{"id":"9820a203-ef28-457b-b320-902b7faf8963","arxiv_id":"2411.17261","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HEIE is an MLLM-based evaluator that predicts defect heatmaps, plausibility scores, and natural-language explanations for AI-generated images, along with a new explainability dataset.","lead":"This paper presents HEIE, a multimodal language model system that finds implausible regions in AI-generated images, scores them, and writes explanations for each defect. It also introduces a dataset of explainable defect annotations built by asking LLMs to analyze images with circled problem areas.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Explanation-quality results may measure imitation of LLM-generated labels rather than accurate explanation of image defects, leaving the paper's central 'explainable' claim under-supported.","rationale":"The reader's weakest-assumption analysis identifies the same core issue: the explanation branch is trained and evaluated against LLM-generated annotations whose provenance and quality are not independently validated. I agree with the conditional verdict. The heatmap and score contributions are internally consistent and are supported by the RichHF-18K and AbHuman experiments, including ablations in Tables 5 and 7, so a rejection would be too strong. However, the explainability claim is the paper's distinctive selling point, and it currently depends on circular evidence. The proposed human-grounded test would settle whether the explanation results are substantive or an artifact of label generation. Since the reader already conditioned the verdict on this class of concern, no change to the verdict is needed; if anything, the condition should explicitly require human-grounded validation of explanations before the explainability claim is accepted. Additional weaknesses, such as missing error bars, an incompletely specified loss, and unreleased code, support keeping the verdict conditional rather than accepting it outright, but they are secondary to the circularity of the explanation benchmark.","tokens_in":15177,"tokens_out":5589,"duration_ms":57294,"concrete_test":"Run a blind, claim-level human evaluation on 200 held-out Expl-AIGI-Eval images with a pre-registered protocol and reported inter-annotator agreement. For each image, independent annotators mark every defect claim in a generated explanation as (a) present and correctly localized, (b) plausible but not visible in the image, or (c) hallucinated or wrong. Compare HEIE, GPT-4o, and Claude-3.5-Sonnet on precision/recall of defect claims and on blind preference. If HEIE does not significantly outperform the two LLMs that generated the gold labels, Table 2's explanation superiority is better explained by training on LLM-generated text than by more accurate explanation of real image defects.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline novelty is explainable implausibility evaluation (title, abstract, CoT-Driven Explainable Trinity Evaluator). That claim rests almost entirely on Table 2. The gold explanations in Expl-AIGI-Eval were produced by Claude-3.5-sonnet free-form and reformatted by GPT-4o (Sec. 3.4, Fig. 3), and HEIE is fine-tuned on those exact texts. The first two metrics in Table 2 are perplexity against those LLM-generated labels and GPT-4o similarity to those labels; both therefore measure distribution matching to the annotation pipeline, not fidelity to actual image defects. The comparison is especially fragile because GPT-4o and Claude-3.5-Sonnet are the same models that generated and reformatted the gold text, so HEIE's large margin over them in Table 2 is exactly what one would expect from training on their output style. A human evaluation is mentioned but its size, rubric, and inter-annotator agreement are not reported, so it cannot independently break the circularity. The heatmap and score results in Tables 1, 3, and 4 are independent of this issue and give real support to the system; the specific claim of explanation superiority, however, is not secured by the experiments as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HEIE, an MLLM-based evaluator that jointly outputs implausibility heatmaps, a verisimilitude score, and free-form textual explanations for AIGC images. The method introduces an Adaptive Hierarchical Implausibility Mapper, which uses two-way cross-attention between [MAP] token features and image features and fuses local and global heatmaps via a Laplace-uncertainty weighted sum, plus a CoT-Driven Explainable Trinity Evaluator that sequences image description, problem-region identification, map-token injection, issue analysis, and score-token regression. The authors also construct Expl-AIGI-Eval by adding LLM-generated explanations to RichHF-18K and AbHuman. Experiments report state-of-the-art heatmap metrics on RichHF-18K (MSE 0.00825, KLD 1.634, CC 0.574, SIM 0.417, AUC-Judd 0.915) and AbHuman, improved score correlations, zero-shot cross-dataset generalization, and ablations for each proposed component.","tokens_in":15503,"tokens_out":5420,"duration_ms":47566,"significance":"If the results hold, HEIE is a meaningful advance: it is an MLLM-based system that produces pixel-level defect heatmaps together with textual rationales, and it outperforms specialized baselines on the heatmap and score benchmarks. The heatmap and score evaluations are grounded in external human-annotated masks from RichHF-18K and AbHuman, and the component ablations in Tables 5, 6, and 7 are internally consistent and support the main design choices. The principal caveat is that the explanation-quality evaluation is not independent of the label-generation process, so the headline 'explainable' claim is only partially secured by the current experiments.","major_comments":[{"comment":"The explanation branch is trained on and evaluated against texts produced by Claude-3.5-sonnet and reformatted by GPT-4o, and the two automated metrics in Table 2 are perplexity against those same texts and GPT-4o similarity to those same texts. Because HEIE is fine-tuned on exactly this annotation pipeline, the large margins in Table 2 likely measure how well HEIE imitates the LLM annotation style rather than how accurately it explains the actual image defects; the comparison against GPT-4o and Claude-3.5-Sonnet is particularly fragile because those are the models that generated the gold text. The human evaluation cited in Table 2 is reported without sample size, annotation protocol, rubric, or inter-annotator agreement, so it does not break the circularity. To support the central explainability claim, the authors should evaluate explanation factuality against the human-annotated masks (for example, whether each mentioned region overlaps a mask), use held-out LLM labelers that were not involved in training-data generation, or provide a complete human study with agreement statistics.","section":"§3.4, Table 2, Fig. 3"},{"comment":"The loss used to train the Implausibility Mapper is not precisely specified. In §3.2.1 the text says 'we simply adopt focal loss to overcome the data imbalance issue,' but Eq. (3) is a Laplace negative-log-likelihood with an additional log(σ) term, and no equation or text states how focal loss is combined with the uncertainty loss, nor how σ_l and σ_g are parameterized and optimized. The hyperparameters of the focal loss (α, γ) and the relative weighting between the heatmap loss and the uncertainty terms are also omitted. Because Table 5 attributes the final gain (Exp. e vs. d) to the uncertainty-based adaptive fusion, the reader cannot determine whether the reported improvements come from the fusion rule or from unstated training details; please provide the full combined loss and all hyperparameters.","section":"§3.2.1, Eq. (3)"},{"comment":"All quantitative claims are reported as point estimates without error bars, multiple seeds, or significance tests. For example, the headline improvements over RAHF (augmented prompt) in Table 1 are small on several metrics (AUC-Judd 0.915 vs 0.913; PLCC 0.697 vs 0.693; SRCC 0.683 vs 0.681), and the zero-shot gains in Table 4 are smaller still; without variance estimates these differences cannot be distinguished from training noise. Please report mean±std over at least three runs or a paired significance test for the main comparisons.","section":"Tables 1–7"},{"comment":"The verisimilitude scorer is not fully specified: the calibration function S = Calib(S_token, S_map) is introduced without giving the concrete form used in the experiments, and the text states only that 'weighted summation and dynamic fusion' exhibited similar performance. Since Table 6 reports that combining token and heatmap features (Exp. d) is a key component of the score result, the exact calibration function and any learned parameters must be reported to make the method reproducible and to interpret the ablation.","section":"§3.3.1, Table 6"}],"minor_comments":[{"comment":"There are typographical errors: 'possessing extensive knowledge and common' in §1 should end with 'sense,' and 'Co-driven' in §3.1 should be 'CoT-driven.'","section":"§1, §3.1"},{"comment":"Baseline names are inconsistent across tables (InternViT, InternViT-300MB, InternViT-300M), and the text does not state whether these refer to the same model variant; please unify the notation.","section":"Tables 3 and 4"},{"comment":"The RAHF numbers are taken from the original paper; the manuscript should state explicitly that the evaluation split, preprocessing, and metrics are identical, since RAHF is not available for re-evaluation.","section":"§4.2, Table 1"},{"comment":"The Expl-AIGI-Eval dataset description lacks basic statistics, such as number of images, number of explanations, average explanation length, and number of regions per image, as well as quality-control details beyond the ethical filtering mentioned in the text; these should be reported for the dataset contribution to be assessable.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"The heatmap and score contributions are solid and well aligned with the journal's scope; the main risk is the explanation benchmark's circularity. If the authors add an independent evaluation of explanation factuality or a properly reported human study, I would support publication. The manuscript would also benefit from full loss and calibration specifications and from variance estimates for the quantitative claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"HEIE is a competent systems paper whose central heatmap/score contribution mostly holds up; the explainability claim is the soft spot, and the stress-test note is right about why. The MLLM with MAP/SCORE tokens plus two-way cross-attention and uncertainty-based hierarchical fusion is genuinely new relative to RichHF, HumanRefiner, Q-Align, and VisualCritic. The ablations in Tables 5-7 are internally consistent and show each component matters. Table 1 and Table 3 show solid gains over specialized baselines, and the zero-shot transfer in Table 4 is a nice plus. The Expl-AIGI-Eval dataset, even with LLM-generated texts, is a useful resource if released.\n\nThe worry in the stress-test is real but confined to the explanation branch. Table 2 compares fine-tuned HEIE against untuned general MLLMs using perplexity on Claude/GPT-4o-generated gold texts and GPT-4o similarity to those same texts. That measures style imitation, not defect explanation. The human eval is the only independent evidence and its size, rubric, and agreement are not reported, so it cannot carry the claim. The paper should be revised to compare against fine-tuned MLLM baselines, use human-written or independently annotated explanations, and report the human eval protocol.\n\nMinor but worth fixing: the loss description is inconsistent (Sec 3.2.1 says focal loss, Eq. 3 gives a Laplace uncertainty loss - presumably one is for the heatmap head and one for fusion weighting, but it should be stated precisely). No error bars or significance tests appear anywhere. No code or data link despite the project page, so the reader cannot independently check the SOTA numbers. Also, 'first to use MLLM for heatmap prediction' is a strong phrasing; the novelty is in the specific mechanism, not the broad category.\n\nNone of this undermines the heatmap/score results, which are independent of the explanation circularity. The paper deserves a serious referee. I would send it out, but with a clear request for major revision focused on the explanation evaluation and on releasing artifacts.","headline":"Solid heatmap/score system paper; explanation claim needs independent evaluation before it can be believed.","tokens_in":16024,"tokens_out":2465,"would_cite":true,"duration_ms":23893,"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":"The paper claims that an MLLM fine-tuned with chain-of-thought prompting can produce defect heatmaps, plausibility scores, and written explanations of AI-generated image flaws, and that these three outputs improve one another enough to…","keywords":["AIGC image quality assessment","defect heatmap prediction","explainable image evaluation","multimodal large language models","chain-of-thought reasoning","hierarchical implausibility mapping","uncertainty-based fusion"],"falsifier":"Construct a test set of AIGC images with defects that are independently verified by multiple humans and whose causes are known, then check whether HEIE's explanation names the verified cause for each defect while ignoring non-defective areas; if explanations match plausible-sounding generic text as often as verified causes, the explainability claim fails. A simpler version: if human raters cannot distinguish HEIE's explanations from fluent template text, the explanation evaluation is rewarding style, not content.","tokens_in":14977,"feed_emoji":"🔍","tokens_out":6953,"duration_ms":60491,"temperature":0.7,"pith_summary":"HEIE is a single multimodal large language model that is fine-tuned to grade AI-generated images in three coordinated ways: a pixel-level heatmap of implausible regions, an overall plausibility score, and a written explanation that names the problem areas and why they fail. The paper argues that specialized defect-localization models cannot explain their findings or generalize well, while general-purpose MLLMs cannot localize tiny defects or emit pixel-level maps. Its central claim is that a chain-of-thought structure that makes the model describe the image, locate problematic regions, analyze their causes, and then score them lets all three outputs reinforce one another, producing state-of-the-art heatmap and score predictions while adding explanation that users can act on. If true, quality assessment of AI-generated images moves from a bare score or heatmap to a form that directly supports human understanding and targeted model repair.","feed_headline":"One model localizes, scores, and explains AI-image defects","feed_subtitle":"It localizes and explains AI-image flaws while beating specialized baselines on AIGC quality benchmarks.","key_machinery":"The load-bearing mechanism is a pair of special tokens inside the LLM sequence. A dynamic number of [MAP] tokens carries implausibility-region information: the hidden state of each token is extracted from the final LLM layer and fused with vision-encoder features through stacked two-way cross-attention to predict a heatmap, with more tokens assigned to higher-resolution images. A separate global [MAP] token yields a coarse whole-image heatmap, and an uncertainty-weighted fusion, using Laplace distribution standard deviations, combines the local and global maps into the final output. A [SCORE] token's hidden state is regressed through a small network and then calibrated against a score derived from the predicted heatmap, while the chain-of-thought prompt organizes the text generation so that description, localization, and analysis feed directly into the map and score tokens.","core_discovery":"The paper's central claim is that an MLLM can be adapted into a full explainable image implausibility evaluator rather than merely a textual critic. The CoT-Driven Explainable Trinity Evaluator decomposes the task into five ordered subtasks: image description, problematic-region identification, map-token injection, issue analysis, and score-token regression; the earlier textual steps are not decorative but feed the map and score tokens that produce the spatial and numeric outputs. The Adaptive Hierarchical Implausibility Mapper predicts a global heatmap and several local patch heatmaps, then fuses them with weights derived from Laplace-distribution uncertainty, so that small defects such as a distorted hand or eye do not get washed out by the global view. On the RichHF-18K benchmark the authors report the best heatmap metrics among the compared models, including an all-data MSE of 0.00825, and on their new Expl-AIGI-Eval dataset they report the best explanation scores by both GPT-4o-based comparison and human evaluation.","pith_inferences":["Editorial inference: the explanation metrics likely measure agreement with LLM-written reference text; they do not directly prove the explanations name the true physical cause of each defect.","Editorial inference: because the dataset's explanations are generated by a commercial LLM from red-circled regions, HEIE's explanation quality ceiling is that teacher LLM's ability to describe circled defects.","Editorial inference: the uncertainty-based fusion is a self-contained module that could be reused in non-MLLM heatmap predictors, for example segmenting localized artifacts in video frames or medical images.","Editorial inference: a sharper test would ask humans to localize the defect the explanation refers to or to judge whether the explanation's cause matches the heatmap region; the current human evaluation score does not separate fluency from accuracy."],"forward_implications":["If the central claim holds, one fine-tuned MLLM can replace the usual pipeline of separate heatmap and scoring models for AIGC quality control.","The chain-of-thought text becomes a working part of the architecture: forcing the model to describe and localize before scoring should make the spatial and scalar outputs more reliable, not just more readable.","Uncertainty-weighted local-global heatmap fusion should transfer to images of arbitrary resolution and aspect ratio without re-tuning the patch grid.","The explainable dataset makes it possible to train and benchmark other models on jointly predicting heatmaps, scores, and textual causes, so the contribution outlives this particular model.","Zero-shot transfer results between two defect datasets suggest that the world knowledge in the MLLM supplies generalization that small specialized models lack."],"supporting_citations":[{"why":"supplies the RichHF-18K benchmark, the task definition, the evaluation protocol, and the multi-head baseline that HEIE must beat.","marker":"[26]"},{"why":"supplies the AbHuman dataset used for additional finetuning and zero-shot cross-domain generalization tests.","marker":"[13]"},{"why":"provides the pretrained multimodal backbone that HEIE fine-tunes; all comparisons inherit its base capabilities.","marker":"[9]"},{"why":"writes the free-form defect analyses that become the ground-truth explanation labels in Expl-AIGI-Eval.","marker":"[3]"},{"why":"motivates the free-form output stage by showing that fixed format restrictions hurt LLM output quality.","marker":"[38]"},{"why":"supplies the chain-of-thought prompting idea that the five-step CoT evaluator builds on.","marker":"[44]"},{"why":"defines the KLD, SIM, CC, and AUC-Judd metrics used to score heatmap quality.","marker":"[6]"},{"why":"provides the Laplace-uncertainty formulation used to fuse local and global heatmaps adaptively.","marker":"[28]"}],"fun_headline_variants":["CoT-guided MLLM gives heatmaps, scores, and explanations for AI images","Heatmaps plus reasons: MLLM judges AI image quality","MLLM maps and explains AI image flaws with uncertainty","CoT-driven MLLM evaluates AI images with pixel-level reasons","Explainable AI image defects: localize, score, justify"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the human-annotated defect masks are correct and that the LLM-written explanations, produced by prompting a commercial model on red-circled defect regions, are a trustworthy stand-in for true explanations of those defects; if either fails, the reported heatmap or explanation numbers measure something other than real implausibility understanding.","fun_headline_variants_meta":{"raw":{"variants":["CoT-guided MLLM gives heatmaps, scores, and explanations for AI images","Heatmaps plus reasons: MLLM judges AI image quality","MLLM maps and explains AI image flaws with uncertainty","CoT-driven MLLM evaluates AI images with pixel-level reasons","Explainable AI image defects: localize, score, justify"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000806,"raw_usage":{"total_tokens":3583,"prompt_tokens":1032,"completion_tokens":2551,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":2460}},"tokens_in":648,"tokens_out":2551,"duration_ms":16880,"temperature":1.0,"reasoning_tokens":2460,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:19:24.046250+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a test set of AIGC images with defects that are independently verified by multiple humans and whose causes are known, then check whether HEIE's explanation names the verified cause for each defect while ignoring non-defective areas; if explanations match plausible-sounding generic text as often as verified causes, the explainability claim fails. A simpler version: if human raters cannot distinguish HEIE's explanations from fluent template text, the explanation evaluation is rewarding style, not content.","supporting_citations":[{"cited_title":"Rich human feedback for 9 text-to-image generation","cited_arxiv_id":null,"evidence_quote":"supplies the RichHF-18K benchmark, the task definition, the evaluation protocol, and the multi-head baseline that HEIE must beat."},{"cited_title":"Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks","cited_arxiv_id":null,"evidence_quote":"provides the pretrained multimodal backbone that HEIE fine-tunes; all comparisons inherit its base capabilities."},{"cited_title":"Claude 3.5 sonnet: A next-generation language model","cited_arxiv_id":null,"evidence_quote":"writes the free-form defect analyses that become the ground-truth explanation labels in Expl-AIGI-Eval."},{"cited_title":"Chain-of-thought prompting elicits reasoning in large lan- guage models.Advances in Neural Information Processing Systems, 35:24824–24837, 2022","cited_arxiv_id":null,"evidence_quote":"supplies the chain-of-thought prompting idea that the five-step CoT evaluator builds on."},{"cited_title":"What do different evaluation metrics tell us about saliency models?IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):740–757, 2018","cited_arxiv_id":null,"evidence_quote":"defines the KLD, SIM, CC, and AUC-Judd metrics used to score heatmap quality."},{"cited_title":"Geometry uncer- tainty projection network for monocular 3d object detection","cited_arxiv_id":null,"evidence_quote":"provides the Laplace-uncertainty formulation used to fuse local and global heatmaps adaptively."}],"review_version":1}