the empirical claims as presented are not verifiable, and the FVC leak is load-bearing. This paper is worth a serious referee — the idea is novel enough and the writing is clear — but the evaluation needs to be redone: deduplicate FVC by unique video, select thresholds on a validation split, and report variance. If the results survive that, the method is a solid incremental contribution; in its current form, I would not quote the numbers.","headline":"Useful idea, unconvincing numbers: the TPareto fusion-level gradient integration is a clean extension of Pareto MTL, but FVC suffers from near-duplicate leakage and FakeSV hyperparameters are tuned on the test set, so the headline gains are unverified.","tokens_in":8974,"tokens_out":2489,"would_cite":false,"duration_ms":26014,"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":"Targeted Pareto gradient integration across three fusion levels lets a multimodal fake news detector outperform baselines, with 2.40% and 1.89% accuracy gains over its own hierarchical network.","keywords":["multimodal fake news detection","Pareto optimization","gradient integration","hierarchical fusion","fake news video detection","multi-task learning","cross-modal conflict"],"falsifier":"Retrain the hierarchical fusion network with and without TPareto on FakeSV while selecting $\\gamma$ and $k$ only on a validation split, then report test accuracy; if the 2.40% gap shrinks to near zero or reverses, the central claim fails. The same check on FVC with thresholds fixed in advance would confirm whether the improvement transfers.","tokens_in":7983,"feed_emoji":"📰","tokens_out":7176,"duration_ms":65723,"temperature":0.7,"pith_summary":"Multimodal fake news detectors often optimize worse as they add modalities, because gradients from text, audio, and images point in conflicting directions and plain additive training degrades fusion. This paper proposes MTPareto, which treats each hierarchical fusion stage as a task with its own loss and integrates the per-stage gradients using a Targeted Pareto rule. The rule keeps the all-modal objective dominant by rescaling every gradient against the all-modal gradient, discarding intermediate gradients whose direction is too far from it and truncating their combined weight. On the FakeSV and FVC datasets, the authors report that this rule raises accuracy by 2.40% and 1.89% over the same architecture without it, and that the full framework beats prior baselines.","feed_headline":"Targeted Pareto fusion lifts fake-news video accuracy 2.4%","feed_subtitle":"The method beats strong baselines by steering hierarchical fusion gradients toward the all-modal objective.","key_machinery":"The central object is the TPareto gradient integration rule for a hierarchical fusion network. For each fusion module shared by several level losses, the algorithm computes Pareto-optimal weights for the candidate gradients by minimizing the norm of their weighted sum, using Frank-Wolfe iterations with an analytic two-task line search, then rewrites the final gradient so the all-modal gradient's weight is forced to 1: non-all-modal gradients are included as ratios $\\alpha_i^j / \\alpha_{\\mathrm{All}}^j$ only if their cosine with the all-modal gradient is above a threshold $\\gamma$, and their total relative weight is truncated at a threshold $k$. These two thresholds — set to $0.25$ and $1$ in the experiments — are what make the integration 'targeted': they block intermediate fusion levels from dominating or pulling shared parameters away from the all-modal objective.","core_discovery":"On the paper's own terms, the central discovery is that optimization conflicts inside a multimodal fusion network are not a fixed cost but a gradient-integration problem with a targeted remedy. The authors design a three-level hierarchical fusion network—text plus audio, then adding image, then adding extra information—and assign a classification loss to each fused level. During backpropagation, each fusion module has several candidate gradients, one per level loss; TPareto solves a Pareto weighting problem per module and then rescales all weights relative to the all-modal gradient, whose weight is fixed at 1. Intermediate-level gradients are kept only when their cosine angle to the all-modal gradient exceeds 0.25 and their summed relative weight stays below 1. The result reported is that each intermediate fusion level improves and the final all-modal classifier reaches 84.50% accuracy on FakeSV and 94.35% on FVC, compared with 82.10% and 92.46% without TPareto.","pith_inferences":["An implication the paper leaves implicit is that the threshold pair ($\\gamma=0.25$, $k=1$) was selected on FakeSV by inspecting FakeSV accuracy, so the honest generalization test is whether the same pair transfers to new data; the FVC result is suggestive but does not fully rule out tuning effects.","A testable extension would be to replace the fixed thresholds with a small validation-driven schedule or a soft penalty on gradient conflict, which could remove the need to hand-tune these two numbers.","The same all-modal-anchored Pareto integration could apply to other multi-stage fusion problems where intermediate fusion levels arise naturally, such as multimodal sentiment analysis, video captioning, or medical image fusion, all of which exhibit analogous cross-modal optimization conflicts.","A sharper test of the mechanism would compare TPareto against simple gradient clipping or fixed loss weighting at the same thresholds; if those reproduce the gain, the Pareto solution of the weights may not be the active ingredient."],"forward_implications":["If TPareto's reported gains are genuine, then the framework's accuracy improvement over its own no-TPareto network is 2.40 percentage points on FakeSV and 1.89 on FVC.","The full MTPareto framework outperforms all four baselines—TikTec, FANVN, SV-FEND, and MMAD—on both datasets across accuracy, f1, recall, and precision.","The FakeSV ablation shows TPareto improves every fusion level, with level-I accuracy rising to 82.84%, level-II to 84.23%, and level-III to 84.50%.","Because the all-modal gradient always carries weight 1, the intermediate fusion objectives can assist but never displace the final all-modal objective in shared parameters.","The exploration study indicates the two thresholds are tunable and that the chosen values give the best all-modal accuracy in the tested ranges."],"supporting_citations":[{"why":"Supplies the multi-task-as-multi-objective formulation and the quadratic gradient-weighting problem that TPareto solves.","marker":"[23]"},{"why":"Establishes the Pareto multi-task optimization view that lets fusion levels be treated as tasks.","marker":"[21]"},{"why":"Provides the Frank-Wolfe algorithm used to obtain Pareto weights when more than two gradients are integrated.","marker":"[30]"},{"why":"Introduces the FakeSV benchmark and the SV-FEND baseline that the framework compares against and outperforms.","marker":"[16]"},{"why":"Provides the strongest baseline (MMAD), which the framework surpasses only when TPareto is added.","marker":"[17]"},{"why":"Supplies the FVC corpus of debunked and verified user-generated videos used for the second evaluation.","marker":"[31]"},{"why":"Defines the multi-head attention mechanism used inside the fusion modules.","marker":"[27]"},{"why":"Supplies the cross-attention design used in the two-stream text-audio fusion module.","marker":"[28]"},{"why":"Introduces the gated multimodal unit that gates auxiliary information in the third fusion level.","marker":"[29]"}],"fun_headline_variants":["Targeted Pareto fusion lifts fake-news accuracy 2.4%","Pareto gradient fix beats fusion conflicts, adds 2.4% accuracy","Hierarchical Pareto boosts fake-news video detection 2.4%","Selective gradient steering wins 2.4% on multimodal fake-news"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two TPareto thresholds—angle cosine 0.25 and non-all-modal weight 1—chosen by looking at accuracy on FakeSV, also work on new data and on FVC; if those thresholds were effectively tuned to FakeSV's test split, the 2.40% gain could reflect overfitting rather than a general property of the algorithm.","fun_headline_variants_meta":{"raw":{"variants":["Targeted Pareto fusion lifts fake-news accuracy 2.4%","Pareto gradient fix beats fusion conflicts, adds 2.4% accuracy","Hierarchical Pareto boosts fake-news video detection 2.4%","Selective gradient steering wins 2.4% on multimodal fake-news"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000874,"raw_usage":{"total_tokens":3758,"prompt_tokens":895,"completion_tokens":2863,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":2781}},"tokens_in":511,"tokens_out":2863,"duration_ms":22088,"temperature":1.0,"reasoning_tokens":2781,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:18.876015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the hierarchical fusion network with and without TPareto on FakeSV while selecting $\\gamma$ and $k$ only on a validation split, then report test accuracy; if the 2.40% gap shrinks to near zero or reverses, the central claim fails. The same check on FVC with thresholds fixed in advance would confirm whether the improvement transfers.","supporting_citations":[{"cited_title":"Pareto multi-task learning,","cited_arxiv_id":null,"evidence_quote":"Establishes the Pareto multi-task optimization view that lets fusion levels be treated as tasks."},{"cited_title":"Revisiting frank-wolfe: Projection-free sparse convex opti- mization,","cited_arxiv_id":null,"evidence_quote":"Provides the Frank-Wolfe algorithm used to obtain Pareto weights when more than two gradients are integrated."},{"cited_title":"Fakesv: A multimodal benchmark with rich social context for fake news detection on short video platforms,","cited_arxiv_id":null,"evidence_quote":"Introduces the FakeSV benchmark and the SV-FEND baseline that the framework compares against and outperforms."},{"cited_title":"Mitigating world biases: A multimodal multi-view debiasing framework for fake news video detection,","cited_arxiv_id":null,"evidence_quote":"Provides the strongest baseline (MMAD), which the framework surpasses only when TPareto is added."},{"cited_title":"A corpus of debunked and verified user-generated videos,","cited_arxiv_id":null,"evidence_quote":"Supplies the FVC corpus of debunked and verified user-generated videos used for the second evaluation."}],"review_version":1}