acceptodds
Under review as a conference paper at ICLR 2027

From Prompts to Question Graphs: Structure-Aware Evaluation and Selection of Text-to-Image Models

Abstract

Text-to-image evaluation usually compresses the satisfaction of multiple, interacting requirements into a single score for each prompt. This flat treatment creates three ambiguities: a missing entity can make all of its attributes and relations unanswerable, prompts containing more checks can receive disproportionate influence, and the same error may reflect either a basic generation failure or an inability to preserve a concept under composition. We introduce QC-T2I-Bench, a framework that represents each prompt as a graph of attributed atomic questions. Built on Davidsonian Scene Graphs, this representation explicitly connects entities, attributes, relations, and higher-level semantics through their prerequisite structure. We develop a dependency-aware aggregation rule that evaluates a question only when its prerequisites are satisfied and balances contributions across prompts of different complexity. The resulting graph-structured evidence supports analyses unavailable from conventional prompt scores: it separates initial concept realization from conditional preservation, measures the probability of jointly satisfying connected requirements, and provides capability profiles for model selection. Experiments across diverse text-to-image generators and bilingual English–Chinese prompts reveal a sharp compositional bottleneck: joint completion decreases from 80.7% for components involving two capabilities to 37.2% for those involving seven or more. Finally, without training a routing model, question-level profiles select generators under quality–cost constraints, matching ERNIE's 89.51-point estimate while reducing inference cost by 21.3% in GPU-s/MP. These results show that question graphs provide a common representation for reliable comparison, interpretable failure analysis, and efficient deployment of text-to-image models.

Primary area:
datasets and benchmarks
Keywords:
text-to-image generation, multimodal evaluation, compositional generation, question graphs, structured evaluation, Davidsonian scene graphs, model selection, cost-aware routing

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_RM1gOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses a useful question: whether question-level dependency graphs can support more informative text-to-image evaluation and model selection. However, several concerns limit the conclusions.

Critical strengths

The paper includes useful controls, particularly pre-masking outcomes for dependency analysis, prompt-cluster bootstrapping, and appropriately qualified routing claims.

Critical weaknesses
  • Model-dependent selection in scoring. Dependency masking evaluates each model’s downstream capabilities on a different surviving subset. This avoids repeated penalties but can favor models that satisfy prerequisites only on easier examples. Report prerequisite coverage and unconditional success alongside conditional scores. High rank correlation with an alternative score does not resolve this issue.
  • Insufficient validation in the supplied manuscript. Construction rules and automated auditing do not establish question atomicity, coverage, graph correctness, or evaluator accuracy. Independent validation is needed because errors in prerequisite judgments propagate into downstream eligibility and capability scores. Relevant evidence may be in the referenced supplement, which is not included in this PDF.
  • Cross-prompt matching does not isolate composition effects. Matching an entity across otherwise different prompts does not control scene difficulty or attached constraints. The reported root-success differences should be interpreted as associations unless supported by matched prompt variants and clearly specified aggregation.
  • The aggregation result has limited scope. Lower noise variance under independent, equal-variance errors does not establish that question weighting is preferable to prompt weighting. These rules weight the evaluation population differently, and correlated evaluator errors require additional analysis.
  • Routing comparisons are incomplete. Fixed generators and uniform random selection do not isolate the benefit of question-level profiles. Compare against a cost-matched prompt-independent policy and a simple prompt-level router.
Questions for authors

See weaknesses

3 comments

Thank you for the detailed critique and requested controls.

W1: Model-dependent selection

"Report prerequisite coverage and unconditional success alongside conditional scores."

Response. We now report coverage (eligible/all questions), conditional success (eligible correct/eligible), and unconditional success (eligible correct/all). Existing non-text records give:

Language/modelCoverage (%)Conditional (%)Unconditional (%)
EN FLUX.2-dev99.2593.5292.82
EN ERNIE-Image99.2493.3992.68
ZH ERNIE-Image99.4393.8693.33
ZH FLUX.2-dev99.2793.6292.94
ZH HiDream-I190.2777.5369.98
ZH FLUX.1-dev39.5034.1813.50

These non-text averages differ from HCQ, which balances capabilities and includes OCR. Missing outputs stay in the full-set denominator, separately from prerequisite failures. We also report all three quantities per non-text capability.

The masking ablation counts dependency-invalid questions as failures; OCR is unchanged and missing/unparsable answers stay excluded. HCQ drops 0.64-1.82 pp in EN and 0.49-1.21 pp for 11 ZH models; ZH exceptions are HiDream-I1 (65.02% to 59.57%) and FLUX.1-dev (32.75% to 14.77%). EN ranks stay unchanged; two adjacent ZH pairs swap. Coverage makes the scoring subsets explicit.

W2: Independent validation

"Independent validation is needed"

Response. The original bilingual 200-record challenge audit used two human reviewers and adjudication. We added an audit of 300 DSG edges from 30 concepts with paired EN/ZH graphs. A/B independently checked necessity/direction; human C adjudicated.

Human validationResultRate
Atomicity (original)183/18897.3%
Tag correctness (original)192/20096.0%
Record validity (original)192/20096.0%
VQA answer agreement (original)152/17885.4%
VQA validity agreement (original)187/18899.5%
Edge necessity/direction (new)298/30099.33%
Confirmed missing / estimated valid recorded edges (new)14/4313.25%
We include this summary in the main paper.

W3: Cross-prompt interpretation

"Matching an entity across otherwise different prompts does not control scene difficulty or attached constraints."

Response. For each model/language, the reference averages the same entity's success in at least three other evaluable prompts, excluding the target. References are not selected by scene, outcome, or constraint count and need not be simpler. Local includes all evaluable roots in eligible components; cross-prompt includes roots meeting the reference threshold. Each prompt has equal weight, split among eligible components and included roots.

Unlike the overall entity-view score, this analysis links the same entity across combinations of attributes, relations, and other requirements. Higher rates reflect more reliable entity presence in these contexts, one aspect of compositional ability. Their gap describes the observed contexts and root populations. We clarify these rules beside Figure 1.

W4: Aggregation and correlated errors

"correlated evaluator errors require additional analysis."

Response. HCQ weights valid atomic requirements equally within capabilities and balances five reporting views. Prompt-first gives a one-question prompt the same within-capability weight as one with several questions. In our 1,000-prompt analysis, this gives a short-prompt judgment 48.4% more weight in EN and 41.8% more in ZH. We choose question weighting because the atomic requirement is our evaluation unit. The variance result assumes independent, equal-variance judging errors; HCQ itself does not. Correlated errors limit the variance result, without making prompt-first weighting preferable. We clarify this distinction.

W5: Routing baselines

"Compare against a cost-matched prompt-independent policy and a simple prompt-level router."

Response. Our goal is to demonstrate useful routing signals with a simple rule. Q-Profile-C maps 21 question tags to five groups and uses active groups and training-fold performance to select generators. We added both controls to the same five-fold evaluation of 13,132 held-out EN/ZH requests:

MethodGPU-s/MPHCQ (%)HCQ gap: 95% CI (pp)
Q-Profile-C22.1689.51-
Cost-matched prompt-independent mixture22.1388.83[+0.397,+0.997]
Prompt-text nearest-neighbor20.1289.22[+0.016,+0.577]

Costs count generation only. Intervals are Q-Profile-C minus each baseline, using paired conceptual-prompt bootstrap. The annotations support effective routing on held-out prompts.

Thanks to the authors for the new results. The rebuttal is encouraging at the same time I think it is important to have rigor in the newly reported metrics for example inter annotator agreement for the human evaluation. Some case analysis into the first table would be useful. I am willing to reconsider my score if the rigor is improved

Thank you for the encouraging follow-up. We add Cohen's κ\kappa for the human reviews, followed by a case addressing your concern about different surviving subsets.

1. Cohen's kappa between human reviewers

"inter annotator agreement for the human evaluation."

Response. The paper's VQA answer agreement, 152/178 (85.4%), compares the automatic judge with the final adjudicated human answers. Your request concerns consistency between the two human reviewers. We therefore add Cohen's κ\kappa for their judgments before adjudication:

Human reviewCohen's κ\kappa
Question quality, tags and image answers0.825
Recorded DSG edges0.664

The question/tag/VQA audit covers 200 bilingual records, checking question quality, capability tags, VQA answers, and question validity. Across these checks, the two reviewers gave identical labels on 1,081 of 1,164 applicable judgments. This is the separate two-reviewer result in the original audit; κ=0.825\kappa=0.825 comes from the same audit's reviewer statistics.

For DSG, we calculated κ\kappa from the existing two annotators' judgments on 300 edges from 30 concepts with paired English/Chinese graphs. Necessity and direction were judged separately and had identical counts: both annotators accepted 296 edges, rejected 2, and disagreed on 2. These labels give κ=0.664\kappa=0.664.

The κ\kappa values use the two human reviewers' labels before adjudication. Specifically, κ=(po−pe)/(1−pe)\kappa=(p_o-p_e)/(1-p_e), where pop_o is the proportion of matching human labels and pep_e is chance agreement calculated from the two reviewers' label proportions. A third reviewer resolved disagreements; in the question/tag/VQA audit, this reviewer also checked 20 fully agreed records, overturning none.

Finding missing edges is an open-ended search and is harder than judging a specified edge. We therefore report confirmed omissions: edges independently identified as missing by both annotators and retained after human adjudication. This gives 14 confirmed missing edges relative to 431 estimated valid recorded edges in the sampled graphs (3.25%).

2. Case analysis: different surviving subsets

"Some case analysis into the first table would be useful." Original concern: "can favor models that satisfy prerequisites only on easier examples."

Response. We agree that a high conditional score can reflect a small surviving subset. Consider the same Chinese prompt requesting an elliptical bedroom with a pyramid-shaped photo frame at the top, contrasting curved and angular forms:

Question (English translation)HiDream-I1ERNIE-Image
Is there a room?YesYes
Is the room elliptical?YesNo
Is there a photo frame?No (excluded)Yes
Is the photo frame pyramid-shaped?No (excluded)Yes
Is the photo frame at the top of the room?No (excluded)Yes
Is there a contrast between curved and angular forms?YesYes

These are the saved VQA decisions; "excluded" means marked unevaluable. On the two shape questions, HiDream-I1's conditional success is 100% (1/1), versus ERNIE-Image's 50% (1/2), because HiDream-I1's frame-shape question is excluded when the frame is absent. Both models complete one of two shape requirements: unconditional success is 50% for both, while coverage is 50% versus 100%.

For all six non-text questions, the three metrics in the first table give:

ModelCoverageConditional successUnconditional success
HiDream-I13/6 (50%)3/3 (100%)3/6 (50%)
ERNIE-Image6/6 (100%)5/6 (83.33%)5/6 (83.33%)

This single-prompt contrast is large. Over the full benchmark, replacing conditional non-text scores with unconditional scores in HCQ (OCR unchanged) preserves the English ordering; only two adjacent Chinese model pairs swap. Overall model differences are broadly consistent, while score changes remain larger for low-coverage models.

Conditional scoring measures capability when a requirement can be evaluated, avoiding repeated penalties for an absent prerequisite. Coverage and unconditional success show how often the model reaches the task and succeeds over the full question set. The supplement and its accompanying data already report all three for every non-text capability, model, and language, alongside HCQ. They make the different surviving subsets visible.

We will add these human-reviewer κ\kappa values and this worked case to the supplementary material.

Thank you again for your constructive feedback and willingness to reconsider your assessment. Your suggestions have helped us present the evidence more clearly and rigorously.

Reviewer_s6wmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces QC-T2I-Bench, a question-centric framework for evaluating text-to-image models. It decomposes prompts into atomic questions with capability labels and dependency relations, and proposes Hierarchy-Constrained Question Aggregation (HCQ) to account for these structures during scoring. The same question-level records are used for compositional failure analysis and cost-aware model selection. Experiments on 13 text-to-image models with English and Chinese prompts demonstrate the framework's use for model ranking, fine-grained diagnosis, and training-free routing, with a reported 21.3% reduction in generation cost at a comparable quality point estimate.

Critical strengths
  • The benchmark provides a useful and reusable question-level evaluation framework. It preserves capability labels, dependency validity, and structural information across English and Chinese prompts. The detailed taxonomy, construction audit, and masking ablation provide useful evidence for the reliability of the evaluation framework.

  • The paper includes careful statistical analysis and acknowledges important limitations. The aggregation analysis, paired bootstrap evaluation, and structural controls provide useful evidence beyond aggregate rankings. The routing experiments also use a controlled evaluation protocol and appropriately acknowledge that the reported cost savings do not establish lossless quality preservation.

Critical weaknesses
  • The reliability of the automated evaluator needs further validation. The challenge audit reports agreement with human judgments on 152 of 178 binary questions, but does not establish whether evaluation errors vary systematically across models or capabilities. Such biases could affect the reported rankings and small structural effects, even when bootstrap confidence intervals are narrow.

  • The structural analysis does not clearly separate basic generation failures from failures caused by additional constraints. The cross-prompt comparison matches the same entities across different prompts, but does not ensure that the reference prompts are simpler. This makes it unclear whether the reported differences reflect compositional difficulty. Also, parent–child questions are inherently related (e.g., generating a blue mug requires generating a mug), which may partly explain the observed correlations even after comparison with unrelated questions. More careful controls are needed to support the conclusions.

  • The writing feels overly polished, defensive, and repetitive. The paper frequently uses lengthy qualifications to explain what its results do not establish (e.g., lines 296-297, 416-481, and 462–463). While these caveats are reasonable, their repeated use makes the writing feel formulaic and distracts from the main technical contributions. A more direct and concise presentation would improve readability.

Questions for authors
  • How are the target and reference prompts selected for the cross-prompt root comparison? Are the reference prompts necessarily simpler or subject to fewer constraints? Could the authors also check whether the reported deficit remains after accounting for differences in how entities and prompts are weighted?

  • How does the topology-based analysis account for the semantic dependence between parent and child questions? For example, a child question cannot be satisfied if its prerequisite entity is absent. Could the authors clarify how such cases are handled and whether the direct-edge effect remains after controlling for this dependence?

1 comment

Thank you for the careful assessment and concrete questions.

W1: Evaluator reliability

"whether evaluation errors vary systematically across models or capabilities"

Response.

  1. Across models. Qwen3-VL-8B-Instruct receives only the generated image and fixed question, without generator identity. All generators share the same prompts, questions, frozen judge, and decoding settings. This common protocol follows VQA-based evaluation in TIFA and DSG.
  2. Across capabilities. Every judgment is a VQA task. Accuracy may vary with the visual or factual knowledge required by our capability categories. Within each category, all generators face the same questions and judge with the same knowledge and settings, providing a common comparison standard.
  3. Validation. Two human reviewers and adjudication established the challenge-audit answers; VQA agreement was 152/178 valid, decidable cases (85.4%), now reported in the main text. Official results [1] report 71.5 on RealWorldQA and 70.9 on MMStar as external evidence of VQA ability.

Together, these conditions and validation support a reasonable, procedurally fair comparison.

W2 + Q1: Cross-prompt selection and weighting

"Are the reference prompts necessarily simpler or subject to fewer constraints?" "accounting for differences in how entities and prompts are weighted?"

Response. Targets are recorded nonadditive DSG components with at least one evaluable entity root having records in at least three other prompts for the same model and language. Exact canonical entity keys link prompts. The reference averages those other prompts, excluding the target; selection does not depend on scene, outcome, or constraint count. References are not necessarily simpler.

Local root success includes all evaluable roots in eligible components; cross-prompt root success includes roots meeting the three-other-prompt threshold. Each prompt has equal total weight, split among eligible components and then the roots included on each side. Thus, frequent entities receive more total weight, but adding components does not increase a prompt's weight.

Following your suggestion, we changed prompt-equal to graph-equal weighting while keeping the original root populations:

Mean local-minus-reference gap (pp)Prompt-equalGraph-equal
English-0.478-0.477
Chinese-0.812-0.783

The 13-model averages change little. Local describes entity realization in current compositional prompts; cross-prompt provides a reference for matching entities elsewhere. Both characterize entity retention across compositional contexts; their gap describes the observed contexts and root populations. We clarify selection and weighting beside Figure 1.

W2 + Q2: Parent-child dependence

"whether the direct-edge effect remains after controlling for this dependence?"

Response. We analyze how requirements connected by DSG are completed together. Semantic dependence is part of the recorded relation and is included in this analysis, rather than subtracted from it.

Scoring. If an entity is absent, its existence question fails; dependent questions are excluded to avoid repeated penalties. The supplement's alternative, counting excluded questions as failures, leaves all 13 English ranks unchanged and swaps only two adjacent Chinese pairs.

Joint outcomes. The structural analysis uses the original VQA answers to both questions; parent failure does not overwrite the child's answer. We compute Δ as observed joint success minus the product of two leave-one-prompt-out reference success rates. We compare direct edges with disconnected pairs from the same prompt, matched on ordered Tags, root/descendant roles, and depths:

Result (pp)ENZH
Direct-edge Δ+0.99+1.61
Disconnected Δ-0.11+0.44
Difference+1.10+1.17
Difference: 95% CI[+0.69,+1.50][+0.77,+1.58]

These positive differences show stronger outcome association on recorded edges, quantifying how DSG dependencies appear in generation outcomes and supporting graph-based analysis of joint completion. These are differences in association, not decreases in generation accuracy. We clarify this motivation and interpretation in the paper.

W3: Presentation

"A more direct and concise presentation would improve readability."

Response. We retain the paper's structure, shorten repeated qualifications, and directly explain each analysis's purpose, calculation, and result. We clarify selection and aggregation and improve the abstract, composition discussion, and conclusion.

[1] Qwen Team. Qwen3-VL-8B-Instruct: Official Model Card and Performance Results.

Reviewer_eecyOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes QC-T2I-Bench, a bilingual benchmark for evaluating text to image models at the atomic question level instead of the prompt level. Building on Davidsonian Scene Graphs, the authors break down 6573 prompts into around 94.5k questions in English and Chinese across 5 main groups and 21 secondary capabilities. They propose HCQ aggregation to mask invalid child questions when parent dependecies fail and balance weights across prompts of different complexity, while using OCR edit distance for text rendering. They evaluate 13 open source T2I models and show that the question records can be used for three tasks: reliable ranking, diagnosing compositional failures by separating basic entity difficulty from composition difficulty, and a training free router called Q-Profile that matches ERNIE-Image score while cutting GPU inference cost by 21.3%.

Critical strengths
  1. The statistical analysis for question aggregation and composition is very solid. Prior works like TIFA and DSG also decompose prompts into questions, but this paper goes much further in how the evidence is aggregated and analyzed. In particular, Eq 9 nicely shows why prompt first averaging increases noise variance on short prompts, and the marginal controlled coupling in Eq 7 and 8 carefully separates simple multiplication of independent errors from actual excess failure along DSG edges.

  2. Reusing the evaluation records directly for training free model routing is practical and convincing. Instead of just giving another static leaderboard, the authors show that the 21 capability profiles can guide generator selection on unseen requests and save 21.3% GPU cost without training a separate routing policy.

Critical weaknesses
  1. Details and human validation of the main VQA judge are missing in the main text. The paper mentions Qwen3-235B-A22B for question generation and Qwen3-VL-Instruct-30B for OCR text extraction, but it never clearly states which VLM is used for the binary VQA judgment on non text questions. Also, there is no human agreement study in the main paper to verify the accuracy of the automated DSG graphs or VQA answers, and relying on Qwen models as judges might introduce some bias when evaluating Qwen-Image.

  2. The test time pipeline and cost calculation for the Q-Profile router are a bit unclear. Section 4.4 is very brief and lacks formal equations in the main paper. It says Q-Profile uses the request active views to estimate generator quality, but for a new prompt at test time, how do we get those active views? If it requires running the LLM question construction pipeline first, that latency and GPU overhead could be large and it is unclear if it is included in the GPU-s/MP metric in Table 4.

Questions for authors
  1. Which specific VLM is used as the binary VQA adjudicator for non text capabilities, and do you have human agreement numbers for both the DSG question generation and the VQA judgments?

  2. In Section 4.4, how does Q-Profile get the active views for a new prompt during inference? Does it use ground truth benchmark tags or run an LLM parser, and is that parsing overhead included in the GPU-s/MP cost in Table 4?

  3. For cross prompt root matching in Eq 6, how do you match the same entity e across different open prompts, for example by exact word matching or embedding similarity?

1 comment

Thank you for recognizing our question-level analysis and routing application.

W1 + Q1: Evaluator and validation

"Details and human validation of the main VQA judge are missing in the main text."

Response. Binary VQA and validity decisions use Qwen3-VL-8B-Instruct; OCR uses Qwen3-VL-30B-A3B-Instruct. We make the binary judge explicit in the main text.

The original bilingual 200-record audit used two human reviewers and adjudication. We additionally audited 300 DSG edges from 30 concepts with paired EN/ZH graphs: A and B independently checked dependency necessity and direction, and human adjudicator C resolved disagreements.

Human validationResultRate
Atomicity (original)183/18897.3%
Tag correctness (original)192/20096.0%
Record validity (original)192/20096.0%
VQA answer agreement (original)152/17885.4%
VQA validity agreement (original)187/18899.5%
Edge necessity and direction (new)298/30099.33%
Confirmed missing / estimated valid recorded edges (new)14/4313.25%

The 14 missing edges were independently identified by both annotators and retained after adjudication. We include this summary in the main paper.

Qwen-family concern. We distinguish two possible mechanisms:

  1. Shared text encoders. Prompt representations condition diffusion, but the judge receives only the final image and fixed question, without encoder representations or generator identity. All generators share the same prompts, questions, judge, and settings. Moreover, Qwen-family encoders are used by Qwen-Image and LongCat-Image (Qwen2.5-VL-7B), FLUX.2-Klein-4B and Z-Image/Base/Turbo (Qwen3-4B), and FLUX.2-Klein-9B (Qwen3-8B). HiDream-O1 also initializes its unified backbone from Qwen3-VL-8B. This shared background is not unique to Qwen-Image, reducing concern about encoder-family preference.
  2. Judge training on Qwen-Image outputs. The public Qwen3-VL report [1] does not explicitly disclose such use. Its synthetic HTML, pointing annotations, and code-rendered diagrams do not establish this. Considering its large, diverse training data, published VQA evaluations, and our common judging protocol, we consider shared family membership alone unlikely to cause substantial preference for Qwen-Image.

W2 + Q2: Routing input and parsing cost

"how do we get those active views?" "is that parsing overhead included in the GPU-s/MP cost in Table 4?"

Response. The original experiment used offline-constructed questions, with no test-time LLM parser. Each question has one of 21 secondary tags, mapped to five first-level groups. A group is active if any question belongs to it, counted once: Color and Material & Texture activate Attributes; Spatial Relations activates Relations. The router equally averages each generator's training-data scores over the active groups in the request's language. Q-Profile-C selects the cheapest generator within a validation-chosen tolerance of the best predicted quality. We add this rule to the main text.

The original 21.3% saving counts generation GPU time only, excluding parsing. We additionally measured Qwen3-8B on 2,630 held-out bilingual requests: either predict the five groups directly, or predict applicable tags among 21 and map them to distinct groups. We reran routing and added measured parsing cost; HCQ uses existing image results.

PolicyGenerationParsingTotalHCQ (%)
Fixed ERNIE-Image28.153028.15389.253
Fixed FLUX.2-dev63.340063.34089.116
8B: five groups26.7600.42027.18089.118
8B: 21 tags to groups24.5820.76025.34389.317

Costs are GPU-s/MP; FLUX.2-dev includes both GPUs. Parser-inclusive savings versus ERNIE/FLUX.2-dev are 3.45%/57.09% for direct groups and 9.98%/59.99% for 21 tags. ERNIE is already inexpensive; users defaulting to FLUX.2-dev retain substantial savings. These results demonstrate stable signals for simple routing. Parsing results go in the supplement; the original saving remains generator-only.

Q3: Entity matching

"exact word matching or embedding similarity?"

Response. We lowercase text, remove extra spaces/trailing punctuation, standardize quotation marks and full-width characters, then remove existence-question wording and listed articles/quantity expressions using language-specific rules. "Is there a cat?" and "Does the image show a cat?" both give entity:cat. Keys match exactly, without embeddings. Within the same model and language, a reference averages evaluable records from at least three other prompts, excluding the current prompt. We clarify matching and root populations in Figure 1.

[1] Qwen Team. Qwen3-VL Technical Report. arXiv:2511.21631, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.