Hospitality-VQA: Decision-Oriented Informativeness Evaluation for Vision-Language Models

Hospitality-VQA: Decision-Oriented Informativeness Evaluation for Vision–Language Models Jeongwoo Lee 1,∗ , Duhyeong Baek 1 , Eungyeol Han 1 , Soyeon Shin 1 , Gukin Han 2 , Seungduk Kim 2 , Jaehyun Jeon 1,† , Taewoo Jeong 2,† 1 leejeongwoo9941,glzeng99,condense,shin020810,jaehyun.jeon@yonsei.ac.kr 2 bryan.han,seungduk.kim,taewoo.jeong@yanolja.com 1 Yonsei University, 2 Yanolja NEXT Abstract Recent advances in Vision–Language Models (VLMs) have demonstrated impressive multi- modal understanding in general domains. How- ever, their applicability to decision-oriented do- mains such as hospitality remains largely un- explored. In this work, we investigate how well VLMs can perform visual question an- swering (VQA) about hotel and facility im- ages that are central to consumer decision- making. While many existing VQA bench- marks focus on factual correctness, they rarely capture what information users actually find useful. To address this, we first introduce In- formativeness as a formal framework to quan- tify how much hospitality-relevant information an image–question pair provides. Guided by this framework, we construct a new hospitality- specific VQA dataset that covers various fa- cility types, where questions are specifically designed to reflect key user information needs. Using this benchmark, we conduct experiments with several state-of-the-art VLMs, reveal- ing that VLMs are not intrinsically decision- aware—key visual signals remain underuti- lized, and reliable informativeness reasoning emerges only after modest domain-specific fine- tuning. 1 Introduction Images play a central role in the hospitality in- dustry, serving as the primary medium through which guests evaluate and compare accommoda- tions (Zhang et al., 2022). When consumers choose where to stay, they often rely more on visual impressions—such as room layout, view, light- ing, and cleanliness—than on textual descriptions. These images convey both factual and atmospheric cues that shape user decisions, making visual un- derstanding a crucial aspect of hospitality intelli- gence (Cuesta-Valiño et al., 2023). ∗ Main contributor. † Corresponding author. Existing VQA Dataset : General Questions Q: What color is the slide? A: Orange Q1: What type of this facility is this space? A1: Room Interior Q2: What activities do this space support? A2 : Sleeping, Sitting Ours: Hospitality-VQA : Decision-Oriented Questions Figure 1: Comparison between general VQA (top) and decision-oriented Hospitality-VQA (bottom). Previous studies in the hospitality domain have predominantly relied on text-based analytics of on- line reviews to model customer satisfaction, pref- erences, and demand patterns (Li et al., 2013; Xi- ang et al., 2015). In parallel, a growing body of work has examined visual information in accom- modation images by extracting predefined or low- level features—such as aesthetics, composition, or object categories—and relating them to outcomes such as booking decisions, user intentions, or per- ceived accommodation quality (Ren et al., 2021; He et al., 2023). More recently, while some studies have leveraged Large Language Models (LLMs) arXiv:2603.07868v1 [cs.AI] 9 Mar 2026 for hospitality analysis, they remain primarily fo- cused on textual inputs such as reviews or descrip- tions (Guidotti et al., 2025). Despite these advance- ments, existing methods—whether text-centric or feature-based—remain limited in their ability to perform integrated multimodal reasoning. Specif- ically, they often fail to capture the interplay be- tween higher-level spatial organization and func- tional semantics in images, factors that are central to how humans evaluate hospitality environments. Meanwhile, recent advances in Vision-Language Models (VLMs) have significantly improved multi- modal reasoning across general domains. Modern models (Comanici et al., 2025; Hurst et al., 2024; Bai et al., 2025) can generate contextualized im- age descriptions and answer open-ended questions that go beyond traditional visual recognition, sug- gesting strong potential for domain-specific appli- cations. While these models have demonstrated promising results in specialized fields such as e- commerce (Trabelsi et al., 2025) and medical imag- ing (Tu et al., 2024), their use in the hospitality domain has been relatively limited with respect to decision-oriented evaluation settings. When examining these domain-specific applica- tions, one important insight emerges: the perfor- mance of VLMs often depends on how information needs are framed. Generic prompts (e.g., "What is in this image?") yield vague descriptions that are in- sufficient for hospitality purposes. As illustrated in Figure 1, appearance-level questions alone provide limited insight into whether a space meaningfully supports guest activities or experiences. Mean- ingful evaluation requires domain-specific ques- tions that elicit decision-relevant insights—not just whether a room contains furniture, but how its lay- out supports guest activities; not merely whether a window exists, but what type of view it provides. This raises a key design challenge: how to for- malize the kinds of visual evidence that actually support user decisions. To address this challenge, we introduce Hospi- tality Informativeness, a domain-grounded frame- work that quantifies how much decision-relevant information a hospitality image–question pair pro- vides. Because user information needs vary across facility types—such as spatial clarity in rooms, amenity completeness in bathrooms, or functional elements in shared facilities (Wakefield and Blod- gett, 1996)—we first identify the facility type and design domain-specific questions accordingly. Al- though these needs appear diverse, we observe that the visual cues influencing booking decisions con- sistently fall into a small set of structural, func- tional, and view-related dimensions. Building on this observation, we define four fundamental visual axes (spatial legibility, activity affordance, contex- tual openness, and geometric completeness). To- gether, these axes capture the dominant cues that shape user perception and decision-making in hos- pitality imagery, providing a principled basis for evaluating VLM responses (Greene et al., 2016). We use these axes to construct Hospitality-VQA, a new VQA benchmark aligned with decision-centric evaluation rather than generic scene description. Our main contributions are: • We formalize Informativeness in the hospital- ity domain as a set of four interpretable axes that capture decision-relevant visual cues in hotel and facility imagery. •We build Hospitality-VQA, a VQA dataset whose questions and labels are derived from these axes and tailored to diverse facility types. •We benchmark eight general-purpose VLMs and show that they struggle with fine-grained hospitality informativeness. Our dataset en- ables measurable performance gains through lightweight domain adaptation, highlighting its value as a foundation for future model de- velopment on hospitality domain. 2 Related Works 2.1 Visual Analysis in Hospitality Research in hospitality AI has largely focused on structured prediction tasks such as room-type clas- sification and price estimation using CNN-based frameworks, treating images as static inputs and overlooking richer semantic cues relevant to user assessment. In parallel, a growing line of work ex- tracts computable visual descriptors—ranging from low-level color statistics to mid-level attributes such as aesthetics and composition—and relates them to outcomes like booking intentions or de- mand (Zhang et al., 2022; He et al., 2023; Cuesta- Valiño et al., 2023). However, these approaches are not designed to evaluate whether models can answer decision-relevant questions about an im- age. Recent work has also explored multimodal hotel retrieval and preference matching (Askari Bad Good Spatial LegibilityActivity AffordanceContextual OpennessGeometric Completeness Figure 2: Bad vs. Good examples for each informativeness dimension. Bad images lack decision-relevant visual cues—resulting in low spatial legibility, weak activity affordance, obstructed or unbalanced contextual openness, or incomplete geometric completeness. Good images exhibit high spatial legibility, clear activity affordances, well-balanced contextual openness, and strong geometric completeness, enabling more reliable assessment of hospitality informativeness. et al., 2025), but focuses on similarity or rele- vance rather than explicit question answering and decision-oriented evaluation. Existing approaches rarely model how guests simulate a potential stay experience from visual evidence. Although the presentation of accom- modation photographs can sway selection behav- ior (Sánchez-Torres et al., 2024), the notion of visual utility—how visual elements convey func- tional and spatial affordances—remains under- specified. Consequently, evaluation typically cen- ters on prediction accuracy or correlational signals rather than decision-oriented reasoning. Our work addresses this gap by shifting the focus to the sys- tematic evaluation of decision-relevant information through a VQA benchmark grounded in Hospitality Informativeness. 2.2 Vision–Language Models and Domain Adaptation Recent general-purpose VLMs, including GPT- 4o (Hurst et al., 2024) and Gemini 2.5 Pro (Co- manici et al., 2025), demonstrate impressive capa- bilities in image captioning and open-ended QA. However, these models are trained primarily on web-scale, caption-style data to describe “what ex- ists,” often lacking the specialized reasoning re- quired to evaluate “how useful it is” in a vertical domain. In hospitality, visual understanding goes beyond object recognition; it requires inferring spa- tial habitability and functional affordance. Since standard VLMs are not inherently optimized for such evaluative reasoning, it remains unclear to what extent they can interpret the nuanced visual evidence essential for consumers—motivating the need for a domain-grounded benchmark. 2.3From Factuality to Decision-Centric VQA Standard VQA benchmarks (e.g., VQA v2 (Goyal et al., 2017), GQA (Hudson and Manning, 2019)) have driven progress in multimodal reasoning but primarily evaluate factual correctness or common- sense knowledge. While recent goal-oriented VQA tasks explore navigation or physical manipula- tion (Das et al., 2018; Gurari et al., 2018), they rarely address consumer-facing decisions in which the goal is to assess the suitability of a space or service. Existing benchmarks are not designed to measure whether an image provides the type of evi- dence needed to support informed accommodation choices (Cuesta-Valiño et al., 2023). Addressing this limitation, we introduce Informativeness as a metric to quantify the specific visual signals—such as layout clarity and functional completeness—that facilitate reliable accommodation assessment. 3 Quantifying Informativeness in Hospitality We argue that true understanding in the hospitality domain requires quantifying the visual evidence that supports user decision-making. While gen- eral VQA benchmarks focus on factual correctness (e.g., “is there a window?”), hospitality users rely on images to envision their stay—judging layout, Facility TypeSL A CO GC Room Interior• Indoor Facility• Outdoor Facility• Accommodation Exterior• Table 1: Facility types and applicable informativeness dimensions (SL: Spatial Legibility; A: Activity Af- fordance; CO: Contextual Openness; GC: Geometric Completeness). affordance, and atmosphere. Because these subjec- tive assessments directly drive booking decisions, mere descriptions are insufficient (Cuesta-Valiño et al., 2023). To address this, we formalize Infor- mativeness as a measurable metric. We propose that the vague notion of “a useful hotel image” can be decomposed into specific, quantifiable axes that act as proxies for the user’s envisioned stay experi- ence (Greene et al., 2016). 3.1 Facility Taxonomy and Informativeness Dimensions Hospitality imagery encompasses diverse scenes, ranging from critical facility views to irrelevant content. To structure our analysis, we categorize images into five main facility types: Room Interior, Indoor Facility, Outdoor Facility, Accommodation Exterior, and Irrelevant. To capture more specific functional contexts, images are additionally anno- tated with finer-grained sub-facility labels; the full taxonomy is provided in the Appendix A. We define an image as informative if it provides quantifiable visual cues along four fundamental dimensions: Spatial Legibility (SL), Activity Af- fordance (A), Contextual Openness (CO), and Geometric Completeness (GC). Figure 2 illus- trates the characteristic visual patterns correspond- ing to each dimension. The applicability of these dimensions depends on the facility type, and Ta- ble 1 specifies which dimensions serve as valid evaluative criteria for each category. Beyond these dimensions, Room Interior im- ages are additionally annotated with two semantic attributes—view type and room type—to capture preferences not fully represented by geometric or functional cues. Conversely, the Irrelevant cat- egory contains images lacking decision-relevant visual evidence and is excluded from further evalu- ation. Hospitality-VQA Annotation Schema (1) Hierarchical Labels main: Primary facility category sub : Fine-grained sub-facility (2) Informativeness Axes SL, A, CO, GC *Mapped based on Table 1 Figure 3: The formal annotation schema used in Hospitality-VQA. We record hierarchical facility labels and quantify visual utility across the four informative- ness dimensions. 3.2 Axis Definitions We define the four axes as quantifiable prediction targets to measure visual utility. Spatial Legibility. Defined as the count of dis- tinct planar surfaces (floor, walls, ceiling), this met- ric serves as a proxy for spatial comprehension, distinguishing ambiguous close-ups from structural views that reveal room volume (Oliva and Torralba, 2001). Activity Affordance. We quantify meaningful components—functional objects that explicitly af- ford guest activities (e.g., desks, seating, storage surfaces)—to capture the space’s functional hab- itability while filtering out purely decorative ele- ments (Greene et al., 2016). Contextual Openness. Measured as the ratio of non-facility elements (sky, nature, background structures), this metric assesses contextual bal- ance, identifying overly occluded views or exces- sively distant shots that hinder environmental inter- pretation (Cuesta-Valiño et al., 2023). Geometric Completeness. Approximating a building as a dominant cuboid, we assess the visi- bility of its three canonical faces—front, side, and roof—to evaluate geometric integrity and the per- ceptibility of its 3D form (Sánchez-Torres et al., 2024). For Room Interior images, we supplement these structural axes with two semantic attributes—View Type and Room Type—which capture domain- specific preferences essential for booking deci- sions. Room Interior 25.3% Indoor Facility 20.1% Outdoor Facility 25.7% Accommodation Exterior 22.3% Irrelevant 6.6% (a) Main Category 0123 Planes 0 20 40 60 80 100 Proportion (%) 0.6% 1.4% 4.2% 93.8% (b) Spatial Legibility 01234567 # Components 0 5 10 15 20 25 30 35 Proportion (%) 31.4% 12.1% 21.1% 21.0% 11.2% 3.0% 0.1% 0.0% (c) Activity Affordance 0 1920394059607980100 Proportion Range (%) 0 10 20 30 40 50 60 Proportion (%) 8.4% 17.8% 14.9% 54.1% 4.9% (d) Contextual Openness VisiblePartialNot in ViewCut Building Visibility State 0 10 20 30 40 50 60 Proportion (%) 52.8% 18.2% 27.9% 1.1% (e) Geometric Completeness Figure 4: Dataset statistics of Hospitality-VQA. (a) Distribution of main facility categories. (b–e) Distributions of the four informativeness axes, reflecting characteristic properties of professionally curated hospitality listing images. 4 Hospitality-VQA Dataset To translate the informativeness framework into a measurable benchmark, we introduce Hospitality-VQA. As existing VQA datasets lack the hospitality-specific imagery and annotations aligned with the four informativeness dimensions, they are ill-suited for evaluating decision-oriented visual reasoning. To address this gap, our dataset provides expert-annotated supervision explicitly tailored to these axes. The following subsections detail our pipeline for image collection, hierar- chical annotation, and the derivation of instruc- tion–answer pairs. 4.1 Data Collection A total of 5,000 hospitality images were collected fromnol.yanolja.comthrough random sampling of listing pages. Because the pool was not pre- filtered by facility type, categorization into facility types and relevance labels was performed during annotation (Section 3). 4.2 Data Annotation Annotation was conducted by five annotators who were instructed in the Informativeness Framework defined in Section 3. Figure 3 summarizes the anno- tation schema used throughout dataset construction. A pilot round was conducted prior to the main an- notation phase to calibrate labeling practices and align annotators’ interpretations. All 5,000 images were then independently labeled by all annotators. For quality control, we adopted a strict consen- sus protocol. Labels with high agreement (at least 4 out of 5 annotators)—covering 86.4% of all an- notations—were accepted as ground truth. Cases with lower agreement were flagged and resolved through consensus discussion, producing finalized facility-type and axis annotations for every image. For VLM assessment, each label is converted into a concise instruction–answer pair using fixed templates specifically defined for each facility type and axis. These templates are designed to ensure consistent and scalable evaluation by mapping clas- sification targets into a VQA format while con- trolling variation in question phrasing. Only the templates applicable to an image’s facility type are instantiated, and example templates are shown in Appendix B. 4.3 Dataset Analysis Across the 5,000 collected images, a total of 19,729 QA pairs are generated by applying the fixed tem- plates to the facility-type and axis-level annotations. Figure 4 summarizes the overall distributions of fa- cility categories and informativeness annotations. As shown in Fig. 4a, the main facility categories are relatively well balanced, each accounting for roughly a quarter of the dataset. Figures 4b–e report the distributions of the four informativeness dimensions. Spatial Legibility (Fig. 4b) and Activity Affordance (Fig. 4c) are summarized as discrete counts reflecting visible planes and meaningful components, respectively, while Contextual Openness (Fig. 4d) and Geomet- ric Completeness (Fig. 4e) are reported using pre- defined categorical bins. Across these informativeness axes, we observe skewed distributions toward higher levels of visual informativeness. Such tendencies are characteristic of official listing imagery provided by hospitality platforms, which typically relies on professional photography to enhance spatial clarity, contextual visibility, and overall visual appeal for potential guests. Unlike user-generated review photos, these images are intentionally composed to reveal room structure, spatial volume, and surrounding context. As a result, the observed distributions reflect the visual evidence that users commonly encounter dur- ing actual booking decisions, supporting the eco- logical validity of our benchmark. This structured coverage enables models to be evaluated not only on generic scene understanding, but also on the decision-relevant visual properties that matter in hospitality settings. In Section 5, we use this dataset to benchmark several state-of-the-art VLMs and analyze their per- formance across facility types and informativeness axes. 5 Experiments We evaluate a range of general-purpose Vision– Language Models (VLMs) on Hospitality-VQA to examine how well they capture the domain-specific informativeness axes introduced in Section 3. 5.1 Experimental Setup Data split. Hospitality-VQA contains 5,000 la- beled accommodation images. We reserve 300 images for evaluation. The remaining 4,700 im- ages are used for training. The evaluation split is sampled to preserve the overall distribution of fa- cility types and informativeness factors, with class proportions matched within a 5% margin relative to the full dataset. Models.We evaluate eight vision–language mod- els that span both commercial APIs and open- weight systems: GPT-5 (OpenAI, 2025), GPT-4o- mini (Hurst et al., 2024), Gemini 2.5 Pro (Comanici et al., 2025), GLM-4.1V-9B-Thinking (Hong et al., 2025), Qwen2.5-VL-3B and Qwen2.5-VL-7B (Bai et al., 2025), LLaVA-NeXT-7B (Li et al., 2024), and Gemma-3-12B (Team et al., 2025). The pro- prietary models (GPT-5, GPT-4o-mini, Gemini 2.5 Pro, and GLM-4.1V-9B-Thinking) serve as strong general-purpose assistants that have been optimized for broad, web-scale multimodal use, whereas the open-weight models (Qwen2.5-VL-3B, Qwen2.5-VL-7B, LLaVA-NeXT-7B, and Gemma- 3-12B) provide instruction-tuned checkpoints with varying capacities and training pipelines that are accessible for research and adaptation. This combi- nation allows us to examine how both deployment setting and model family affect performance on hospitality-oriented VQA. Beyond zero-shot evaluation, we also derive task- adapted variants of Qwen2.5-VL-3B and Qwen2.5- VL-7B by applying LoRA fine-tuning (Hu et al., 2022) on Hospitality-VQA. In this configuration, the models are trained to predict the discrete axis la- bels in our framework from an image–prompt pair, aligning their outputs with our informativeness- oriented, classification-style supervision rather than generic captioning or open-ended generation. Tasks and metrics. To align with real-world hospitality applications (e.g., booking platforms) that require discrete, interpretable attributes rather than free-form text, we formulate all tasks as classification problems.We evaluate six core tasks—main facility type, main+sub facility type, visible plane count, meaningful component count, discretized scenery proportion, and building-face visibility—along with two auxiliary interior at- tributes: room and view type. Models are prompted with a natural-language instruction template and must output a single cat- egorical label. We report exact-match accuracy, reflecting the binary nature of practical decision- making; predictions that fail to map to a valid label are strictly penalized, mirroring real-world failure modes in attribute extraction systems. For API- based models, we use deterministic decoding (tem- perature = 0). ModelFacilityInformativeness Main Main&SubSLAACOGCRoom View Gemini 2.5 Pro90.6675.0011.51 9.43 46.81 7.3580.65 50.00 GPT-592.3382.5546.76 18.87 31.91 20.5983.8764.10 GPT-4o-mini 92.3384.9197.12 38.21 56.03 8.8270.97 79.49 GLM-4.1V-9B-Thinking93.6679.2589.21 35.85 57.4516.1861.29 57.69 LLaVA-NeXT-7B73.3353.7794.24 8.02 19.86 5.8825.81 79.49 Gemma-3-12B 92.0082.0886.33 15.09 55.32 22.0672.19 43.59 Qwen2.5-VL-3B64.6644.3468.35 19.34 39.72 1.4741.94 70.51 Qwen2.5-VL-3B Finetuned86.6681.1394.9642.9257.4526.4780.65 76.92 Qwen2.5-VL-7B78.6664.1543.88 25.94 48.94 5.8825.81 69.23 Qwen2.5-VL-7B Finetuned92.0085.3797.12 44.34 67.37 32.3587.10 74.36 Table 2: Comparison of VLM performance across facility types and informativeness categories. Best in each column is in bold and second-best is underlined. 5.2 Overall Results Table 2 reports accuracy across facility recognition and all informativeness-related tasks. We summa- rize the results by (i) task difficulty across axes, (i) model-family trends, and (i) the effect of domain adaptation. 5.2.1 Task Difficulty Across Axes Table 2 shows that main facility classification trans- fers well across most evaluated VLMs, with sev- eral models exceeding 90% accuracy. In contrast, main&sub recognition is consistently lower, indi- cating that fine-grained sub-category prediction is more demanding than coarse scene categorization under the same prompting and evaluation protocol. Axis-level tasks exhibit a sharper drop in perfor- mance than facility recognition. Across models, Spatial Legibility (SL) is generally more stable than the other informativeness axes, whereas Ac- tivity Affordance (A) and Geometric Complete- ness (GC) are notably weaker for many models. Contextual Openness (CO) falls between these ex- tremes but still remains substantially below facility recognition performance, suggesting that decision- relevant attributes are not reliably recovered from generic multimodal capabilities alone. Room and view attributes for Room Interior show additional variability across models. While some models achieve strong accuracy on these aux- iliary tasks, others lag despite high facility recog- nition, reinforcing that success on global catego- rization does not guarantee robust prediction of hospitality-relevant fine-grained attributes. 5.2.2 Model Family Trends Model families show broadly similar behavior on coarse facility recognition but diverge more on axis-level prediction. Several proprietary models achieve high accuracy on main facility classifica- tion, and some open-weight models also reach com- parable levels, indicating that recognizing the over- all facility category is not the primary bottleneck in this benchmark. Differences become more pronounced for infor- mativeness axes. For instance, GPT-4o-mini attains very high SL accuracy (97.12), yet A and GC remain much lower (38.21 and 8.82). A similar pattern appears in multiple open-weight baselines (e.g., Qwen2.5-VL-7B: SL 43.88 vs. A 25.94 and GC 5.88), where axis-level prediction does not track facility recognition. These results suggest that axis performance reflects additional reasoning requirements beyond generic scene labeling. We avoid attributing these gaps to a single cause, as controlled ablations over training data, vision encoders, and instruction-tuning procedures are outside the scope of this work. Nonetheless, the consistent separation between facility recognition and axis-level performance across both propri- etary and open-weight systems motivates explicit domain-grounded supervision for decision-oriented attributes. 5.2.3 Effect of Domain Adaptation Domain adaptation via LoRA fine-tuning (Hu et al., 2022) consistently improves Qwen2.5-VL models across all evaluated tasks. Table 3 reports absolute gains (Finetuned−Base) computed from Table 2. Improvements are observed for both coarse facil- Task3B (∆ Acc) 7B (∆ Acc) Main+22.00+13.34 Main&Sub+36.79+21.22 SL+26.61+53.24 A+23.58+18.40 CO+17.73+18.43 GC+25.00+26.47 Room+38.71+61.29 View+6.41+5.13 Table 3: Absolute accuracy gains (%) from domain adaptation via LoRA fine-tuning for Qwen2.5-VL mod- els (Finetuned−Base). ity recognition and fine-grained facility prediction, with particularly large gains on main&sub classifi- cation. Gains are also evident on informativeness axes, which are challenging in the zero-shot setting. No- tably, both model sizes improve on A, CO, and GC, while the 7B model shows a pronounced in- crease on SL. Interior attributes benefit as well: room type accuracy increases substantially for both models, whereas view type shows smaller but con- sistent gains. Overall, these results indicate that axis-aligned supervision in Hospitality-VQA pro- vides an effective signal for aligning VLM outputs with decision-oriented hospitality attributes under a strict label-matching evaluation. 6 Conclusion This work addressed the gap between general- purpose visual understanding and the kinds of fine- grained, decision-relevant reasoning required in the hospitality domain. While images play a central role in shaping guest expectations and booking de- cisions, existing multimodal systems lack the struc- tured grounding necessary to interpret the spatial, functional, and view-related cues that matter in real domain use cases, just interpreting surface-level visual scenes. To bridge this gap, we introduced Hospitality In- formativeness, a domain-grounded framework that formalizes four fundamental visual axes—spatial legibility, activity affordance, contextual open- ness, and geometric completeness, whom are in- terpretable and measurable.Building on this framework, we constructed Hospitality-VQA, a decision-centric VQA benchmark designed to elicit and evaluate the kinds of visual evidence that influ- ence guest perception across diverse facility types. E.g., whether models capture layout, functional components, scenery, and exterior visibility that matter for booking decisions. Together, these con- tributions provide the first structured basis for mea- suring how well VLMs interpret hospitality im- agery beyond generic scene recognition. Our empirical study revealed that state-of-the- art general-purpose VLMs struggle with the fine- grained informativeness reasoning that the hos- pitality domain demands.However, we also showed that lightweight domain adaptation using our dataset leads to consistent and measurable im- provements, highlighting both the challenge of the task and the value of the benchmark as a foundation for future model development. Future Directions Looking ahead, Hospitality- VQA and the hospitality informativeness frame- work open several research directions, including domain-aware representation learning, prompt opti- mization, and test-time reasoning strategies. A par- ticularly promising extension is modeling human- preferred accommodation attractiveness, as user impressions are often shaped by images. This line of work carries clear practical value: B2C applications include displaying more appealing im- ages to improve user experience and booking rates, while B2B applications involve curating and rank- ing property images based on user appeal. We hope our benchmark provides a foundation for future advances in hospitality-aware multimodal intelligence that benefits both users and service providers. Limitations This work has several limitations. First, Hospitality- VQA focuses on static images collected from a spe- cific set of hotels and platforms, and the proposed informativeness axes represent a pragmatic but nec- essarily incomplete abstraction of real-world user information needs. In particular, while our frame- work emphasizes functional, spatial, and contextual visual cues, it does not explicitly capture aesthetic qualities such as visual style, ambiance, or emo- tional appeal, which can also influence user prefer- ences in hospitality settings. Second, our study does not model additional modalities or contextual factors commonly in- volved in accommodation decisions, such as textual reviews, pricing information, temporal media (e.g., videos), or personalized user preferences. As a re- sult, the evaluation is limited to image-based visual reasoning under a controlled decision setting. Third, all model evaluations are conducted un- der a single annotation protocol and question for- mulation. We do not assess the robustness of the reported results under alternative labeling schemes, prompt designs, or downstream task definitions. Finally, although the dataset contains 5,000 an- notated images in total, quantitative evaluation is performed on a held-out subset of 300 images. This relatively small evaluation set may limit statistical power and reduce sensitivity to rare or long-tail cases. Acknowledgments The views and conclusions expressed in this paper are those of the authors and should not be inter- preted as representing the official views, policies, or products of their affiliated organization. References Arian Askari, Emmanouil Stergiadis, Ilya Gusev, and Moran Beladev. 2025. Hotelmatch-llm: Joint multi- task training of small and large language models for efficient multimodal hotel retrieval. arXiv preprint arXiv:2506.07296. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wen- bin Ge, Sibo Song, and 1 others. 2025. Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Mar- cel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261. Pedro Cuesta-Valiño, Sergey Kazakov, Pablo Gutiérrez- Rodríguez, and Orlando Lima Rua. 2023. The effects of the aesthetics and composition of hotels’ digital photo images on online booking decisions. Humani- ties and Social Sciences Communications, 10:59. Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. 2018. Embodied question answering. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR). Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Michelle R. Greene, Christopher Baldassano, Andre Esteva, Diane M. Beck, and Li Fei-Fei. 2016. Vi- sual scenes are categorized by function. Journal of Experimental Psychology: General, 145(1):82–94. Dario Guidotti, Laura Pandolfo, and Luca Pulina. 2025. Discovering sentiment insights: streamlining tourism review analysis with large language models. Infor- mation Technology & Tourism, 27(1):227–261. Danna Gurari, Quchen Li, Anthony J Stangl, Yongsen Guo, Chuan-He Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. 2018. Vizwiz: Nearly real- time answers to visual questions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Jiaxiu He, Bingqing Li, and Xin Shane Wang. 2023. Image features and demand in the sharing economy: A study of airbnb. International Journal of Research in Marketing, 40(4):760–780. Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, and 1 others. 2025. Glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning. arXiv preprint arXiv:2507.01006. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3. Drew A Hudson and Christopher D Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Aaron Hurst and 1 others. 2024. GPT-4o System Card. arXiv preprint arXiv:2410.21276. Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Ren- rui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. 2024. Llava-next: Stronger llms super- charge multimodal capabilities in the wild. Huiying Li, Qiang Ye, and Rob Law. 2013. Determi- nants of customer satisfaction in the hotel industry: An application of online review analysis. Asia Pacific journal of tourism research, 18(7):784–802. Aude Oliva and Antonio Torralba. 2001. Modeling the shape of the scene: A holistic representation of the spatial envelope. International journal of computer vision, 42(3):145–175. OpenAI. 2025. Gpt-5 system card. Meng Ren, Huy Quan Vu, Gang Li, and Rob Law. 2021. Large-scale comparative analyses of hotel photo con- tent posted by managers and customers to review platforms based on deep learning: implications for hospitality marketers. Journal of Hospitality Market- ing & Management, 30(1):96–119. Javier A. Sánchez-Torres, Sandra-Milena Palacio- López, Yuri Hernandez-Fernandez, Francisco J. Arroyo-Cañada, and Ana Argila-Irurita. 2024. Vi- sual photography’s influences on hotel selection: an analysis using e-booking as a comparative platform. International Journal of Electronic Customer Rela- tionship Management, 14(2):128–142. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, and 1 others. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. Ameni Trabelsi, Maria Zontak, Yiming Qian, Brian Jackson, Suleiman Khan, and Umit Batur. 2025. What matters when building vision language mod- els for product image analysis? In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACV Workshops). Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaek- ermann, Mohamed Amin, Pi-Chuan Chang, Andrew Carroll, Charles Lau, Ryutaro Tanno, Ira Ktena, and 1 others. 2024. Towards generalist biomedical ai. Nejm Ai, 1(3):AIoa2300138. Kirk L. Wakefield and Jeffrey G. Blodgett. 1996. The effect of the servicescape on customers’ behavioral intentions in leisure service settings. Journal of Ser- vices Marketing, 10(6):45–61. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elic- its reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837. Zheng Xiang, Zvi Schwartz, John H Gerdes Jr, and Muzaffer Uysal. 2015. What can big data and text analytics tell us about hotel guest experience and satisfaction?International journal of hospitality management, 44:120–130. Shunyuan Zhang, Dokyun Lee, Param Vir Singh, and Kannan Srinivasan. 2022.What makes a good image?airbnb demand analytics leveraging in- terpretable image features. Management Science, 68(8):5644–5666. A Detailed Facility Taxonomy To support consistent annotation and evaluation, we define a hierarchical facility taxonomy with clear operational criteria. Images are first assigned to one of five main facility types based on accessibility and visible structural boundaries: Room Interior, Indoor Facility, Outdoor Facility, Accommodation Exterior, and Irrelevant. Room Interior is restricted to private guest spaces. In cases of spatial overlap (e.g., studio-type rooms), a fixed priority order is applied (Bedroom >Kitchen>Bathroom>Living room) to ensure unique assignment. Indoor and Outdoor Facili- ties are distinguished by whether the space is fully enclosed, with outdoor facilities required to be the primary visual focus rather than part of a general landscape. Accommodation Exterior is assigned only when the building itself constitutes the main subject with identifiable accommodation features. Images lacking discernible hospitality context are grouped into the Irrelevant category, which func- tions as a noise class. For finer-grained functional analysis, each main category (except Accommodation Exterior and Ir- relevant) is further annotated with sub-facility la- bels, summarized in Table 4. This granularity en- ables evaluation of whether models can recognize specific functional contexts relevant to hospitality decision-making. Main CategorySub-category Room InteriorBedroom; Kitchen; Bathroom; Living room Indoor FacilityGuest lounge; Reception desk; Hallway; Restaurant & Cafe; In- door pool; Indoor parking lot; Other amenities Outdoor Facil- ity Outdoor pool & Spa; Outdoor lounge & BBQ area; Sports & Recreation facility; Outdoor parking lot; Camping area Accommodation Exterior — Irrelevant— Table 4: Full taxonomy of hospitality facility classifica- tion and sub-category labels. GeneralInstruction–AnswerTemplate Format Task Target classification or assessment objective. Prompt Natural language instruction defining task semantics and decision rules. Answer Format Strictly constrained output schema (e.g., class ID or fixed key–value pairs). Answer Example output following the specified format. Figure 5: General structure of instruction–answer tem- plates shared across all evaluation tasks. B Instruction-Answer Construction Template This appendix provides details on how the expert- verified labels are mapped into instruction–answer pairs using our fixed templates. As described in Section 4.3, these templates are designed to ensure consistency across the dataset by formatting classi- fication targets into a standardized VQA format. To maintain evaluation rigor, each template con- sists of a task-specific prompt and a constrained an- swer format. Figure 5 illustrates the general struc- ture of these templates. Representative examples for facility-type classification and informativeness axis evaluation are presented in Figures 6 and 7, respectively. Task: Facility Type (Main) Prompt: Your task is to classify given image. Definitions and specific instructions for each category are as follows: 1. Private accommodation room interior space for guest sleeping/living functions. Includes bedrooms, bathrooms, living rooms, kitchens, and photos taken from inside rooms. Shared areas or facilities do not belong to this category. 2. Shared "indoor" facilities within accommodation (e.g. customer lounges, reception desks, corridors, restaurants/cafes, indoor pools, indoor parking, other amenities (gyms, indoor golf, saunas, conve- nience stores, seminar rooms etc.)) 3. Specific "outdoor" facilities that falls into following cases:outdoor pools/spas, outdoor lounges/garden/terrace/BBQ areas, outdoor sports/recreation facilities, outdoor parking, outdoor camping areas. Must be clearly identifiable as one of these facility types and be the image’s primary focus, not part of general accommodation or landscape views. Exclude: overall build- ing/accommodation views even if outdoor facilities are visible, pure nature shots without specific facilities. 4. Image showing accommodation building exterior AS THE MAIN SUBJECT. Buidling must occupy significant portion of image with clear structural elements (walls, windows, roof) and typical accommodation features (guestroom windows, balconies, nearby amenities) to be identifiable as accommodation. Only for photos that do NOT fall into categories 1, 2, 3, or 5 AND where the building itself is the primary focus, not background. Excluded: no visible building, appears to be non-residential buildings due to lack of accommodation features, main accommodation building unclear among multiple scattered buildings, accommodation too small/distant to recognize (e.g. tiny in a wide drone shot). 5. Image lacking any clues identifying them as prior 4 categories of accommodation. Includes pure nature shots, pet/person-focused shots without spatial clues, notice/text-based images (e.g., posters, receipts), and building exteriors not meeting criteria of 4. For close-up images, prefer classifying to 1-4 over 5 when possible, if there are any clues suggesting accommodation context, even if subtle (e.g., cushion seems to be on bed→ 1, food seems to be in restaurant→ 2). ANSWER FORMAT Output a single number: <1-5> Do not include any explanation, spaces, or other characters. Answer: 1 Figure 6: Example of constructed instruction–answer template of main facility. Task: Geometric Completeness Prompt: You are an expert image analyst specializing in architectural assessments. Your task is to analyze the visible faces of the single most plausible and visually prominent lodging building in an image. For the selected building, output a visibility status code (1–4) for each of these three faces: - ’1’ = Front Facade - ’2’ = Side Wall - ’3’ = Roof Apply the rules in this order for each face: 1. Assign 1 if the face is absent or not visible at all. 2. Assign 2 if a clear, identifiable portion of the face is cut off by the image’s edges. 3. Assign 3 if the face is visible but significantly blocked by an external object. 4. Assign 4 if the face is clearly visible and unobstructed (roof only if distinct and unambiguous). ANSWER FORMAT Output exactly in this format, with no spaces or extra text: ’1’: <1-4>, ’2’: <1-4>, ’3’: <1-4> Answer: ’1’: 3, ’2’: 3, ’3’: 4 Figure 7: Example of constructed instruction–answer template of geometric completeness. C Additional Experimental Details We provide implementation details to support the reproducibility of the results in Section 5. All fine- tuning experiments on open-weight models were conducted on a single NVIDIA RTX 4090 GPU using the unsloth framework. C.1 Training Setup Models were fine-tuned for two epochs using su- pervised learning on image–instruction pairs. We used the AdamW optimizer with a learning rate of2× 10 −5 , 5% linear warmup, and cosine decay. The effective batch size was 16 (batch size 2 per device with gradient accumulation of 8). Training was performed in bfloat16 precision with a maxi- mum context length of 8,192 tokens. Weight decay and gradient clipping were not applied. C.2 LoRA Configuration We adopted Low-Rank Adaptation (LoRA) (Hu et al., 2022) with rankr = 16and scaling factor α = 32. Adapters were inserted into the vision encoder and language decoder, covering the atten- tion projections and MLP layers. LoRA dropout was set to 0, and all other model parameters were frozen. D CoT vs. No CoT We explicitly investigated the impact of incorpo- rating Chain-of-Thought (CoT) (Wei et al., 2022) reasoning during the supervised fine-tuning pro- cess. Table 5 presents a performance comparison between the base models, models fine-tuned with CoT supervision, and models fine-tuned with direct answers (w/o CoT). In the 3B setting, CoT supervision provides small gains on a few attributes (e.g., Scenery and Building Faces), but these improvements are nei- ther consistent across tasks nor robust across model scales. Overall, direct-answer supervision with- out CoT yields more reliable performance for our classification-oriented evaluation. ModelFacilityInformativeness Main Main&SubSLAACOGCRoom View Qwen2.5-VL-3B64.6644.3468.35 19.34 39.72 1.4741.94 70.51 Qwen2.5-VL-3B FT (w/o CoT) 86.6681.1394.96 42.92 57.45 26.4780.65 76.92 Qwen2.5-VL-3B FT (w/ CoT)85.6673.5893.53 34.91 63.38 27.9464.52 70.51 Qwen2.5-VL-7B78.6664.1543.88 25.94 48.94 5.8825.81 69.23 Qwen2.5-VL-7B FT (w/o CoT)92.0085.3797.12 44.34 67.37 32.3587.10 74.36 Qwen2.5-VL-7B FT (w/ CoT)91.3383.0294.24 42.45 59.57 26.4783.87 76.92 Table 5: Comparison of VLM performance with and without CoT. Best in each column is highlighted in bold.