acceptodds
Under review as a conference paper at ICLR 2027

Evaluating the Evaluators: Diagnosing Large Multimodal Models for AI-Generated Image Assessment

Abstract

With the rapid advancement of text-to-image (T2I) generation, robust evaluation becomes critical yet challenging, as traditional metrics fail to capture fine-grained alignment and generative artifacts. While large multi-modal models (LMMs) are increasingly adopted as evaluators, existing benchmarks typically study semantic understanding, quality perception, and authenticity identification in isolation, while largely neglecting responsibility detection. This leaves a gap in unified and comprehensive validation. To bridge this gap, we introduce **SQUARE-Bench**, a comprehensive benchmark that systematically evaluates LMM capabilities as evaluators of AI-generated images across four aspects, including ***S**emantics*, ***Qu**ality*, ***A**uthenticity*, and ***Re**sponsibility*. SQUARE-Bench introduces a granular taxonomy of 38 sub-dimensions to evaluate nearly 10K AIGIs sampled from 22 diverse models, ranging from legacy to state-of-the-art generators, complemented by over 3K real-world images. The images are annotated with curated question-answering pairs. Extensive experiments on 23 LMMs reveal that top proprietary models (*e.g.*, Gemini-3-Pro) already outperform the individual human expert baseline. However, the performance gap between models remains significant, exhibiting notable disparities in fine-grained inference and domain-specific robustness. Beyond benchmarking, we conduct a proof-of-concept study of LMM-guided iterative editing, in which dimension-specific LMMs provide diagnostic feedback to fixed image editors. The resulting guided system yields selective improvements in semantics, authenticity, and responsibility, while exhibiting a consistent visual-quality trade-off. SQUARE-Bench can serve as both a diagnostic tool for characterizing LMM evaluator capabilities and studying their use in T2I generation refinement. The benchmark and dataset will be released upon publication.

Primary area:
datasets and benchmarks
Keywords:
Large Multimodal Models, AI-Generated Images, Multimodal Evaluation.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_V9v3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

SQUARE-Bench evaluates LMMs as judges of AI-generated images. It organizes approximately 18K question-answer pairs into four aspects and 38 subcategories using generated and real images. Dual-answer annotations distinguish observed image content from the intended generation outcome. The paper evaluates 23 models, compares them with five independent human experts, and fine-tunes three models on the benchmark. An auxiliary experiment uses adapted models to guide iterative image editing. This produces mixed gains and lower scores on all six quality metrics for both editors relative to single-pass editing.

Critical strengths
  1. A useful evaluation objective and broad resource scope. Assessing the reliability of multimodal evaluators is important. Bringing semantic, quality, authenticity, and responsibility tasks into a shared evaluation framework, with diverse image sources and a broad model comparison, could support more informed evaluator selection.
  2. Explicit separation of depicted content and generation targets. The dual-answer annotations provide a useful way to distinguish whether the generator failed to realize the target or the evaluator misread the resulting image. The paper also reports substantial variation across evaluation aspects and acknowledges negative downstream editing results, providing a more informative picture than an aggregate leaderboard alone.
Critical weaknesses
  1. The construction process does not sufficiently establish what the benchmark measures. To support conclusions about evaluator reliability, the taxonomy and QA strategy need to capture consequential errors in actual image assessment. The paper defines 38 subcategories but provides limited validation of this coverage. Diagnosis within these categories also requires reliable labels. Expert cross-checking and adjudication are described, yet agreement rates and the nature of disputed judgments are unreported. This matters particularly when independent human experts score below several models: without examining those disagreements and human errors, it is unclear how much the gap reflects evaluator capability versus differences in interpretation or judgment standards.
  2. Incomplete test composition and numerical inconsistencies undermine score reliability. Test counts for most of the 38 subcategories and the aggregation rule are missing, preventing assessment of balance and reliability, but the disclosed 6-question and 12-question subsets already yield unstable category estimates. Several reported values also raise concerns about numerical consistency. For example, Table 5 reports 81.82% Human Bias accuracy for fine-tuned Ovis2.5-9B, while Appendix F (p.24, L1270–1278) specifies 12 test questions, with invalid responses counted as incorrect. No integer correct count on 12 questions can produce this percentage. Such discrepancies require clarification of the scoring procedure across models and categories.
Questions for authors
  1. What evidence shows that the taxonomy and QA strategy capture important errors in actual T2I evaluation? Please report annotation agreement, representative disputed label, and human-error cases to clarify the gap between human and LMM judges.
  2. Please provide category counts, per-model evaluation denominators, and the aggregation formula. How do the authors explain the numerical inconsistencies, exemplified by 81.82% under the stated 12-question protocol? Please check the tables for similar discrepancies and explain whether the same questions and scoring rules were used across models.
2 comments
AuthorsOpenReview ↗

Comment Part1

We thank the reviewer for the constructive questions about what SQUARE-Bench measures, how disagreements were resolved, and how the reported scores were computed. We agree that the submission should have made the evidence for these claims more explicit.

Response to W1a & Q1: Relevance of the taxonomy and QA design

We appreciate the reviewer’s question about how the taxonomy relates to actual T2I evaluation. Table 7 documents the sources used to curate the benchmark, including established benchmarks for image generation, synthetic-image detection, safety, and image editing. We use prompts and captions from these sources to bring established evaluation challenges into SQUARE-Bench, and we generate new outputs using 22 T2I models. Where source category metadata is available, it informs the initial categorization. The final question targets a property or failure visible in the resulting image.

To check whether these questions capture recurring failures that matter to the evaluation, we examined 796 questions from the test set with both ground-truth labels, traceable to the original generation batch, after screening for prompt cues. Visual GT and Intended GT differed on 516 questions across all 22 generators. In each of the 35 subcategories represented in this analysis, such discrepancies occurred in outputs from at least two generators. Attribute binding illustrates why this matters. One prompt requests a red giraffe, but the generated image depicts a naturally brown-and-orange giraffe. The image can look plausible and still fail an explicit requirement. The QA item asks about the giraffe’s visible color, so the benchmark can record the generation failure and separately assess whether the evaluator recognized it. Among the attribute-binding questions, Visual GT and Intended GT differed on 19 items, whose images came from 12 different T2I generators. We will add this analysis to show how the taxonomy groups recurring failures and how the questions test the specific visual evidence needed to judge them.

Response to W1b & Q1: Reviewer agreement, disputed labels, and the human–LMM gap

We appreciate the reviewer’s question about how much of the human–LMM gap reflects evaluator capability and how much may reflect differences in interpretation. We first quantified agreement during QA review. Three reviewers independently assessed each proposed question and answer and submitted their decisions before discussion or adjudication. This analysis covers all question, excluding the automatically generated binary real/fake questions. Of all review decisions, 83.09% approved the item, 12.24% flagged an incorrect answer, and 4.67% flagged a problem with the question. All three reviewers gave the same decision on 78.76% of items. Pairwise agreement was 85.37%, compared with 70.76% expected by chance under the observed vote distribution, yielding a Fleiss’ κ of approximately 0.50. To examine the score gap, we compared Qwen3-VL-32B with the Human (Best) reference on the same 2,875 test questions. We grouped the non-binary items by the three pre-adjudication review decisions that could be matched to the test version.

Question groupQuestionsOnly Qwen3-VL-32B correctOnly Human (Best) correctNet difference (Qwen − Human)
Automatically generated binary real/fake questions93417399+74
All three reviewers approved the test version1,200165159+6
Three reviews of the test version, with disagreement before adjudication4209140+51
All three reviewers flagged the same type of issue in the test version before adjudication100239+14
No complete set of three reviews could be matched to the test version2212812+16
Total2,875480319+161

The provenance-based real/fake questions account for 74 of the 161 additional correct answers. On the 1,200 non-binary questions unanimously approved by all three reviewers, the gap is six answers, or 0.50 percentage points. Larger differences appear in groups where reviewers disagreed or flagged problems. These groups may also differ in difficulty and task composition, so we will report the breakdown alongside the aggregate score and qualify our interpretation of the overall gap. The individual cases illustrate why this distinction matters. In one binary item, the human reference labeled a generated image as a real photograph, while Qwen correctly identified its synthetic provenance. In a bicycle-wheel item, the initial assessment and check disagreed about whether the hub and axle were visible. Human adjudication added an option stating that the components could not be distinguished. The human reference chose that option, while Qwen chose the option stating that the components were missing or merged.

AuthorsOpenReview ↗

Comment Part2

Response to W2 & Q2: Test composition and scoring

The test set contains 2,875 questions, with 412 in Semantics, 368 in Quality, 1,792 in Authenticity, and 303 in Responsibility. In the revised full-test comparison, accuracy within each aspect will be the number of correct answers divided by the number of scheduled questions in that aspect. Overall accuracy will be the total number of correct answers divided by 2,875. Thus, each question contributes equally to the overall score.

Thank you for catching the inconsistency in the fine-tuned Ovis2.5-9B Human Bias score. One of the 12 questions triggered a CUDA out-of-memory error before a prediction was saved. The original aggregation divided the nine correct answers by the 11 saved predictions, yielding 81.82%. We reran the missing question, and the model answered it incorrectly. We will correct Table 5 to 75.00% (9/12) and clarify that empty, unparseable, or missing predictions receive zero credit under the fixed-denominator protocol.

The submitted Responsibility scores for LLaVA-NeXT and CogAgent-18B excluded the two multi-image categories they could not process, Cultural Fairness and Human Bias. We agree that this coverage difference should have been explicit. In the revision, we will identify unsupported questions and report both a comparison on the same supported questions for every model and a full-test comparison with zero credit for unsupported items. We will state the denominators used and apply the same scoring rules when updating the affected tables and rankings.

Reviewer_RYPkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose a new benchmark SQUARE-Bench to evaluating large multi-modal models as evaluators for AI-generated images, covering four aspects including Semantics, Quality, Authenticity, and Responsibility. The SQUARE-Bench introduce a introduces a granular taxonomy of 38 sub-dimensions from the four aspects. A hybrid data consisting of Nearly 10,000 AIGIs curated from 22 diverse Models are constructed. In addition, a proof-of-concept study of LMM-guided iterative editing is designed and conducted to provide further diagnostic feedback to fixed image editors.

Critical strengths
  1. A systematic and comprehensive evaluation benchmark is proposed. Especially, the Responsibility aspect has been incorporated, which has not been extensively studied and investigated.
  2. Diverse emulation tasks have been devised and evaluated, ranging from semantic question answering with multiple formats and dual-answer view design.
  3. A proof-of-concept study of LMM-guided iterative editing is designed and conducted to provide further diagnostic feedback
Critical weaknesses
  1. The novelty of the propose evaluation should be further highlighted. The Semantics, Quality, Authenticity are extensively investigated and evaluated in related Benchmarks. What characters and specific features do the proposed benchmark involve? What specific features can this benchmark provide, in comparison with similar benchmark?
  2. The curated dataset is not large enough. The authors should provide some curated data samples even if the whole dataset cannot be released, showing the quality and importance of the curated data.
  3. Although this paper provides evaluations results on diverse LLMs with human evaluation scores, further comparisons with similar benchmarks are not further included.
Questions for authors

refer to the weakness

2 comments
AuthorsOpenReview ↗

Comment Part1

We thank the reviewer for the constructive questions about what SQUARE-Bench adds to existing benchmarks, how readers can inspect the curated data, and how our results compare with those of a related benchmark.

Response to W1 & W3: Diagnostic contribution and comparison with A-Bench

We thank the reviewer for asking what SQUARE-Bench adds to existing benchmarks. Our contribution is to define more precisely what it means for an LMM to be a capable evaluator of AI-generated images. We start from the requirements for text-to-image generation. An image should depict the requested content where appropriate, meet visual quality standards, appear physically plausible when realism is intended, and satisfy safety and fairness criteria. We translate these requirements and the failures that can arise during generation into 38 fine-grained sub-dimensions, each testing whether an LMM can judge if a specific criterion is met. The same annotation team applies shared guidelines to items requiring human judgment across all four aspects. We choose question formats that fit the judgment, such as multi-image questions for evaluating patterns of representation. Where dual labels are available, they let us assess whether an LMM correctly describes the image even when the generator misses the intended outcome.

FakeBench, LOKI, and FakeClue focus on detecting synthetic content and explaining forgery cues. As summarized in Table 1, they organize parts of their evaluations around question types, modalities, or image categories. In SQUARE-Bench, binary provenance classification is one scored task. The other Authenticity tasks ask whether the visible content satisfies specific realism criteria that matter for generation. For example, we test whether an LMM can judge if a prism disperses light in a physically plausible way. That judgment tells us whether the LMM can identify a particular generation failure. FakeBench · LOKI · FakeClue

Q-Bench+ evaluates general low-level visual perception, description, and quality assessment. About two-thirds of the source images for its perception and description tasks are in-the-wild photographs. SQUARE-Bench defines its Quality tasks around what an evaluator needs to check in generated images, including technical defects, aesthetic issues, and generation-specific failures such as malformed anatomy. Q-Bench+

A-Bench shares our focus on LMMs as evaluators of generated images and directly covers Semantics and Quality. SQUARE-Bench extends that evaluation to Authenticity and Responsibility within the same framework, including multi-image fairness judgments. This overlap makes A-Bench the most direct comparison for testing what SQUARE-Bench adds within the shared scope and across all four aspects. A-Bench

We agree that the submission lacked a direct empirical comparison. We evaluated the same eight base models on both benchmarks: Qwen3-VL-32B, Ovis2.5-9B, MiniCPM-V4.5, InternVL3-14B, LLaVA-OneVision-1.5-8B, Qwen3-VL-8B, InternVL3.5-14B, and InternVL3.5-8B. We also report results with the three SQUARE-Bench fine-tuned variants of Ovis2.5-9B, MiniCPM-V4.5, and Qwen3-VL-8B. We compare A-Bench Overall with both the closest task scope, SQUARE-Bench Semantics + Quality, and the full four-aspect SQUARE-Bench test set.

Model setA-Bench Overall compared withSpearman ρKendall τ
8 base modelsSQUARE-Bench Semantics + Quality0.7860.643
8 base modelsSQUARE-Bench Overall0.7620.643
All 11 modelsSQUARE-Bench Semantics + Quality0.8910.745
All 11 modelsSQUARE-Bench Overall0.8730.745

Among the eight base models, the accuracy range was 5.69 percentage points on A-Bench and 15.55 percentage points on the full SQUARE-Bench test set. Across all 11 models, 52.55% of A-Bench questions and 25.60% of SQUARE-Bench questions were answered correctly by every model. They show that SQUARE-Bench provides additional model-separation and ranking information under a unified protocol while extending the evaluated scope to Authenticity and Responsibility. The submitted results also illustrate why the choice of evaluator depends on the task. Qwen3-VL-32B scores higher than Gemini-3-Pro-Preview on Quality (75.27% vs. 60.27%, Table 3), while Gemini scores higher on Authenticity (84.40% vs. 77.40%, Table 4). We will add this controlled cross-benchmark comparison to the paper, explicitly name the evaluated models, and recalibrate the contribution statement to emphasize unified and controlled evaluator profiling rather than the novelty of each individual component task.

AuthorsOpenReview ↗

Comment Part2

Response to W2: Curated examples

Thank you for asking to see more examples. We agree that the seven worked examples in Figure 4 gave readers a limited view of the benchmark. We have expanded the preview to 38 examples, one for each sub-dimension, so readers can inspect the images, questions, answer choices, and annotations across all four aspects. The expanded preview is available in the anonymous repository.

Reviewer_6RKsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces SQUARE-Bench for assessing how well large multimodal models evaluate AI-generated images. It covers semantics, quality, authenticity, and responsibility through 38 sub-dimensions, using nearly 10,000 generated images and over 3,000 real images. Its dual-answer design records both what is visible in an image and what the intended generation should have produced. The authors evaluate 23 LMMs against five independent human experts, examine supervised fine-tuning, and use adapted models to guide iterative image editing. The results show substantial differences across evaluation aspects. The editing study reports selective gains in semantics, authenticity, and responsibility, alongside consistent losses in visual quality relative to single-pass editing.

Critical strengths

The benchmark brings several relevant evaluation tasks into one setting and puts substantial effort into annotation. The inclusion of responsibility and multi-image fairness questions broadens the coverage beyond semantic alignment and image quality. For the manually authored QA instances, a primary annotator constructs the question and answers, and at least three additional experts check them before adjudication (Section 3.3 and Appendix D). The distinction between visual content and the intended output is also useful: it allows a model to recognize what an image actually depicts even when generation has failed.

The comparison across 23 LMMs provides useful evidence that evaluator choice should depend on the task. For example, Table 3 reports substantially better quality assessment for Qwen3-VL-32B than for Gemini-3-Pro, while Table 4 shows the reverse ordering for authenticity. These differences are more informative than a single aggregate score. The editing study also reports its negative results clearly: both guided editors lose quality across all six quality metrics. This is a useful caution about assuming that diagnostic feedback will automatically improve the final image.

Critical weaknesses

The scoring protocol needs to be reconciled before some of the ranking claims are reliable. Appendix F states that Human Bias contains 12 test questions and that invalid answers count as incorrect, yet Table 5 reports 81.82% for fine-tuned Ovis2.5-9B. That percentage cannot come from an integer number of correct answers out of 12, so the denominator or aggregation rule needs clarification. There is also a coverage problem: LLaVA-NeXT’s 92.98% responsibility score excludes the two social-fairness subsets it cannot process, but Section 4.3 uses this score to describe the model as fourth in responsibility alongside models evaluated on those subsets. A supported-task score can be useful, but it should not be treated as a directly comparable full-benchmark score. These may be correctable reporting or implementation issues, but they matter for a paper whose main purpose is to compare evaluators.

The dual-answer analyses do not yet separate useful diagnostic findings from effects of the question format and data selection. Section 4.3 reports that 88.64% of incorrect predictions on mismatch cases select the Intended GT. On a two-option question where the two ground truths differ, however, every incorrect valid answer must equal the Intended GT. The paper does not report the option-count distribution within the mismatch subset or a corresponding baseline. This makes the aggregate statistic difficult to interpret as a preference for intended content. The Mixed-Generator Average raises a related interpretation issue: generative-distortion examples are explicitly curated to contain flaws (Appendix C.2), so the 26.72% success rate describes a selected diagnostic corpus, not a representative estimate of generator reliability. This limits the broader conclusion about a severe generative bottleneck. The dual annotations remain useful, but these conclusions need tighter definitions and controls.

Questions for authors

Could you reconcile the 81.82% Human Bias result with the stated 12-question protocol and provide evaluated, valid, and correct counts for each model and subset? Please also report rankings on a common supported subset and on the full benchmark, with unsupported tasks identified explicitly. How are the four aspects weighted in the overall score?

Does the tendency to select Intended GT remain on mismatch questions with at least three options? Please report the mismatch subset’s option-count and aspect distributions, and compare the result with a baseline that accounts for the available incorrect options and answer frequencies.

How exactly is the Mixed-Generator Average computed? Please specify the eligible examples and scoring rule for each format, especially the binary authenticity score of 11.76% in Table 4. Binary real/fake labels are based on provenance, whereas the generator analysis concerns photorealism; a realistic generated image is still synthetic. How is this distinction handled, and how should the aggregate score be interpreted given the deliberate curation of difficult and defective images?

2 comments
AuthorsOpenReview ↗

Comment Part1

We thank the reviewer for the detailed and constructive comments. Your questions helped us identify where the scoring protocol needed clarification and where our interpretation of the results went beyond the evidence.

Response to W1 & Q1: Scoring, coverage, and overall weighting

Thank you for catching the inconsistency in the Human Bias score. Appendix F described how we scored empty or unparseable responses when a prediction record existed. Our aggregation omitted cases where inference failed before saving a record. We reviewed both available run logs for this checkpoint and found a single CUDA out-of-memory error on a shared GPU, which occurred on the ten-image Human Bias question. Of the 11 saved predictions, all were valid and nine were correct, which explains the reported 81.82% (9/11). We reran the missing question, and the model answered it incorrectly. We will correct the Table 5 entry to 75.00% (9/12) and clarify in Appendix F that every scheduled question contributes to the denominator, with empty, unparseable, or missing predictions receiving zero credit.

The reviewer also raises an important point about coverage. LLaVA-NeXT’s 92.98% reflects the Responsibility questions it supports, and the original fourth-place rank was computed among zero-shot models using the scores in Table 5. On the same supported questions for all zero-shot models, LLaVA-NeXT ties for seventh. On the full Responsibility set, where unsupported questions receive zero credit, it scores 87.46% and ranks 18th. We will identify unsupported tasks explicitly.

Under the revised fixed-denominator protocol, each question on the full filtered test set contributes equally to overall accuracy. This gives weights of 14.33% for Semantics, 12.80% for Quality, 62.33% for Authenticity, and 10.54% for Responsibility. We will state this aggregation rule and update the affected scores and rankings in the tables and Figure 5.

Response to W2a & Q2: Dual-answer results and answer-position bias

We appreciate the reviewer’s observation that, on a two-option mismatch question, any valid wrong answer necessarily selects Intended GT. The submitted 88.64% was a macro-average across the 23 zero-shot models that combined questions with different numbers of options. The table shows the 626 questions in that set whose Visual GT and Intended GT differ.

Aspect2 options3 options4 options
Semantics381084
Quality1258
Authenticity2576830
Responsibility61152
Total35795174

To examine the effect of question format, we restricted the analysis to questions with three or four options. Across the 23 models, 1,350 of 1,715 valid wrong responses (78.72%) selected Intended GT. The expected rate under uniform selection among the available wrong options is 41.06%. We also estimated a baseline that accounts for answer-position preferences. For each model and question, we estimated wrong-answer letter frequencies from the other mismatch questions with the same aspect and number of options, then normalized those frequencies over the current question’s wrong options. This gives 77.16%, close to the observed rate. We will report both baselines and describe the result as an error pattern. Whether models favor the intended content itself remains open.

AuthorsOpenReview ↗

Comment Part2

Response to W2b & Q3: Mixed-Generator Average and its scope

Thank you for raising the distinction between visual realism and provenance, and for asking which examples contribute to the generator score. The intended outcome depends on the aspect being assessed. For Semantics, the image should match the content specified in the prompt. For Quality, it should satisfy the visual-quality criterion addressed by the question. For Authenticity, it should satisfy the relevant criterion for visual realism or physical plausibility. For Responsibility, it should satisfy the applicable safety or fairness criterion, including when the prompt requests harmful or biased content.

These criteria define the two labels for each eligible image–question pair. Visual GT describes what the image actually shows, and Intended GT describes the ideal outcome under the relevant criterion. Figure 4 illustrates a semantic failure. The prompt requests four vases, so the Intended GT is four. The generated image shows nine, so the Visual GT is nine.

The Mixed-Generator Average uses only images that we generated from prompts we sampled. The pre-existing images sampled for the Quality benchmark are outside this pool. We compute the fraction for which the parsed Visual GT and Intended GT match. Each eligible question contributes once. We agree that the manuscript should have explained this eligibility and scoring rule. We will clarify both and describe 26.72% as target agreement on this curated set of generation tasks. We will also narrow the broader claim about generator reliability.

A generated image can look photorealistic and still have synthetic provenance. The binary real/fake task uses provenance labels for LMM evaluation, and the generator score uses fine-grained authenticity questions to assess visual plausibility. When transferring the aggregated scores into Table 4, we misaligned the Mixed-Generator Average row with the column headers. As a result, the 11.76% Geometric Scale score appeared under Binary. We will realign the affected entries and mark Binary as not applicable in the generator row.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.