Drishtikon: Indian Multimodal Cultural Benchmark

• DRISHTIKON is a multimodal, multilingual benchmark that assesses AI's cultural understanding of India through 64,000 aligned text-image pairs covering festivals, attire, cuisines, art, and heritage.
• It employs both zero-shot and chain-of-thought evaluation protocols to reveal the limitations of vision-language models in interpreting region-specific cultural cues, especially for low-resource languages.
• The benchmark emphasizes granular, India-specific cultural insights, challenging models to move beyond generic linguistic competence toward authentically grounded cultural reasoning.

Searching arXiv for DRISHTIKON and related culturally grounded multimodal benchmarks. DRISHTIKON is a multimodal and multilingual benchmark centered exclusively on Indian culture and designed to evaluate the cultural understanding of generative AI systems. It was introduced as “a first-of-its-kind multimodal and multilingual benchmark centered exclusively on Indian culture,” with coverage spanning 15 languages, all states and union territories, and “over 64,000 aligned text-image pairs.” Its stated purpose is to test whether models, especially vision-LLMs (VLMs), can identify, reason over, and ground culturally salient visual and textual information tied to India’s regional, linguistic, and historical diversity. Reported evaluations cover zero-shot and chain-of-thought settings and expose limitations in current models’ ability to reason over culturally grounded multimodal inputs, particularly for low-resource languages and less-documented traditions (Maji et al., 23 Sep 2025).

1. Definition and intellectual setting

DRISHTIKON was developed against the background of a broader distinction between multilingual capability and cultural competence. In that framing, strong multilingual performance does not by itself imply reliable understanding of local norms, practices, entities, or visual cues. DRISHTIKON addresses that gap by operationalizing Indian culture through multimodal evidence rather than through text-only questioning. The benchmark targets domains such as festivals, attire, cuisines, art forms, and historical heritage, where models often default to generic or Western-centric interpretations when local context is required (Nezhad, 27 Mar 2026).

Within culturally grounded NLP, DRISHTIKON is positioned as a multimodal counterpart to text- and value-focused benchmarks such as Global-MMLU, CDEval, WorldValuesBench, and CulturalBench. That positioning is methodologically important: it treats culture as distributed across modalities, not confined to lexical or propositional content. A plausible implication is that DRISHTIKON shifts evaluation away from abstract cultural recall toward grounded interpretation of artifacts, rituals, food, clothing, and region-specific visual entities that appear in real communicative ecologies (Nezhad, 27 Mar 2026).

2. Scope, coverage, and represented cultural domains

The benchmark’s reported scope is unusually deep within a single national-cultural context. It spans 15 languages, covers all states and union territories, and contains over 64,000 aligned text-image pairs. Its thematic coverage includes festivals, attire, cuisines, art forms, and historical heritage, with the explicit aim of capturing India’s regional diversity at fine granularity rather than offering thin cross-country coverage (Maji et al., 23 Sep 2025).

Aspect	Reported coverage
Languages	15 languages
Geography	All states and union territories
Modality	Aligned text-image pairs
Scale	Over 64,000 pairs
Themes	Festivals, attire, cuisines, art forms, historical heritage

This design distinguishes DRISHTIKON from global benchmarks that sample many cultural contexts at lower density. The survey literature emphasizes that such depth can expose failures that broader but thinner evaluations miss, especially when models must connect regional language varieties with local visual cues. The benchmark’s India-specific concentration therefore functions not as a limitation of ambition but as a deliberate strategy for fine-grained stratification across language, region, and cultural facet (Nezhad, 27 Mar 2026).

At the same time, several compositional details are not specified in the available summaries. The survey does not report train/dev/test splits, class taxonomies, item counts beyond the “over 64,000” figure, or detailed per-theme distributions. It also does not provide collection provenance, annotation guidelines, translation pipelines, or licensing information. Those omissions matter because they constrain reproducibility and make it harder to audit ecological validity or community representation (Nezhad, 27 Mar 2026).

3. Evaluation design and protocol dimensions

DRISHTIKON evaluates “a wide range of vision-LLMs (VLMs), including open-source small and large models, proprietary systems, reasoning-specialized VLMs, and Indic-focused models,” and does so in both zero-shot and chain-of-thought settings. This evaluation design is intended to test whether scale, specialization, or reasoning-oriented prompting materially improve culturally grounded multimodal understanding in the Indian context (Maji et al., 23 Sep 2025).

The broader literature synthesized alongside DRISHTIKON identifies several protocol variables that are especially relevant for interpreting results on this benchmark. These include prompt language, tokenizer behavior, translated benchmark design, culturally specific supervision, and multimodal context. Although the survey does not report DRISHTIKON-specific prompt templates or tokenization analyses, it emphasizes that such factors materially affect outcomes across culture-focused tasks, particularly for Indic scripts and low-resource varieties. This suggests that DRISHTIKON is not only a benchmark for model comparison but also a probe of infrastructural choices in multilingual modeling, such as script coverage and prompt-language selection (Nezhad, 27 Mar 2026).

The exact scoring metrics and formal evaluation equations for DRISHTIKON are not specified in the survey. Consequently, the available literature foregrounds protocol contrasts and qualitative failure modes rather than benchmark-specific formulas. For an arXiv-reading audience, that absence is itself a relevant methodological fact: the benchmark’s conceptual contribution is clear from the reported summaries, but the full operationalization of evaluation remains only partially visible in those summaries (Nezhad, 27 Mar 2026).

4. Empirical findings and characteristic failure modes

The principal reported empirical finding is concise and consequential: DRISHTIKON’s results “expose key limitations in current models’ ability to reason over culturally grounded, multimodal inputs, particularly for low-resource languages and less-documented traditions.” In other words, performance degrades precisely where local specificity is highest and where data coverage is weakest, indicating that generic multilingual competence does not robustly transfer to culturally dense Indian settings (Maji et al., 23 Sep 2025).

The survey identifies several recurrent failure modes associated with DRISHTIKON and adjacent multimodal resources. These include cultural misreadings in visual contexts, especially when the cues are regionally specific; performance disparities tied to language resourcing, with lower-resource Indian languages faring worse; and fragility on less-documented traditions. The same synthesis argues that training data coverage is a strong determinant of performance but is necessary rather than sufficient. Even when models have broad multilingual exposure, they may still flatten local norms, miss region-specific entities, or misinterpret artifacts and rituals when multimodal grounding is required (Nezhad, 27 Mar 2026).

A broader comparative statement from the survey is especially relevant: CulturalVQA, GIMMICK, WorldCuisines, DRISHTIKON, and CaMMT all show that models struggle with geographically diverse food, dress, artifacts, rituals, and region-specific visual cues. DRISHTIKON’s contribution within that pattern is to demonstrate the same phenomenon at greater depth inside one cultural ecology, thereby making internal Indian disparities more visible than a global benchmark typically can (Nezhad, 27 Mar 2026).

5. Relation to adjacent benchmarks and methods

DRISHTIKON is best understood not as an isolated dataset but as part of a recent cluster of culturally grounded evaluation resources. Its distinctiveness lies in its combination of multimodality, multilingualism, India-specific focus, and region-level depth.

Resource	Core focus	Relation to DRISHTIKON
Global-MMLU	Multilingual MMLU with culturally sensitive subsets	DRISHTIKON is multimodal and region-specific
CDEval / WorldValuesBench	Value-focused, text-only cultural alignment	DRISHTIKON adds grounded visual reasoning
CulturalBench	Human-authored globally diverse textual questions	DRISHTIKON provides visual grounding and local depth
CulturalVQA	2,378 image-question pairs across 11 countries	DRISHTIKON is larger and India-specific
GIMMICK	Global multimodal benchmark over 144 countries	DRISHTIKON narrows scope for fine-grained Indian coverage
WorldCuisines	Massive multilingual VQA on cuisines	DRISHTIKON covers broader cultural facets beyond food
CaMMT	Multimodal translation with culturally specific items	DRISHTIKON targets cultural understanding rather than translation

These comparisons clarify the benchmark’s methodological niche. Relative to Global-MMLU, DRISHTIKON avoids reliance on imported educational priors from English-centered curricula and instead focuses on local visual artifacts and practices. Relative to CDEval and WorldValuesBench, it probes grounded perception and reasoning rather than value statements alone. Relative to CulturalVQA and GIMMICK, it sacrifices geographic breadth for within-country density. Relative to WorldCuisines, it expands beyond cuisine into attire, festivals, art forms, and heritage. Relative to CARE and CLCA, which address cultural alignment through methods or resources, DRISHTIKON primarily serves as an evaluation substrate for testing whether adaptation improves multimodal cultural reasoning in Indian contexts (Nezhad, 27 Mar 2026).

6. Methodological significance, recommended use, and disambiguation

The survey literature treats DRISHTIKON as exemplary of a shift from benchmark design organized around language labels alone toward what it calls communicative ecologies: the institutions, scripts, domains, modalities, and communities through which language is actually used. In that agenda, DRISHTIKON supports culturally stratified evaluation by language, region, and theme; richer contextual metadata; participatory alignment with native raters; analysis of within-language variation; and multimodal community-aware design. Recommended uses include testing zero-shot against chain-of-thought settings, comparing English prompts with local Indic-language prompts, auditing tokenizer coverage for Indic scripts, and reporting stratified performance across cultural themes and resource conditions (Nezhad, 27 Mar 2026).

Several pitfalls are also identified. One is assuming that multilingual fluency implies cultural competence; DRISHTIKON is cited precisely as evidence against that assumption. Another is relying on translated, English-source evaluations that may import external assumptions. A third is ignoring within-language variation, including scripts, dialects, registers, and code-mixing. The survey further recommends complementing closed-form scores with open-ended explanations and native-speaker judgments where possible, because culturally plausible alternatives and uncertainty may not be visible in a single accuracy-oriented summary (Nezhad, 27 Mar 2026).

The name DrishtiKon/Drishtikon is also used by unrelated research artifacts. “DrishtiKon: Multi-Granular Visual Grounding for Text-Rich Document Images” concerns OCR-based visual grounding for multilingual document images and derives a benchmark from CircularsVQA (Kasuba et al., 26 Jun 2025). “Drishtikon: An advanced navigational aid system for visually impaired people” describes a stereo-vision and Google Directions-based assistive navigation prototype for visually impaired users (Kotyan et al., 2019). These works are distinct from DRISHTIKON as a benchmark for testing LLMs’ understanding of Indian culture.

In summary, DRISHTIKON defines a culturally grounded evaluation problem in which multilingual and multimodal competence must be demonstrated together rather than inferred from one another. Its reported scale, India-wide regional coverage, and thematic focus on culturally salient domains make it a significant benchmark for studying how VLMs handle local knowledge, visual grounding, and low-resource linguistic contexts. The available evidence places it at the center of a broader methodological transition toward deeper, region-specific, multimodal evaluation of cultural understanding in AI systems (Maji et al., 23 Sep 2025).