{"id":"85fadb4d-a802-4b40-afc9-4507f62794ee","arxiv_id":"2412.05843","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A contrastive-learning and LLM-based multimodal model reports 88.9% accuracy on Fakeddit, slightly ahead of smaller baselines but with an unclear role for the contrastive component.","lead":"This paper builds a fake-news detector by stacking contrastive image learning, a Q-Former visual alignment module, and the Vicuna language model, then reports 88.9 percent accuracy on the public Fakeddit dataset. The result is a few points better than the strongest baselines it compares against, but the comparison is confounded by the large language model's scale.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own fusion equations exclude the contrastive features, so the claimed self-learning contribution has no defined path into the classifier.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: the paper never demonstrates that the contrastive learning module's features are used by the classifier. This is confirmed by the paper's own Section III.B, which replaces contrastive image features with Q-Former features, and by Equation 3, where only etext and f_Q-Former(eimg) enter the prompt. The ablation study is too coarse to rescue the claim, because Experiment B adds the contrastive module without specifying how it connects to the fusion path, and Experiment C shows the infuse module actually reduces F1-score relative to Experiment B. Without a defined information path or a proper ablation that isolates the contrastive encoder's output, the reported 88.88% accuracy is not attributable to the proposed self-learning mechanism. Secondary issues, including the 7B Vicuna scale confound, the filtered dataset without comparability checks, and the absence of error bars, all reinforce rejection but are not needed for the core concern. I found no independent support such as a code release or machine-checked proof that would mitigate this structural gap. The verdict should remain REJECT, so the reader's verdict is unchanged.","tokens_in":11115,"tokens_out":2455,"duration_ms":26122,"concrete_test":"Re-implement or obtain the authors' code and trace the forward pass: determine whether the output of the contrastive encoder in Section III.A is passed into Q-Former or into Equation 3 at all. Then run three configurations on the same filtered Fakeddit split: (i) the full proposed model, (ii) the same model with the contrastive loss L1 removed and the contrastive encoder disconnected from all inputs to the fusion module, and (iii) a MiniGPT-4/Q-Former + Vicuna model with identical prompts but no contrastive module. If (i) matches (ii) and (iii) within noise, the contrastive self-learning component is decorative and the central claim fails; if (i) significantly outperforms both, the claimed contribution is real.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the self-learning contrastive module is responsible for the reported state-of-the-art accuracy. That requires the contrastive module's learned image representation to actually influence the final prediction. The paper's architecture does not provide such a path. Section III.B explicitly states: 'Instead of using the image features from contrastive learning module, we would like to leverage Q-Former... to bridge the relationship between image and text.' Equations 3 and 4 then define the classifier input as eprompt = etext + f_Q-Former(eimg) and E = LLM(eprompt), with no term originating from the contrastive encoder trained in Section III.A. If the Q-Former operates on a frozen image encoder while the contrastive module trains a separate ViT, the contrastive features are simply not consumed by the fusion or classification modules. The ablations do not resolve this: Experiment B adds the contrastive learning module and improves F1 from 85.16 to 86.02, but if the contrastive features never enter the classifier, this improvement must come from an unspecified interaction, such as weight sharing or a changed training signal, that the paper never describes. Experiment C adds the 'infuse module' and actually lowers F1 from 86.02 to 85.90 while also lowering recall from 91.45 to 85.40. Thus the paper does not establish that the self-learning contrastive component contributes to the headline result; the gain could plausibly come entirely from the 7B Vicuna backbone and Q-Former. The absence of a code release or a data-flow diagram with explicit tensor paths makes this a structural gap in the central argument, not merely a missing detail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multimodal fake-news detector for image-text pairs, combining a contrastive-learning image encoder with a Q-Former-based fusion module, a Vicuna-7B LLM, and a two-layer MLP classifier. The training loss is an uncertainty-weighted sum of InfoNCE and classification losses. Experiments on a filtered subset of Fakeddit report accuracy 88.88%, precision 86.40%, recall 85.40%, and F1-score 85.90%, which the authors claim outperforms EANN, CAFE, SpotFake, SpotFake+, MVAE, HMCAN, and VERITE. The paper also includes single-modality and module-level ablations.","tokens_in":11383,"tokens_out":8828,"duration_ms":81562,"significance":"The problem is important, and the use of a public benchmark and pre-trained components are strengths. The reported gains over the listed baselines are large, but the comparison is confounded by model scale and by the filtered, nonstandard benchmark. The central claim that the self-supervised contrastive module materially contributes to the final prediction is not currently supported: the fusion equations do not show a path from the contrastive module into the classifier, and the full-model F1 is lower than one of the ablated configurations. If these points are resolved, the paper would be a useful contribution to LLM-based multimodal fake-news detection; as written, the evidence is suggestive but not conclusive.","major_comments":[{"comment":"The classification input is defined as eprompt = etext + f_Q-Former(eimg) and E = LLM(eprompt), with no term originating from the contrastive module described in Section III.A. The text explicitly states, 'Instead of using the image features from contrastive learning module, we would like to leverage Q-Former...', and Figure 2 shows the contrastive module as a separate component. If the Q-Former's frozen image encoder is not the ViT trained with the InfoNCE loss, then the contrastive module cannot influence the final prediction, contradicting Contribution 1 and the paper's title. If the Q-Former uses the contrastively pretrained ViT as its frozen encoder, that coupling must be stated explicitly; currently the architecture is ambiguous on this load-bearing point.","section":"III.B, Eqs. (3)-(4)"},{"comment":"The full model (Experiment C) reports F1 85.90, which is lower than Experiment B's 86.02, and recall drops from 91.45 to 85.40 when the infuse module is added. Only accuracy improves from 88.21 to 88.88. The claim in Section V.A that 'Ours' achieves the highest F1 among all configurations is therefore not supported by the ablation, and the conclusion that the learnable alignment module significantly improves the model is at best metric-dependent. This needs to be discussed and reconciled.","section":"Table IV"},{"comment":"The state-of-the-art comparison is on a filtered subset of Fakeddit from which entries without images were removed, with no evaluation on the unfiltered benchmark or comparison to previously reported Fakeddit numbers. The baselines are all substantially smaller models; the paper itself notes model scale as a factor in Section V.A. Moreover, no error bars, multiple seeds, or statistical significance tests are reported. The claim of superiority over 'state-of-the-art' is therefore not yet established; a same-scale LLM-based baseline and variance estimates are needed.","section":"IV.A, IV.C, Table II"}],"minor_comments":[{"comment":"Reference [15] does not describe the CAFE fake-news model, and references [30] and [31] are the same SpotFake citation; please correct and deduplicate.","section":"References [15], [30], [31]"},{"comment":"The manuscript omits the optimizer, learning rate, scheduler, and fine-tuning details for Vicuna, and the reported feature dimension 'batch size × 12 × 1048' is unexplained; please provide full reproducibility details.","section":"IV.B"},{"comment":"The four prompts are randomly selected during training; report whether the selection is per sample or per batch/epoch and add a prompt-sensitivity analysis, since the prompts are a hand-written design choice.","section":"III.D"},{"comment":"Typographical errors such as 'LLMs are excel at this task', 'authentic of news', 'the the', 'Monument configuration' in Fig. 3, and 'approapriate' should be corrected.","section":"Throughout"},{"comment":"The second term appears to be written as 1/σ2^2 times L2 rather than 1/(2σ2^2) times L2; if this is intentional, explain the deviation from the AWL formulation in [17].","section":"Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The reader's stress-test correctly identifies that the contrastive module has no defined forward-path into the classifier. My recommendation of major_revision is based on the possibility that the Q-Former's frozen image encoder is the contrastively pretrained ViT; if that is not the case, the central claim collapses and the paper should be rejected. I would ask the editor to require the authors to clarify this point decisively, ideally with an explicit diagram of weight sharing and a statement of which components are frozen or fine-tuned. The wrong CAFE citation and duplicated SpotFake reference also suggest the manuscript was not carefully proofread."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper reports 88.88% accuracy on a filtered Fakeddit split, which beats the strongest baseline by about 4 points. That's a legitimate empirical result, but the paper's central claim—that a self-learning contrastive module drives the improvement—is not supported by the architecture it describes. In Section III.B the authors explicitly write 'Instead of using the image features from contrastive learning module' and define the classifier input as e_text + f_Q-Former(e_img). No term from the contrastive encoder appears in Equations 3 and 4. The ablation shows adding the contrastive module (Experiment A to B) improves F1 from 85.16 to 86.02, but with no tensor path connecting the contrastive features to the classifier, that gain is unexplained. Adding the 'infuse module' (Experiment C) actually lowers F1 to 85.90. So the headline contribution is decorative, and the paper reduces to a MiniGPT-4-style classifier with a 7B Vicuna backbone.\n\nWhat's good: the paper is clearly written, the experimental setup is reasonably detailed, and the multimodal ablation (image vs text vs both) is a useful sanity check. The baseline comparison, while unfair in scale, includes several standard methods and the numbers are not obviously fabricated.\n\nThe soft spots are serious, though. No error bars, so the 4-point gap over VERITE could be within noise. The dataset is filtered to remove entries without images, with no check against the unfiltered benchmark. The baselines are much smaller models, so the comparison is confounded. There are citation errors: the CAFE reference [15] points to a federated learning paper, and SpotFake+ is cited to the same paper as SpotFake. These are fixable, but they add to the impression of carelessness.\n\nWho is this for? A reader who wants a quick survey of how MiniGPT-4-style classifiers perform on Fakeddit might skim it, but anyone looking for evidence that contrastive self-learning improves fake news detection will be disappointed. The paper deserves a serious referee because the empirical question is real and the issues are identifiable, but it needs major revision: either show that the contrastive module actually shapes the image encoder used by Q-Former, or drop the claim and compare at matched scale.\n\nRecommendation: send it to peer review, but expect the reviewers to catch the missing data-flow path. If the authors can clarify the interaction, there might be a decent empirical paper here.","headline":"The accuracy gain is real but the claimed self-learning mechanism has no path into the classifier, so the paper's central contribution is unsupported.","tokens_in":12007,"tokens_out":3370,"would_cite":false,"duration_ms":30557,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a three-component fake news detector—contrastive image learning, Q-Former fusion into a Vicuna-7B language model, and a two-layer classifier—reaches 88.88% accuracy on a filtered Fakeddit split, beating eight prior…","keywords":["fake news detection","multimodal classification","contrastive learning","large language models","Q-Former","Fakeddit","image-text alignment","automatic weighted loss"],"falsifier":"Run the same experiment with the contrastive module removed while keeping all data augmentation: if accuracy and F1-score do not drop, the contrastive branch is not load-bearing. Tracing gradients or input connections from the classification loss back through the fusion module would also settle whether the contrastive head can affect the prediction at all.","tokens_in":10892,"feed_emoji":"📰","tokens_out":11536,"duration_ms":96635,"temperature":0.7,"pith_summary":"The paper proposes a multimodal fake news classifier that combines contrastive learning on images with a large-language-model fusion stage, and claims it outperforms eight prior detectors on the Fakeddit benchmark. On a filtered Fakeddit split, the model reports accuracy of 88.88%, precision of 86.40%, recall of 85.40%, and F1-score of 85.90%, beating the best baseline by roughly four accuracy points. The authors argue that contrastive learning helps in settings with limited labeled image data, that the learnable-query alignment is what lets the language model use visual evidence, and that an automatically weighted loss keeps the contrastive and classification objectives balanced. If these claims hold, the method would be a new state of the art for paired image-text misinformation detection and a template for combining self-supervised visual learning with LLM reasoning.","feed_headline":"Fake news detector hits 88.88% accuracy on image-text pairs","feed_subtitle":"A 7-billion-parameter multimodal LLM with contrastive pretraining claims new highs on Fakeddit.","key_machinery":"The load-bearing machinery is a three-stage pipeline. A contrastive learning module trains a ViT image encoder under InfoNCE loss with a momentum encoder update $y_t = m y_{t-1} + (1-m)x_t$; a multimodal fusion module uses Q-Former's learnable queries to align image features with text and feeds the combined prompt into the Vicuna-7B language model; and a two-layer MLP with GELU produces the true/false output. The two losses are merged by an automatic weighted loss $\\mathcal{L} \\approx \\frac{1}{2\\sigma_1^2}\\mathcal{L}_1 + \\frac{1}{\\sigma_2^2}\\mathcal{L}_2 + \\log(1+\\sigma_1) + \\log(1+\\sigma_2)$, with learnable uncertainties $\\sigma_1,\\sigma_2$. The fusion stage uses Q-Former features rather than the contrastive module's output, so the contrastive branch and the classifier are trained under a shared total loss rather than through a direct feature path.","core_discovery":"On its own terms, the central discovery is that a three-component model—a contrastive image feature learner, a Q-Former-based multimodal fusion into a pre-trained large language model, and a small classification head—can classify fake news from paired image-text posts at 88.88% accuracy on Fakeddit, with balanced precision (86.40%), recall (85.40%), and F1-score (85.90%). The paper's ablation attributes the gains to multimodal fusion over single-modality input, and to the addition of the contrastive and infuse modules, with accuracy rising from 87.16% to 88.21% to 88.88%. The authors further claim the approach is data-efficient because contrastive learning exploits unlabeled images and the language model requires only modest fine-tuning.","pith_inferences":["A direct test of the paper's self-learning claim would be to route the contrastive encoder's features into the Q-Former alongside the frozen image encoder and re-measure accuracy; if accuracy rises, the contrastive branch is load-bearing, and if not, the gains likely come from the language model backbone and prompts.","The reported results are on a filtered Fakeddit split with entries lacking valid images removed, so an independent check would re-run every baseline on this exact filtered split before comparing with published numbers from the unfiltered benchmark.","The method's randomized use of four paraphrased questions as prompts is a general regularization idea for vision-language classification, though the paper does not ablate the number or wording of prompts."],"forward_implications":["A large multimodal LLM fine-tuned with 563k training pairs can reach 88.88% accuracy on a filtered Fakeddit split, suggesting that model scale and prompt-based fusion are decisive for this task.","Multimodal fusion contributes about 5 to 7 accuracy points over text-only or image-only inputs on this benchmark (88.88% vs. 83.18% and 79.68%).","The ablation sequence (LLM only, then plus contrastive, then plus infuse) implies each added module improves accuracy, from 87.16% to 88.21% to 88.88%.","Automatically weighting the contrastive and classification losses avoids manual tuning and is presented as a way to fine-tune the large language model jointly with the image encoder.","Because the contrastive module is trained without labels, the same architecture could transfer to domains where labeled fake news examples are scarce."],"supporting_citations":[{"why":"The Fakeddit benchmark dataset supplies the paired image-text news posts used for training, validation, and testing.","marker":"[22]"},{"why":"Defines the Q-Former with learnable queries that the fusion module uses to align image features with text.","marker":"[18]"},{"why":"Provides the Vicuna-7B large language model that carries the multimodal reasoning and classification.","marker":"[6]"},{"why":"Introduces the MiniGPT-4 framework on which the paper's fusion architecture is based.","marker":"[37]"},{"why":"Supplies the uncertainty-based automatic weighted loss that combines the contrastive and classification losses.","marker":"[17]"},{"why":"VERITE is the strongest CLIP-based baseline in the comparison, used to establish the proposed model's relative gain.","marker":"[24]"},{"why":"CAFE is a BERT/ResNet baseline with unified embedding space, one of the top comparators the model claims to beat.","marker":"[15]"},{"why":"SpotFake+ is the attention-enhanced version of SpotFake in the baseline table.","marker":"[30]"}],"fun_headline_variants":["Self-learning multimodal model detects fake news with 88.88% accuracy","Contrastive learning plus LLMs spots fake image-text pairs at 88.88%","Self-learning model achieves 88.88% on Fakeddit without labels","Contrastive pretraining + LLM fusion hits 88.88% fake news detection","Data-efficient fake news detector: 88.88% accuracy on image-text pairs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the contrastive learning module's image features actually reach the final classifier; the paper says fusion uses Q-Former features instead of the contrastive module's features, so if the contrastive branch never feeds the classifier, the reported accuracy gains cannot be attributed to self-learning.","fun_headline_variants_meta":{"raw":{"variants":["Self-learning multimodal model detects fake news with 88.88% accuracy","Contrastive learning plus LLMs spots fake image-text pairs at 88.88%","Self-learning model achieves 88.88% on Fakeddit without labels","Contrastive pretraining + LLM fusion hits 88.88% fake news detection","Data-efficient fake news detector: 88.88% accuracy on image-text pairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000782,"raw_usage":{"total_tokens":3426,"prompt_tokens":889,"completion_tokens":2537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":2430}},"tokens_in":505,"tokens_out":2537,"duration_ms":20054,"temperature":1.0,"reasoning_tokens":2430,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:17:04.355982+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same experiment with the contrastive module removed while keeping all data augmentation: if accuracy and F1-score do not drop, the contrastive branch is not load-bearing. Tracing gradients or input connections from the classification loss back through the fusion module would also settle whether the contrastive head can affect the prediction at all.","supporting_citations":[{"cited_title":"Fakeddit: A new multimodal benchmark dataset for fine-grained fake news detection","cited_arxiv_id":null,"evidence_quote":"The Fakeddit benchmark dataset supplies the paired image-text news posts used for training, validation, and testing."},{"cited_title":"Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models","cited_arxiv_id":null,"evidence_quote":"Defines the Q-Former with learnable queries that the fusion module uses to align image features with text."},{"cited_title":"Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, march 2023","cited_arxiv_id":null,"evidence_quote":"Provides the Vicuna-7B large language model that carries the multimodal reasoning and classification."},{"cited_title":"Petrantonakis","cited_arxiv_id":null,"evidence_quote":"VERITE is the strongest CLIP-based baseline in the comparison, used to establish the proposed model's relative gain."},{"cited_title":"Cafe: Catastrophic data leakage in vertical federated learning","cited_arxiv_id":null,"evidence_quote":"CAFE is a BERT/ResNet baseline with unified embedding space, one of the top comparators the model claims to beat."},{"cited_title":"Spotfake: A multi-modal framework for fake news detection","cited_arxiv_id":null,"evidence_quote":"SpotFake+ is the attention-enhanced version of SpotFake in the baseline table."}],"review_version":1}