1 Introduction↩︎

Recommender systems have become a central part of the modern Web, shaping how users navigate, consume, and engage with online content across e-commerce, media, and social platforms. As the Web continues to integrate multimodal interaction and human-centered personalization, Conversational Recommender Systems (CRS) have emerged as an expressive paradigm for enabling natural, context-aware user experiences [1], [2]. However, building effective CRS remains a significant challenge, owing to their inherently subjective nature, requisite multimodality & human-like conversational ability, and need for generalizable user preference understanding [3]. Among various domains, fashion shopping particularly exemplifies these challenges: preferences are highly subjective and often hinge on visual style and fashion sense, which are difficult to capture through text alone. These issues compound, making a comprehensive evaluation particularly challenging.

Existing conversational recommendation datasets often fall short in evaluating CRS on such difficulties, including the fashion shopping domain. Most are limited to: a single modality (typically text) [4]–[6]; simulated interactions that rely heavily on synthetic rather than organic human dialogues [7]–[9]; narrow domains such as film recommendation  [5], [10]; or omit personal user profiles [5], [7], [9], [11]–[14]. Additionally, Large Language Models (LLMs) and Multimodal LLMs (MLLMs) now serve as the backbone of many CRSs [15], offering strong capabilities for interpreting free-form dialogue and generating natural responses. However, existing datasets and metrics largely emphasize surface-level outcomes, such as the final recommended item, while neglecting the explicit item-level preference ratings and user satisfaction signals [8], [12], [13]. As a result, they fail to expose the limitations of MLLMs compared to human assistants and hinder robust, system-to-system comparison.

In this work we introduce Vogue (Visual-recommendation dialOgue with Grounded User Evaluations), a dataset that addresses these gaps by featuring (1) multimodal grounding with both visual and verbal inputs, (2) authentic human-human dialogues in realistic fashion shopping settings (Fig. 2), (3) a broad and open-ended fashion domain, and (4) comprehensive user profiles capturing style history, preferences, and ratings of non-candidate items. Each conversation is visually grounded by a catalog of candidate items (including images and metadata) and annotated with full post-conversation item ratings from both the Seeker (user) and Assistant (recommender), and subjective satisfaction metrics captured through post-task Likert surveys. This design enables evaluation along two complementary dimensions: item-level preference alignment and subjective satisfaction metrics. Each conversation is also tagged at the utterance level with a novel intent taxonomy specific to the fashion shopping domain. These measures move beyond accuracy alone, supporting richer and more detailed evaluations across a broad range of conversations.

By releasing Vogue, we aim to provide a practical benchmark (cf. Fig. 1) for studying multimodal conversational recommendation. Our dataset establishes an invaluable resource for designing recommendation agents as CRS move into multimodality, providing a foundation to study how dialogue, visual grounding, and user-centered evaluation interact in shaping effective systems.

Our contributions are threefold:

1. Open Dataset of Real Human Dialogues: We introduce the first multimodal conversational recommendation dataset in the fashion shopping domain, containing organic human-human dialogues, associated item images and metadata, participant profiles, and detailed rating signals. The dataset provides utterance-level annotations of Assistant and Seeker intents across conversations.

2. Novel Analysis and Insights: We provide a novel analysis for modeling dialogue dynamics, including stage-wise segmentation (Fig. 4) of conversations and intent-tagged utterance flows in the multimodal setting.

3. MLLM Benchmark Evaluation: We benchmark MLLMs against humans on Vogue, revealing persistent challenges in calibration, generalization, and inference, thus positioning it as a challenge dataset for advancing MLLM agents.

2 Related Work↩︎

To develop an effective end-to-end learning methodology for multimodal CRS, as well as to establish benchmark datasets for evaluation, studies have contributed meaningful datasets to support recommendation tasks in various domains, as listed in Tab. 1. However, existing conversational recommendation datasets are limited and fall short of meeting the needs of our study as follows:

• Lack of Fashion Datasets: Most existing conversational recommendation datasets are not about fashion shopping. They usually focus on movies [4], [5], [11], while some cover restaurants, product search/recommendation or travel [12], [13], [16]–[18], which are not transferable to our study.

• Lack of Explicit User Feedback: The rating included along with the existing datasets serves only as an indicator of conversation quality rather than feedback on recommended items during conversations, as well as overall user satisfaction throughout the whole experience  [12], [16], which hinders comprehensive evaluation of recommendation quality.

• Lack of User History and Profiles: Most existing multimodal conversational datasets either lack personal profiles [7], [9], [14], adapt them directly from interaction data [19], or generate them using LLMs [8]. This reflects a lack of a real user profile associated with the speaker. This limits the opportunity to develop stronger elicitation methods, since long-term preferences are missing [2].

• Lack of Real User Dialogue: For current multimodal conversational datasets targeting fashion shopping, they curated dialogue datasets using data generation methods (e.g., dialogue simulator, LLMs) to produce synthetic dialogue data [8], [19]. However, the deficiency of realistic conversational behaviors present in the synthetic datasets limits the opportunities to study dialogue patterns.

Thus, we believe there is a need for a multimodal conversational dataset supporting visual grounding, consisting of organic human-human dialogues with authentic personal profiles and ratings on recommended items, conversation qualities, and user satisfaction.

3 Dataset Construction↩︎

Based on previously described deficiencies of related work, we aim to curate a novel dataset that satisfies the following desiderata:

1. Multimodal grounding: Combine text, images, and metadata to situate dialogues in a realistic setting.

2. Fashion domain focus: Address the lack of existing datasets centered on fashion shopping.

3. Explicit user feedback: Provide comprehensive and quantitative item-level ratings as well as subjective satisfaction.

4. User history and profiles: Include baseline preferences, rationale, and opinions to support personalization.

5. Organic human–human dialogues: Collect real conversations rather than synthetic or simulated ones.

To meet these desiderata, we designed Vogue as six independent conversational recommendation trials between paired participants. We refer to the recommender as the Assistant and the user as the Seeker. Each trial centered on one of 6 unique and orthogonal fashion scenarios, giving the Seeker a temporary decision-making goal. The full scenarios are included in the dataset repository¹.

In each scenario, participants shared a catalog of 12 candidate items drawn from Amazon.ca, spanning outerwear, layering pieces, and shoes. Each item included images and collated metadata (name, description, review score, review summary). Visual grounding is intrinsic to fashion recommendation, since many critical attributes such as style, silhouette, and overall aesthetic are appearance-dependent. As a result, the task is inherently multimodal and cannot be reduced to a purely textual setting without changing its nature.

To reduce cognitive load, descriptions and reviews were summarized by GPT-4o-mini, while size, fit, color, and price were omitted, and participants were asked not to anchor on them.

Throughout the study, the catalog comprised 36 items (12 per scenario) plus 40 additional items used for fashion profiles. The Seeker’s task was to articulate needs and select an item, while the Assistant guided preference elicitation and made recommendations.

3.1 Protocol↩︎

Due to space limitations, we present a concise description of the protocol and annotation schema in the paper, while the full experimental protocol, survey instruments, intent definitions, and annotation guidelines are provided in the repository. Participant pool and number of trials were set to the largest feasible within our computational and personnel resource budget.

Our data collection protocol² proceeded in three steps:

1. Onboarding surveys: All participants completed demographic forms and fashion entrance surveys, including a profile questionnaire and pre-ratings of 40 non-candidate items, which provided baseline preferences and served as user profiles, a feature rarely included in similar datasets. (Desiderata 1, 2, 4). The complete details for all surveys are in the surveys folder of the supplementary material.

2. Conversational trials: In each of the six scenarios, the Seeker and Assistant discussed 12 candidate items, with the Assistant accessing full metadata and the Seeker seeing only images and the scenario description. A web-based User Interface (UI) facilitated catalogue viewing, with corresponding Seeker and Assistant views (Fig. 2). Assistants were instructed not to use other external resources and to elicit context through dialogue. Conversations continued until the Seeker indicated satisfaction with at least one item or rejected all options (Desiderata 5). The complete details are in the item_ratings sub-folder of Vogue.

3. Post-conversation Surveys: After each scenario, the Seeker provided ground-truth preference ratings of all 12 items on a 1 to 5 scale, while the Assistant attempted to predict these ratings. Participants were instructed to rate based on how likely they would be to purchase the item given the conversation and scenario, with 5 being “would purchase” and 1 being “would never purchase”. Neither party saw the other’s ratings. Both parties also completed a short feedback survey on the perceived quality of that conversation. Seekers additionally completed a 1-5 Likert scale survey of their satisfaction with the scenario and conversation. Survey questions are detailed in Tab. 2 (Desiderata 3).

3.2 Participants↩︎

We recruited 20 participants through informal outreach at the University of Toronto and among recent graduates in the surrounding area. Young adults were recruited in order to reflect a core demographic of fashion shopping who are potentially amenable to the use of conversational assistants. Participants were randomly assigned fixed roles as Seeker or Assistant for the entire session. To prepare Assistants for their role, we conducted a short coaching session to familiarize them with the catalog of candidate items. Seeker participants received only the task scenarios beforehand to replicate real-world conditions, necessitating that Seekers articulate their needs without access to full product knowledge.

3.3 Dialogue Intent↩︎

To better understand dialogue patterns with visual support in fashion shopping recommendations, we extended the holistic user and intent definitions from Lyu et al. [20] and also referred to the work of Cai and Chen on utterance-level labeling [21]. Since earlier studies focused on domains outside fashion shopping under conversational recommendation without visual grounding in conversations, we created new taxonomies on both sides to better categorize dialogue intents and allow for more thorough analysis. The taxonomies used in this study, designed for dialogue actions in a fashion shopping setting with shared visual catalogue, are listed in Tab. 3.

We conducted a two-phase labeling process to annotate dialogue utterances for subsequent analysis. In the first phase, we leverage Gemini-2.5-Flash to streamline large-scale tagging. In the second phase, all tagged conversations were manually reviewed by human annotators to correct mislabeled utterances and to refine and expand the taxonomy based on the collected dialogues. Pairwise inter-annotator reliability was assessed using Cohen’s Kappa, yielding a strong agreement of \(\kappa = 0.81\), which indicates high consistency in intent labeling. From these tags, we derive the intents used in Sec. 4.1. For clarity, we primarily discuss overarching intent categories (e.g., Provide or Explain) rather than specific tags.

3.4 Dataset Contents↩︎

Vogue contains 60 two-party dialogue transcripts, transcribed and time-stamped at the turn level. All 2,100 utterances are labeled with a comprehensive intent taxonomy for both roles. Transcripts are additionally annotated with final item choices and catalogue items explicitly mentioned, and each scenario is paired with its corresponding 12-item visual catalogue released with item metadata.

Beyond transcripts, Vogue includes dense preference supervision: post-scenario item-rating surveys over all items from both Seeker (ground truth) and Assistant (predicted), yielding 1,440 total item-level ratings, plus post-task satisfaction surveys. We additionally collected fashion-profile surveys for all participants to contextualize preference priors. We position Vogue as a high-fidelity evaluation and analysis benchmark for multimodal CRS, and as a challenge dataset exposing shortcomings in visual grounding, preference inference, and preference generalization.

All collected data (cf. Sec. 3.1) are available in Vogue’s repository, including user profiles (fashion_profile folder), item rating surveys from both Seeker and Assistant (surveys folder), and conversational transcripts (conversations sub-folder of data folder). Major dataset statistics are listed in Tab. 4.

4 Data Analysis↩︎

In this section, we examine CRS dialogues from both a qualitative and quantitative perspective, considering their overall structure, the alignment between Seeker and Assistant preferences, and the performance of modern MLLMs relative to the human Assistants.

Specifically, we focus our analysis with Vogue on three complementary aspects: visual grounding, which examines how shared images shape dialogue dynamics in inherently multimodal fashion scenarios; preference alignment, which tests whether Assistants’ alignment with both direction and intensity of Seeker preferences improves Seeker satisfaction; and MLLM as recommender, which evaluates whether current models can move beyond surface-level dialogue cues toward robust and generalizable preference modeling.

To this end, we propose three Research Questions (RQs):

• Visual grounding & dynamics. How does visual grounding alter the dynamics and stages of dialogue, compared to text-only settings? In an inherently multimodal domain like fashion, what structures and processes emerge, and how should CRS adapt to these?

• Impact of Human Assistant alignments on Seeker experience. To what extent do the item-level preference ratings predicted by Assistants after a conversation match the ground-truth ratings from Seekers? And does this alignment on preference actually correlate with the subjective satisfaction of Seekers?

• MLLM versus Human Assistant Alignment. How well do MLLMs perform when they are asked to recover a Seeker’s ratings given the conversation? What strengths and weaknesses do they show compared to human Assistants?

By framing our analysis around these questions, we highlight how Vogue both enables deeper evaluation of existing systems and suggests design principles needed for user-centric multimodal CRS.

4.1 RQ1: Visual grounding & dynamics.↩︎

To address RQ1, we examine how dialogue progresses in an inherently visual recommendation setting, and what structural and behavioral changes result from this grounding. Since a direct non-visual ablation is not possible in this domain, we examine the distribution of conversational intent, item mentions, and intent tag transitions across the course of the interaction and compare with related datasets. For the sake of clarity, our analysis will omit Acknowledgment and Others tags, since they do not meaningfully advance the conversation.

As shown in Fig. 3, Assistants consistently introduce new items in two distinct waves: the first following preference and context elicitation, and the second after refinement based on earlier feedback. We also observe that Assistants frequently group recommendations, presenting up to four items simultaneously for discussion. This contrasts with conversational flows in text-only datasets, where item presentation tends to be more linear and sequential. Chiefly, we observe that the conversations flow naturally through a common implicit procedure shaped by visual grounding.

The distributions in Fig. 3 further suggest a clear stage-like structure: activity is clustered into distinct peaks rather than being evenly spread across the dialogue, with a characteristic ebb and flow as initiative flips between participants. We therefore segment the conversations into five stages, each defined by the dominant (i.e., most prevalent) intents and role dynamics at that point. In Fig. 4, we provide an illustrative example of each stage based on conversations in Vogue:

1. Preference Elicitation. Early exchanges are dominated by preference elicitation (Assistant Request; Seeker Provide, Ask for Recommendation, and Answer).

2. Recommendation. This is followed by the first recommendation wave (Assistant Recommend), with Explain rising soon after. Interestingly, we observed that participants often discussed the items in groups according to abstract unifying properties of those item groups.

3. Critique. The conversation transitions to a bridge characterized by Critique, Inquire, and evaluative ratings, while Assistant Explain and Personal Opinion remain high.

4. Refinement. Stage 4 is strongly characterized by Assistants introducing a new, reduced set of items (usually one or two) to refine the recommendation based on Seeker critiques. Concurrently, Seekers refine their preferences based on available item attributes, continuing to utilize property grouping as observed in Stage 2.

5. Final Agreement. Finally, Stage 5 is marked by Seeker critique and Recommendation Rating peaking, particularly Accept tags, signaling choice and resolution.

4.1.0.1 Significance to CRS

Our analysis indicates that visual grounding and multimodality reshape conversational dynamics compared to text-only CRS settings. Dialogues in Voguetend to follow a staged structure, supported by shared images that keep exchanges relatively compact. Conversations average 14.8 turns and 36 tokens per turn, yet still converge on final agreements with strong preference alignment (see Sec. 4.2.0.1).

Anecdotal inspection of the transcripts indicates frequent usage of positional identifying stylistic references in the shared catalogue that reduces the need for verbose description or minute back-and-forth; contrasting with similar studies lacking a shared visual catalogue [20], and allows participants to anchor their discussion and reasoning in concrete visual referents.

Taken together, these observations point to design implications for CRS: multimodal grounding can enable parallel exploration of grouped items, comparative reasoning, and collaborative refinement, interaction patterns that are far less accessible in text-only dialogue flows. These mechanisms could be directly incorporated into CRS design, providing a way to make interactions more efficient and user-friendly. Exploring how such patterns translate into deployed systems offers a promising direction for future work.

4.2 RQ2: Human Assistant Alignment and Impact on Seeker experience.↩︎

We now evaluate the quality of human Assistant interaction in Vogue using two complementary metrics: (1) quantitative alignment between human Assistant predictions and Seeker ground-truth ratings, and (2) post-task satisfaction surveys from Seekers.

Following [2] and [22], we choose alignment as recovery of both the structure and intensity of preferences over accuracy, reporting it via Pearson Correlation (PC) and Mean Absolute Error (MAE) between item-level ratings. Satisfaction is included as an orthogonal but essential user-centered measure and reported on a 1-5 Likert scale.

4.2.0.1 Alignment between Assistants and Seekers

We observed that Seekers strongly favored low scores (44% 1s, only 8.6% 5s), resulting in a highly imbalanced distribution. Human Assistants nonetheless achieved good preference alignment, with a mean PC of 0.64 (median 0.61, SD 0.21) as shown in Fig. 5, indicating reliable recovery of preference directions rather than isolated success cases. MAE varied more across scenarios, from 0.25 to  2.0, averaging 0.90, underscoring persistent challenges in modeling preference strength.

Since each Seeker engaged with the same Assistant in six scenarios, Fig. 5 also analyzes the performance across repeated interactions. PC improved and MAE declined over successive scenarios, suggesting that Assistants became more effective with practice, despite neither party ever seeing the other’s item ratings, likely due to increased dyad-specific calibration over time. While improvement trends were modest, the result underscores the value of long-horizon modeling in CRS and indicates potential for adaptive learning without explicit item-level feedback.

4.2.0.2 Correlation between Alignment and Satisfaction.

Unlike prior CRS resources, our design captures not only objective alignment through item ratings but also subjective satisfaction from Seekers, enabling a more holistic evaluation of recommendation quality. In Fig. 6, we observe that Seeker satisfaction (y-axis) was consistently high. Likert surveys captured perceived decision quality, justification, intent to purchase, and Assistant helpfulness. Most scores exceeded 4.5 (out of 5), with helpfulness especially positive.

We further analyze the correlation between the Assistant’s alignment and the Seeker’s satisfaction (cf. Fig. 6). Encouragingly, higher satisfaction generally coincides with stronger alignment (measured by PC and MAE), indicating that better alignment contributes to better user experiences. This strong correlation also supports the use of alignment metrics as a proxy for preference inference quality when user satisfaction is unknown, such as in the post-conversation MLLM evaluation in the following sections.

Nonetheless, satisfaction varied across perceived decision impact and purchase intent, underscoring that it is not monolithic. This suggests that measuring only alignment, as is common in existing datasets, is insufficient. By contrast, the comprehensive design of Vogue highlights the opportunity for CRS to maximize impact by coupling robust preference modeling with user-centered design. In accordance with prior research [22], a system that feels supportive and transparent, while also aligning with user preferences, is more likely to foster sustained user satisfaction and trust.

4.3 RQ3: MLLM versus Human Assistant Alignment.↩︎

The recent surge of MLLMs in conversational recommendation [15] leads to questions about their ability to operate on human-human dialogue in recommender and preference modelling roles. Having established the impact that alignment can have on user experience, we now examine how well MLLMs perform in this domain. To answer RQ3, we provide an evaluation of several prevalent MLLMs on the proposed dataset. We provide the prompt template as follows:

You are an expert sales associate and fashion expert, with deep experience in apparel, styling, sales, and customer relations, reviewing a recorded conversation between an assistant and a seeker. The assistant provides shopping advice and help to the seeker, who is attempting to make a clothes purchase. You must rate each of the 12 items on a 1-5 scale (inclusive).

Criteria: Rate the items based on how likely you think the Seeker is to purchase them, given the conversation and scenario that just concluded.

<format guidelines>
—
<catalogue, metadata and conversation history here>

The goal of this experiment is to evaluate how well MLLMs can recover Seeker preferences from the available post-conversation evidence, and not end-to-end conversational policy. We do not compare specific fine-tuned or purpose-built MLLM-based CRS for fashion recommendation, as it is beyond the scope of this work. Accordingly, all models are provided with the same prompt as above, the full item catalogue & metadata, and the complete tagged transcript at the conversation level. We then instruct them to complete the same item rating survey task as the Assistants. We also include two baselines: (i) Baseline: Random, which randomly samples ratings from the Seeker’s empirical ground truth distribution, and (ii) Baseline: Mode Rating, which returns the mode (i.e. most common) rating of 1. Our analysis first examines rating distributions, followed by item-level alignment and preference generalization.

4.3.0.1 Performance Overview

As before, we report PC and MAE; however, due to class imbalance (which most strongly affects the MLLMs) we also report per class (\(\text{MAE}[\text{GT}=k]\) for Seeker ground truth (GT) rating \(k\)) and Macro MAE (M-MAE). We report accuracy for all models, though we discuss later that this statistic can be misleading due to class imbalance. Due to the retroactive nature of these evaluations, no model is actively participating in the recommendation process so satisfaction metrics are not relevant here. The results are presented in Tab. 5. We note that the Assistants score best or 2nd best in nearly every measure, even accounting for the learning effects seen in Sec. 4.2.0.1. Critically, no MLLM alternative tested was competitive with real human recommenders.

4.3.0.2 Calibration and Distribution

Examining a distribution of ratings, we note that both families of models experience systemic failure in replicating Seeker ratings; Gemini models heavily over-estimate the proportion of 1 ratings, while heavily collapsing on 3s and 4s; GPT-4o-mini adheres to a strongly normal distribution, peaking at 3, while GPT-5-mini produces a nearly linear decline across ratings (Fig. 7). All models maintain a relatively fair proportion of 5 ratings. This deficit is directly reflected in the PC performance of each model, where they perform very poorly, especially in comparison to human Assistants. Poor performance at capturing natural human rating distributions directly undermines the ability of these models to effectively predict preferences.

4.3.0.3 Weak Preference Generalization

Next we specifically examine how well different recommenders are able to infer and generalize preference knowledge within the conversation. Since it has been shown that most state-of-the-art MLLMs are able to parse explicit dialogue signals well [23], we foreground the ability to extend and generalize explicit preferences onto the full catalogue. Since many of these items may not have been explicitly discussed, the ability to generalize from the discussion to accurately predict any item rating in the full catalogue is paramount for effective recommendation.

From the per-conversation M-MAE for the MLLM models and two baselines, we note a distinct and consistent gap from the Assistants’ performance (cf. Tab. 5 and Fig. 1). Both Gemini models perform worse than the Random baseline. This underscores that extending preference inference beyond explicitly discussed items remains a challenging open problem, even for state-of-the-art MLLMs.

4.3.0.4 Implications for MLLMs in CRS

Viewed from the naive lens of accuracy, many MLLMs appear to perform comparably to human Assistants. However, this apparent parity is largely an artifact of class imbalance and the relative ease of recovering extreme ratings from conversational cues. When results are disaggregated by rating (i.e., \(\text{MAE}[\text{GT}=k]\) in Tab. 5), none clearly surpasses the Assistants. In fact, a trivial and uninformed strategy of predicting all 1s (Baseline: Mode Rating) beats all MLLMs in raw accuracy. Accordingly, accuracy is not a meaningful primary metric for preference alignment in Vogue; we recommend PC as the headline measure.

This dissonance stems in part from systematic distortions in rating distributions, which prevent MLLM-based recommenders from faithfully capturing Seeker ambivalence or partial fit. In turn, this weakens their ability to generalize preference reasoning beyond explicit conversational cues and actively discussed items. We do not observe substantial evidence that any of these models form a coherent preference model of the Seeker.

In direct response to RQ3, these results demonstrate that extending preference inference beyond explicit conversational cues remains an open challenge for state-of-the-art MLLMs. The dataset introduced here directly foregrounds this gap by enabling evaluation across fully featured candidate catalogues and comprehensive and detailed item-level feedback, providing a benchmark that moves beyond surface-level metrics like accuracy and supports rigorous comparison of MLLM-based CRS systems in terms of their capacity for nuanced and generalizable preference modeling.

5 Conclusion↩︎

We introduced Vogue, a multimodal conversational recommendation dataset that addresses current gaps by combining (i) multimodality, with fashion catalogs containing both product images and metadata to ground dialogue in visual and textual context; (ii) human–human conversations, with 60 natural Seeker–Assistant dialogues in contrast to synthetic or model-generated interactions; (iii) domain specificity, focusing on fashion, where subjective, context-sensitive choices make CRS evaluation especially challenging; and (iv) explicit user profiles, derived from pre-task surveys capturing style preferences, values, and behaviors.

Our stage analysis shows a consistent five-step rhythm: Assistant-led elicitation, Seeker critique, and convergence on choices. Visual grounding and grouped recommendations support comparative reasoning and faster convergence, while transparent justification remains crucial when alignment falters. At the item level, human Assistants still outperform large models: while MLLMs can follow explicit cues, they fail to capture relative preferences or generalize across all items. This underscores preference modeling and rating calibration as central open challenges.

Taken together, Vogue foregrounds item-level alignment, conversational stage dynamics, and satisfaction in a realistic, visually grounded setting. Most importantly for the development of future multimodal CRS, our findings establish Vogue as a challenge dataset beyond the current recommendation capabilities of existing top-tier MLLMs such as GPT-4o-mini, GPT-5-mini, and Gemini-2.5-Flash.

References↩︎

---

1. The complete Vogue dataset is available at https://github.com/D3Mlab/vogue ↩︎

2. The protocol received approval from the University of Toronto Research Ethics Board (#00049065).↩︎