Research map / Vision-language and multimodal / Multimodal LLMs

World in a Frame: Understanding Culture Mixing as a New Challenge for Vision-Language Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CultureMix, a food VQA benchmark of ~23k diffusion-generated culture-mixing images (food+food, food+background), evaluates 10 LVLMs, and tests prompting and SFT mitigations.

Problem

LVLMs' ability to recognize and preserve individual cultural identities when cultural items from multiple regions (foods and backgrounds) co-occur in one scene is underexplored.

Approach

Defines culture mixing as a task and builds a diffusion-edited benchmark that systematically varies distractor type (food, background, both) and cultural distance, then measures accuracy, entropy and label shift relative to single-food baselines; also tests prompting and SFT mitigations.

Key results

• Across 10 LVLMs, country accuracy drops by 1-14 points and food accuracy by 1-8 points when distractors are added; background distractors (SFB) shift predictions more than food distractors.
• SFT on 5K culture-mixing images raised Ovis2.5-9B MFB accuracy from 6.14% to 8.95% and InternVL3-8B MFB accuracy from 3.33% to 5.14%, with significant entropy reduction.
• Gemini-2.5-Pro and GPT-5 outperform open models; Gemini achieves 0.457 country accuracy on single food vs 0.286 on SFB.

Surprising findings

• Chain-of-thought prompting sometimes lowers label consistency and amplifies reliance on background cues.
• Higher-accuracy Gemini shows larger label shifts toward distractor countries than GPT-5.

Limitations

• Focuses only on food and scene cues, not festivals, clothing, or daily-life elements.
• Low-resource and underrepresented cultures are underexplored; 30 countries only.
• Food accuracy relies on a character n-gram similarity threshold, and country-level labels are coarse proxies for culture.

Experimental setup

Datasets: WorldCuisines WorldWideDishes VIPPGeo Google Landmarks Dataset v2

Benchmarks: CultureMix CultureMix-Real

Metrics: food name accuracy country accuracy entropy label shift (%)

Baselines: Gemini-2.5-Pro GPT-5 InternVL3-78B QwenVL3-32B Ovis2.5-9B Molmo-72B

Builds on: FLUX.1-Kontext Qwen-Image-Edit WorldCuisines WorldWideDishes

Closest papers by meaning

• No Filter: Cultural and Socioeconomic Diversity in Contrastive Vision-Language Models (NeurIPS 2024 · similarity 0.90)
• See It from My Perspective: How Language Affects Cultural Bias in Image Understanding (ICLR 2025 · similarity 0.89)
• Culture Affordance Atlas: Reconciling Object Diversity Through Functional Mapping (AAAI 2026 · similarity 0.88)
• CVQA: Culturally-diverse Multilingual Visual Question Answering Benchmark (NeurIPS 2024 · similarity 0.88)
• All Languages Matter: Evaluating LMMs on Culturally Diverse 100 Languages (CVPR 2025 · similarity 0.88)
• Culture In a Frame: C$^3$B as a Comic-Based Benchmark for Multimodal Culturally Awareness (ICLR 2026 · similarity 0.88)