{"id":"7f053264-72b4-406f-b55a-08860461d16e","arxiv_id":"2508.01402","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"ForenX detects AI-generated images with MLLMs guided by a forensic prompt and trained on a new explanation dataset, ForgReason.","lead":"ForenX uses multimodal large language models with a specialized forensic prompt to tell whether an image is AI-generated and to explain its reasoning. The paper also introduces ForgReason, a human- and LLM-curated dataset of forgery explanations, and reports gains on two benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing concern is circularity: if ForgReason annotations derive from LLM heuristics already present in the MLLM, the reported explainability may show self-consistency rather than independent forensic reasoning.","rationale":"The reader's weakest assumption exactly matches the most load-bearing concern: the curated annotations may encode the same heuristics the model already has, making explanations self-consistent rather than independently grounded. Given only the abstract, there is no way to verify the annotation process, the benchmark composition, or the subjective evaluation protocol. The concern is not that the paper is wrong, but that its central claim is currently unverified. The proposed concrete test would settle whether fine-tuning on ForgReason provides genuine forensic signal or merely reinforces pre-existing biases. Since the reader's verdict is already UNVERDICTED, this stress-test does not change that verdict; it sharpens what evidence would be needed.","tokens_in":653,"tokens_out":2322,"duration_ms":26917,"concrete_test":"Construct a held-out test set of AI-generated images from generators not represented in ForgReason, with independent expert annotations of the actual manipulation cues. Run ForenX on this set and measure explanation agreement (e.g., recall of expert-identified cues) and classification accuracy. If explanation agreement is not significantly above chance or accuracy drops to baseline, the prompt/annotation loop did not encode independent forensic knowledge.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ForenX's central contribution is that the specialized forensic prompt and ForgReason dataset improve both generalization and explanation quality. This requires the annotations to direct the model to genuinely forgery-indicative visual cues. However, the abstract states ForgReason is curated by an LLM-based agent with human annotators; if that agent shares the MLLM's inductive biases, tuning on these labels can amplify existing priors without adding forensic ground truth. The paper's subjective evaluation would then measure perceived plausibility, not actual forensic accuracy. Since only the abstract is available, no internal contradiction can be checked, but the abstract does not rule out this self-consistency loop. The risk is external: the claimed generalization and explainability may not transfer to unseen generators or to images where benchmark-correlated artifacts are absent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents ForenX, a method that uses multimodal large language models (MLLMs) for AI-generated image detection, augmented by a specialized forensic prompt and a new dataset called ForgReason. The dataset is curated through collaboration between an LLM-based agent and human annotators. The authors claim that ForenX improves generalization over standard MLLMs for forgery detection and provides explanations that are accurate, relevant, and comprehensive, as verified by subjective evaluations on two major benchmarks.","tokens_in":848,"tokens_out":2311,"duration_ms":27900,"significance":"If the claims are substantiated, ForenX would represent a meaningful step toward explainable AI-generated image detection, addressing a recognized limitation of black-box classifiers. The introduction of the ForgReason dataset with human-in-the-loop curation is a potentially useful resource for the community. However, the significance is entirely conditional at this stage: the abstract reports no quantitative results, no error bars, and no details of the evaluation protocol, so the reader cannot assess whether the claimed improvements are real or an artifact of benchmark-specific correlations or self-consistent evaluation.","major_comments":[{"comment":"The central claim that ForenX 'improves the generalization of forgery detection' is not supported by any quantitative result in the abstract. No detection accuracy, AUC, F1, or comparison against existing methods is reported, nor are error bars or significance tests mentioned. Without such evidence, the claim is unverified and cannot be evaluated by the reader.","section":"Abstract, first paragraph"},{"comment":"The statement that 'explainability is verified by comprehensive subjective evaluations' is too vague to be assessed. The abstract does not describe the number of participants, the rating scale, inter-annotator agreement, the comparison baselines, or the exact criteria for 'accurate, relevant, and comprehensive.' Without this information, the reader cannot judge whether the claimed explanation quality is meaningful or whether the evaluation was biased by the evaluators' familiarity with the method.","section":"Abstract, third paragraph"}],"minor_comments":[{"comment":"The sentence 'This approach not only enhance the generalization...' contains a subject-verb agreement error; it should read 'not only enhances.'","section":"Abstract, first paragraph"},{"comment":"The phrase 'this process provides refined data that further enhances our model's performance' is unclear about what 'refined' means operationally; specifying the curation pipeline or giving an example would improve clarity.","section":"Abstract, second paragraph"},{"comment":"The term 'two major benchmarks' is underspecified; naming the benchmarks would allow the reader to judge the generality of the evaluation.","section":"Abstract, third paragraph"}],"recommendation":"major_revision","confidential_remarks":"The submission as provided consists only of an abstract, with no full text to review. The abstract's claims are not yet supported by quantitative evidence, and the circularity risk in the dataset curation is substantial. I would recommend that the editor confirm the existence of a full manuscript and ask the authors to address the major comments above, particularly the need for quantitative generalization results and a clear demonstration that the ForgReason annotations provide independent forensic signal rather than self-consistent LLM heuristics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked for a read on arXiv:2508.01402. What's actually new here is the combination: ForenX, an MLLM-based detector steered by a specialized forensic prompt, and ForgReason, a dataset of forgery explanations built by an LLM agent plus human annotators. That pairing is sensible and fills a real gap—most detectors say \"fake\" without saying why. The claim that even a little human annotation improves explanation quality is concrete and testable, and the subjective evaluation approach for explainability is defensible.\n\nThe paper's strengths at this level are mostly design choices. Focusing on forensic cues that \"resonate with human thoughts\" is a good framing, and involving human annotators in the dataset gives the project independent grounding that a purely synthetic pipeline would lack. I also credit the authors for targeting generalization, not just benchmark accuracy.\n\nNow the soft spots. We only have the abstract, so there are no numbers, no comparisons, no ablation details, no inter-annotator agreement, and no description of the subjective evaluation protocol. The central claims are unsupported at this evidence level. That alone would normally make me hesitant to send a paper to review, but the abstract is a claim, not the whole manuscript.\n\nThe stress-test note flags circularity: if the LLM agent that curated ForgReason shares inductive biases with the MLLM being fine-tuned, gains could reflect self-consistency rather than independent forensic reasoning. That is a legitimate concern, but it is not a knockout at this stage. The abstract does not say the agent and the MLLM are the same model, and the human annotators provide a partial check. The paper will need to show that explanations track ground-truth forgery indicators, not just human-rated plausibility. I'd want to see ablations where tuning on ForgReason is compared against tuning on non-expert descriptions, and at least one test on a generator not seen in training.\n\nThe citation pattern I can't judge from the abstract. No red flags, but also no way to verify novelty against the cited literature.\n\nWho is this for? Researchers in AI-generated image detection, explainable AI, and MLLM evaluation. If the full paper ships the dataset, has quantitative comparisons, and tackles the circularity concern, it's a useful contribution to a subfield that cares about explainability.\n\nMy recommendation: do not desk reject. Send it to peer review, but tell the reviewers to focus on dataset construction, the source of the LLM agent's priors, and whether the subjective evaluation measures correctness or just fluency. If the full text holds up on those points, it's publishable. If it doesn't, the reviewers will catch it.","headline":"A plausible explainable-detection package, but the abstract alone can't support the soundness claims; the circularity worry about ForgReason deserves scrutiny in the full paper.","tokens_in":1280,"tokens_out":1559,"would_cite":false,"duration_ms":19846,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ForenX uses a forensic prompt to make multimodal language models detect AI-generated images and explain why they are fake.","keywords":["AI-generated image detection","multimodal large language models","explainable forensics","forensic prompt","ForgReason dataset","image forgery","generalization"],"falsifier":"A concrete test would be to hold out an entire family of generative models not represented in the training data, run ForenX on those images, and have independent forensic examiners judge whether the model's explanations point to the true generative artifacts (for example, the specific texture or noise patterns of that generator). If explanation accuracy collapses on such out-of-distribution images while detection accuracy stays high, the explanations are not tracking independent forensic evidence.","tokens_in":527,"feed_emoji":"🔎","tokens_out":3304,"duration_ms":39569,"temperature":0.7,"pith_summary":"The paper argues that AI-generated image detection should not stop at a binary verdict; it should also produce explanations a human examiner can follow. To that end, it proposes ForenX, a method that uses multimodal large language models (MLLMs) to analyze an image and generate a justified authenticity judgment. The key additions are a specialized forensic prompt that directs the model to forgery-indicative attributes, and ForgReason, a dataset of forgery-evidence descriptions built by an LLM-based agent and refined by human annotators. The paper claims that this combined approach improves generalization on two major benchmarks and that even a modest amount of human-annotated data substantially boosts explanation quality.","feed_headline":"Forensic prompt steers MLLMs to detect and explain AI images","feed_subtitle":"ForenX pairs a specialized prompt with human-curated evidence descriptions to boost detection generalization and explanation quality.","key_machinery":"Two coupled components carry the argument. The first is a specialized forensic prompt, a fixed instruction that steers the MLLM's attention toward forgery-indicative attributes, counteracting the model's default tendency to overlook subtle artifacts. The second is ForgReason, a dataset of image-forgery evidence descriptions produced by an LLM-based agent and then refined by human annotators; fine-tuning on this data gives the model a vocabulary and reasoning pattern for generating explanations that are accurate, relevant, and comprehensive. Together they transform a general-purpose vision-language model into a forensically aware detector with an explanatory output.","core_discovery":"The central claim is that a multimodal large language model, guided by a forensic prompt and fine-tuned on the ForgReason dataset, can both detect AI-generated images with better generalization than standard classifiers and explain its decisions in terms that resonate with human forensic reasoning. The paper reports that this method outperforms classifier baselines on two major benchmarks, and that the explanations it produces are accurate, relevant, and comprehensive according to subjective evaluation. The authors see this as closing a gap between automated detection and the kind of interpretive analysis a human forensic examiner would perform.","pith_inferences":["A natural next test would be to apply ForenX to images from generators entirely absent from the training benchmarks; the paper's explanation quality claim could break down if the model has only learned benchmark-specific artifacts.","The same forensic-prompt-plus-dataset recipe might transfer to other evidence-oriented tasks, such as deepfake video attribution or splice localization, where MLLMs currently lack a vocabulary for describing artifacts.","The subjective evaluation of explanations leaves open whether the explanations are causally tied to the detection decision or are plausible post-hoc narratives; a controlled comparison against a classifier trained without the explanation loss would separate the two.","If explanation quality degrades under adversarial perturbations that preserve the image content, that would suggest the explanations reflect surface statistics rather than genuine forensic evidence."],"forward_implications":["A single model can output both a verdict and a human-readable explanation, so forensic analysts can inspect the reasoning behind each decision instead of trusting an opaque score.","Because the method builds on MLLMs, it inherits broad visual knowledge from pretrained models, potentially transferring to unseen forgery types without retraining the entire vision backbone.","The ForgReason dataset, built cheaply with an LLM agent and refined by limited human annotations, could lower the cost of producing explainable detectors for other forensic tasks.","If the reported generalization holds, explanation-guided training acts as a regularizer that improves detection beyond what a standard classifier would achieve on the same benchmarks."],"supporting_citations":[],"fun_headline_variants":["MLLMs get a forensic eye for fake images","Explainable AI forgery detection via forensics","ForenX: AI that explains how it spots fakes","Forensic prompts make MLLM detectors explainable","Seeing like an examiner: MLLMs explain fake images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the forensic prompt and the ForgReason annotations direct the model to visual evidence that genuinely indicates forgery, rather than to artifacts that merely correlate with the two benchmarks used for evaluation.","fun_headline_variants_meta":{"raw":{"variants":["MLLMs get a forensic eye for fake images","Explainable AI forgery detection via forensics","ForenX: AI that explains how it spots fakes","Forensic prompts make MLLM detectors explainable","Seeing like an examiner: MLLMs explain fake images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00057,"raw_usage":{"total_tokens":2659,"prompt_tokens":870,"completion_tokens":1789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1711}},"tokens_in":486,"tokens_out":1789,"duration_ms":14650,"temperature":1.0,"reasoning_tokens":1711,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:36:23.063242+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test would be to hold out an entire family of generative models not represented in the training data, run ForenX on those images, and have independent forensic examiners judge whether the model's explanations point to the true generative artifacts (for example, the specific texture or noise patterns of that generator). If explanation accuracy collapses on such out-of-distribution images while detection accuracy stays high, the explanations are not tracking independent forensic evidence.","supporting_citations":[],"review_version":1}