{"id":"a98d3c65-5e6d-4e12-bc5d-607018cd339b","arxiv_id":"2508.19639","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"FakeSV-VLM reaches 90.22% and 89.30% accuracy on FakeSV and FakeTT by adding a two-stage MoE adapter and contrastive alignment to InternVL2.5-8B.","lead":"FakeSV-VLM is a system that fine-tunes a vision-language model with a mixture-of-experts adapter and a cross-modal contrastive loss to automatically judge whether short news videos are real or fake, reporting top accuracy on two public benchmarks. It matters because platforms need scalable, accurate fact-checking as fake short videos multiply, and this work shows a practical recipe for adapting general-purpose VLMs to that task.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Attribution MoE's hard-argmax router is unsupervised and receives no gradient, so the claimed four-way manipulation-type inference cannot be learned as written; this, not ADEC alone, breaks the central PMOE mechanism claim.","rationale":"FakeSV-VLM's headline claim is SOTA accuracy plus a mechanistic explanation ('PMOE and ADEC enable robust manipulation reasoning and event-level cross-modal consistency checking'). The mechanistic explanation is the scientific content; if it is wrong, the empirical accuracy alone does not support the paper's conclusions. The Attribution MoE failure is the most load-bearing because it is not a matter of interpretation: Eq. 7's hard argmax and the absence of any loss on the attribution probabilities jointly imply zero gradient to the attribution gate. The four experts then cannot learn to correspond to the four scenarios, contradicting the method's core design and the Appendix C.5.2 validation. The ADEC issue raised by the reader is real but less decisive: fake pairs receive no direct positive/negative supervision, and the loss only aligns real pairs while using fake items as negatives; this does not implement 'inconsistency checking' as claimed, but it does provide some indirect learning signal. Because both flaws are correctable if the authors add appropriate losses/soft routing and rerun experiments, the appropriate disposition is rejection of the current version (or conditional acceptance after major revision), not acceptance. I agree with the reader's overall REJECT verdict, though my primary concern is the Attribution MoE routing rather than ADEC.","tokens_in":21029,"tokens_out":9174,"duration_ms":105223,"concrete_test":"Inspect the released code and compute torch.autograd.grad(L_total, params of the Attribution MoE gating network) on the first training batch. If the gradient norm is zero — or if the implementation actually uses soft routing / an extra four-way loss absent from §2.3 — the paper's central mechanism claim is unsupported. A complementary check: freeze the attribution gate at initialization and retrain; if FakeSV/FakeTT accuracy is unchanged, the claimed manipulation-type routing is not responsible for the reported gains.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central contribution is that PMOE performs two-stage manipulation reasoning: Detection MoE judges real/fake and Attribution MoE infers the specific manipulation type (real, visual forgery, textual forgery, full forgery). As written in §2.3, the Attribution MoE router computes softmax probabilities p_i^r, p_i^{fv}, p_i^{ft}, p_i^{fm}, selects z = argmax, and applies only Exp_z (Eqs. 7–8). No loss is defined on those four attribution probabilities: L_APG (Eq. 6) supervises only the Detection MoE's two probabilities, and L_ACL (Eq. 11) supervises the attention-pooled output, not the attribution gate. Because argmax is piecewise constant, the attribution gating network receives zero gradient from L_CE/L_ACL via Eq. 8. Thus, under the described training scheme, the four attribution experts cannot become specialized to the four forgery scenarios; the claimed manipulation-type inference is not learnable. The Appendix C.5.2 report of 82% manual-annotation accuracy for these types is not derivable from the specified losses and routing, so either the text omits a differentiable or supervised routing mechanism or that evaluation is not meaningful. Separately, the reader's ADEC objection is partially correct: fake video-text pairs have I(v,t)=0, so no term directly increases their distance; fake items are only pushed away from real anchors as in-batch negatives, which does not amount to learning event-level inconsistency. The Attribution MoE issue is the more decisive equation-level failure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FakeSV-VLM, a Vision-Language Model framework for fake short-video news detection. The method inserts learnable Artifact Tokens into an InternVL2.5-8B backbone, feeds them through a two-stage Progressive Mixture-of-Experts Adapter (PMOE) consisting of a Detection MoE and an Attribution MoE, and adds an Alignment-driven Event Checking (ADEC) contrastive module. Training uses L_total = L_CE + L_PMOE + L_ADEC. On FakeSV and FakeTT, the paper reports Accuracy gains of +3.32% and +5.02% over the previous SOTA, with ablations attributing gains to PMOE and ADEC. The paper also reports a manual-annotation evaluation of the Attribution MoE's four-way manipulation-type predictions, and includes case studies and failure cases.","tokens_in":21387,"tokens_out":3910,"duration_ms":45416,"significance":"If the reported results and mechanism claims held, this would be a meaningful advance: it would establish a new SOTA on two standard fake short-video benchmarks and provide a reusable VLM-based architecture with fine-grained manipulation attribution. The manuscript's strengths include a broad comparison against 13 baselines, repeated runs, component ablations, code release, and a candid discussion of failure cases. However, the paper's central explanatory claims are not supported by the equations defining the losses. The Attribution MoE has no supervision or gradient path for its four-way routing, and the ADEC contrastive loss, as written, minimizes an objective over positive real pairs only and does not directly learn event-level inconsistency. These are load-bearing issues: the reported gains cannot, on the evidence in the manuscript, be attributed to the proposed mechanisms.","major_comments":[{"comment":"The Attribution MoE is claimed to infer four manipulation types (real, visual forgery, textual forgery, full forgery), but no loss is defined on the four routing probabilities p_i^r, p_i^{fv}, p_i^{ft}, p_i^{fm}. Eq. (7) selects the expert by argmax and Eq. (8) applies only the selected expert. L_APG (Eq. 6) supervises the two Detection-MoE probabilities, and L_ACL (Eq. 11) supervises the attention-pooled binary output, not the attribution gate. Because argmax is piecewise constant, gradients from L_CE/L_ACL cannot train the Attribution-MoE router or the four expert networks through Eq. (8). Thus the four experts cannot specialize to the four forgery scenarios under the described training scheme. The 82% manual-annotation result in Table 7 is not derivable from the specified losses and routing; either the manuscript omits a differentiable or supervised routing mechanism, or the Attributi","section":"§2.3, Eqs. (7)–(8), and Appendix C.5.2"},{"comment":"The ADEC module is claimed to capture event-level cross-modal inconsistency, but the match label I(v,t) is defined as 1 only when the video and text belong to the same news and the news is real. For fake pairs and cross-news pairs, I(v,t)=0, so the corresponding term -I log s in Eqs. (17)–(18) is identically zero. Consequently, the loss provides no positive gradient signal that directly separates inconsistent video-text pairs; fake items enter only as negatives in the denominators of real-anchor contrastive terms. This does not amount to learning 'inconsistency' in the claimed sense. The ablation gains credited to ADEC in Table 2 therefore cannot be attributed to the stated event-level inconsistency-checking mechanism. The authors should either reformulate the contrastive loss (e.g., with an explicit mismatch class or with I=0 paired with a push-away term) or revise the claim.","section":"§2.4, Eqs. (15)–(19)"},{"comment":"Because both central mechanisms are not realized as written, the component ablation in Table 2 does not establish that the PMOE/ADEC designs improve performance for the stated reasons. The empirical gains over the fine-tuned backbone may come from the artifact tokens, the additional binary classification heads, LoRA fine-tuning, or the prompt template. As the paper stands, the novelty claim — 'the PMOE and ADEC modules enable robust manipulation reasoning and event-level cross-modal consistency checking' — is unsupported.","section":"§3.3, Table 2 and overall interpretation"}],"minor_comments":[{"comment":"Typo: 'Detetction MoE' should be 'Detection MoE'.","section":"Table 2"},{"comment":"Typo: 'The t case' should be 'The second case'.","section":"§3.4"},{"comment":"The text says 'To validate the effectiveness of CFGE' but the model is FakeSV-VLM; this appears to be a leftover from another draft.","section":"Appendix C.2"},{"comment":"Figure 1 labels and Figure 3 axis text render as unreadable '/uni...' glyph paths in the PDF; the figures need to be regenerated with proper fonts.","section":"Figure 1 and Figure 3"},{"comment":"The notation fc ⊗ A is used for both the input to LLM_{1:l} and its output, which is confusing; use a different symbol for the contextualized features.","section":"§2.3, Eq. (1)"},{"comment":"Typo: 'preceive' should be 'perceive'.","section":"§2.4"},{"comment":"The manual annotation of 100 samples lacks details on annotator agreement, selection bias, and class imbalance; even as an auxiliary evaluation, more protocol detail is needed.","section":"Appendix C.5.2"},{"comment":"These tables are only referenced as 'Appendix C.3, C.4' in the main text; the reader would benefit from explicit in-text discussion of the entropy-loss and layer-placement ablations.","section":"Tables 4–5"}],"recommendation":"reject","confidential_remarks":"The paper cites the authors' own CA-FVD work, but only as a baseline, so circularity is not the concern. The decisive issue is internal inconsistency: the equations do not implement the claimed learning mechanisms. This is not a matter of presentation; the Attribution MoE is not trainable under the stated objective and the ADEC loss does not model inconsistency. Even if the empirical numbers are reproducible, the manuscript's central scientific claims would require substantially re-designed losses and new experiments, which is beyond a normal minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked about FakeSV-VLM. Here's my reading.\n\nThe paper is an empirical system for fake-short-video news detection, building on InternVL2.5 with a two-stage MoE adapter plus a contrastive alignment loss. It reports a solid SOTA jump on FakeSV and FakeTT—about +3.3 and +5.0 accuracy over ExMRD. The architecture is a reasonable combination of established ideas: artifact tokens, sparse MoE with hard routing, LoRA tuning, and a contrastive loss. If the numbers are reproducible (code is available), this is a meaningful engineering contribution to the fake-news-detection subfield.\n\nThat said, the scientific claims do not survive a close read of the equations. Two problems stand out.\n\nFirst, the Attribution MoE, which the paper bills as the core of \"progressive\" reasoning, has no supervision for its four-way routing. The router computes softmax probabilities, selects the argmax expert, and only the selected expert's output flows to the final loss. No loss is defined on the four routing probabilities, and argmax is piecewise constant, so the gating network receives zero gradient. The experts can still be trained as separate MLPs, but the router never learns to specialize them to the four manipulation scenarios. The manual-annotation accuracy of 82% in Appendix C.5.2 is not derivable from the training objective as written. Either there's a missing auxiliary loss or a straight-through estimator, or that evaluation is meaningless.\n\nSecond, ADEC is claimed to \"perceive fake news by capturing inconsistency between modalities.\" But the match label I(v,t) is zero for every fake and cross-news pair, so those pairs contribute nothing to the loss. The contrastive normalization does push negatives away from real anchors, but fake items are never directly updated. So the loss can only tighten the real cluster; it cannot learn event-level inconsistency, and the ablation gains attributed to ADEC cannot be credited to that mechanism.\n\nThere are smaller issues: no error bars despite the claim of three runs, a baseline protocol that mixes zero-shot and fine-tuned models, some corrupted figures, and a few typos. But the two load-bearing flaws above are the real problem.\n\nMy take: the empirical direction is worth pursuing, and a corrected version—with a proper classification loss on the attribution gate and a contrastive loss that actually penalizes mismatched pairs—could be a solid paper. As written, the method description is internally inconsistent with the reported results. I would send it to a serious referee because the system is substantial and the flaws require careful checking, but I'd expect a major revision or a reject in current form. I'd bring it to a reading group as a case study in how hard routing can silently break a paper.\n\nRecommendation: engage with it if you're working on multimodal misinformation, but treat the mechanism claims with skepticism until the losses are fixed.","headline":"Solid SOTA numbers and a useful recipe, but the two central mechanism claims (Attribution MoE inferring manipulation types, ADEC capturing cross-modal inconsistency) are contradicted by the paper's own equations; deserves referee attention but not acceptance as written.","tokens_in":21878,"tokens_out":5258,"would_cite":false,"duration_ms":57314,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fine-tuning a VLM with two expert stages beats prior fake-video news detectors.","keywords":["fake news detection","short video","vision-language model","mixture of experts","artifact tokens","cross-modal consistency","contrastive learning","FakeSV/FakeTT"],"falsifier":"Look at the gradient of L_ADEC for a fake-labeled sample: since I(v,t)=0 for every fake and cross-news pair, the per-sample gradient is identically zero, so no inconsistency signal reaches the model through this loss. An experiment that replaces the matched-only contrastive loss with one that also penalizes mismatched pairs—or that measures ADEC's gradient contribution separated by label—would settle whether the reported ablation gains come from detecting inconsistencies or merely from aligning real pairs.","tokens_in":20923,"feed_emoji":"📰","tokens_out":7590,"duration_ms":78322,"temperature":0.7,"pith_summary":"The paper sets out to show that a large vision-language model, instead of being used only for zero-shot answers or to pre-generate clues, can be fine-tuned end-to-end to detect fake news in short videos. It organizes the problem into four video-text scenarios—both real, only the text fake, only the video fake, both fake—and builds a Progressive Mixture-of-Experts Adapter (PMOE) so the model first judges real/fake and then attributes the manipulation type. A second module, Alignment-driven Event Checking (ADEC), aligns visual and textual event embeddings with a contrastive objective. On the FakeSV and FakeTT benchmarks, the method reports 90.22% and 89.30% accuracy, improving over the previous best approaches by 3.32 and 5.02 percentage points, and the paper documents cases where the modules overturn mistakes made by earlier models.","feed_headline":"VLM adapter beats fake-news baselines by up to 5%","feed_subtitle":"Two-stage expert reasoning plus cross-modal checking lifts accuracy to 90.22% on FakeSV and 89.30% on FakeTT.","key_machinery":"The load-bearing machinery is the PMOE adapter on top of the fine-tuned VLM: learnable Artifact Tokens are concatenated with visual and text features, passed through early LLM layers, then routed through two sparse mixtures of experts—Detection MoE (real/fake) followed by Attribution MoE (real, visual-only fake, textual-only fake, both fake)—with authenticity probability guidance and artifact classification losses. The second piece is ADEC, which average-pools LLM-contextualized text and visual features and trains them with a symmetric contrastive loss so that matched real pairs sit close in embedding space. Together they give the VLM a manipulation-aware reasoning path plus a cross-modal co","core_discovery":"The paper argues that a pretrained vision-language model, fine-tuned end-to-end rather than used zero-shot or as a text generator, can detect fake short-video news better than dedicated baselines. It introduces two mechanisms: PMOE, which inserts learnable Artifact Tokens into early LLM layers and routes them through a real/fake Detection MoE and then a four-expert Attribution MoE (real, text-only fake, video-only fake, both fake), and ADEC, which uses contrastive learning to align pooled visual and textual event embeddings. On the FakeSV and FakeTT benchmarks, the resulting system reports 90.22% and 89.30% accuracy, surpassing the previous best methods by 3.32 and 5.02 points, with ablation","pith_inferences":["Editorial: Eqs. 17-18 assign zero weight to every fake and cross-news pair, so ADEC as written can only pull matched real pairs together; a variant with explicit negative pairs would clarify whether the ablation gain comes from alignment of real pairs or from separation of inconsistent ones.","The authors note that no fine-grained dataset exists to label which part of a video is fabricated, so the Attribution MoE's four-way diagnoses remain validated only on a 100-sample manual set; building such annotations would be a direct next step.","A testable extension is to apply the same progressive MoE adapter to out-of-context image-news pairs or audio-video mismatches, where event-level inconsistency is the dominant failure mode.","The reported margins are relative to specific prior methods; scaling beyond the 8B backbone used here may shift the gap, since the paper could not test larger VLMs under its compute budget."],"forward_implications":["Fine-tuned open-source VLMs can outperform proprietary zero-shot models on this task, so top results do not require closed commercial APIs.","Because PMOE and ADEC are not needed at inference, the accuracy gain does not add per-video compute or latency.","The four-scenario expert decomposition gives the model a way to report the manipulation type, not just the real/fake label.","The gains on both a Chinese and an English dataset suggest the recipe transfers across languages and platforms."],"supporting_citations":[{"why":"Supplies the InternVL2.5-8B backbone that all experiments fine-tune.","marker":"Chen et al., 2024"},{"why":"Provides LoRA, the parameter-efficient fine-tuning method used throughout.","marker":"Hu et al., 2022"},{"why":"Defines the sparse mixture-of-experts routing paradigm that the Detection and Attribution MoE modules build on.","marker":"Fedus et al., 2022"},{"why":"Provides the FakingRecipe baseline it must beat and the prompt format it extends with event metadata.","marker":"Bu et al., 2024"},{"why":"Introduces the FakeSV dataset and the SV-FEND baseline, and defines the temporal train/validation/test split.","marker":"Qi et al., 2023a"},{"why":"CA-FVD baseline that also uses VLMs for modality alignment; the consistency-aware comparison point.","marker":"Wang et al., 2025"},{"why":"ExMRD baseline with VLM chain-of-thought reasoning; the strongest previous method the paper claims to beat.","marker":"Hong et al., 2025"}],"fun_headline_variants":["VLM with progressive MoE adapter outdetects fake-video baselines","Progressive MoE adapter wins fake-video news detection by 5%","Mixture-of-experts adapter helps VLM spot fake short-video news","Expert-routed VLM catches fake news in short videos better"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The crucial assumption is that the ADEC module's contrastive loss teaches the model to notice mismatched video-text pairs; but in Eqs. 17-18 every unmatched or fake pair is assigned I(v,t)=0, so those pairs contribute exactly zero gradient and the module as written can only pull matched real pairs closer together, not push inconsistent pairs apart.","fun_headline_variants_meta":{"raw":{"variants":["VLM with progressive MoE adapter outdetects fake-video baselines","Progressive MoE adapter wins fake-video news detection by 5%","Mixture-of-experts adapter helps VLM spot fake short-video news","Expert-routed VLM catches fake news in short videos better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000317,"raw_usage":{"total_tokens":1687,"prompt_tokens":861,"completion_tokens":826,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":746}},"tokens_in":605,"tokens_out":826,"duration_ms":8066,"temperature":1.0,"reasoning_tokens":746,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:36:21.635332+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Look at the gradient of L_ADEC for a fake-labeled sample: since I(v,t)=0 for every fake and cross-news pair, the per-sample gradient is identically zero, so no inconsistency signal reaches the model through this loss. An experiment that replaces the matched-only contrastive loss with one that also penalizes mismatched pairs—or that measures ADEC's gradient contribution separated by label—would settle whether the reported ablation gains come from detecting inconsistencies or merely from aligning real pairs.","supporting_citations":[],"review_version":1}