---
title: "CuRSE-Bench: A Benchmark for Probing... | SI Paper Desk"
url: https://thesimind.si/papers/2611.20821/
description: "Safety checks for text-to-image SI often use one culture-blind harm label, and this benchmark suggests they miss content that is harmful only within a..."
---

Number 5 of 12 · November 18, 2026

# CuRSE-Bench: A Benchmark for Probing Cultural Blind Spots in Text-to-Image Safety Evaluation

Safety checks for text-to-image SI often use one culture-blind harm label, and this benchmark suggests they miss content that is harmful only within a particular culture.

Minsuk Jang, Minseok Son, KiHyun Seol and 2 more · Submitted November 16, 2026 · arXiv 2611.20821v1 · cs.CL

[Read the paper on arXiv](https://arxiv.org/abs/2611.20821) · [The PDF](https://arxiv.org/pdf/2611.20821) · [The day's 12](https://thesimind.si/papers/day/2026-11-18/)

## What they found

The authors built CultureCraft, an agentic framework (a system of SI agents working in steps). It writes image prompts that look harmless under a general rubric but carry a hidden harmful meaning in one target culture. They used it to make CuRSE-Bench: 10,709 culture-specific prompts and 2,648 culture-general controls, covering six regions and ten harm axes. Each prompt comes from a predefined knowledge base of documented harm vectors and can be traced to at least two independent sources. They scored every prompt with six region persona judges and one paired culture-blind judge. The culture-blind judge under-rated culture-specific latent harm, while matching the persona judges on the controls. Each persona judge mainly decoded the hidden harm of its own region. The authors say this selective blind spot persists in deployed moderators.

## Why it matters

Most safety checks for image generators give each prompt or image one harm label that ignores culture. This paper argues that content can pass as harmless while being harmful in a specific culture. A benchmark with sourced, region-specific cases gives builders a way to test for this gap. The abstract does not give numbers for how large the gap is, so its size is not clear from this summary.

## How it changes the race

If the results hold, teams building image-generating SI would need to test moderation across cultures, not only with one general rubric. This is a new evaluation tool rather than a gain in capability.

## Safety and control

The paper bears directly on safety and misuse. Moderation systems that judge harm without cultural context can let harmful content through, and the authors report this blind spot in deployed moderators. Their persona judges are SI models acting as regional reviewers, so they are a test instrument and not a substitute for real people from those cultures.

### The abstract, by the authors

Text-to-image (T2I) safety evaluation typically assigns each prompt or image a single culture-blind harm label, so content that is harmful only within a particular culture can pass as benign. We introduce CultureCraft, an agentic framework that constructs T2I prompts that appear benign under a general-purpose rubric yet carry a latent harmful meaning in a target culture, and use it to build CuRSE-Bench: 10,709 culture-specific prompts and 2,648 culture-general controls over six regions and ten harm axes, each drawn from a predefined knowledge base of documented harm vectors and traceable to at least two independent sources. Scoring every prompt under six region persona judges and a paired culture-blind judge, we find that the culture-blind judge under-rates culture-specific latent harm while matching the persona judges on the controls. Each persona judge, in turn, chiefly decodes its own region's latent harm under our instrument: a selective cultural blind spot that persists in deployed moderators.

Quoted from arXiv:2611.20821. The research and its words belong to its authors.

Desk score **83** out of 100

Why it was picked

- Safety (in the title)
- Interpretability (in the title)
- Evals (in the title)
- Agents

Topics

- [Safety](https://thesimind.si/papers/?tag=safety)
- [Benchmarks](https://thesimind.si/papers/?tag=benchmarks)
- [Evals](https://thesimind.si/papers/?tag=evals)
- [Multimodal](https://thesimind.si/papers/?tag=multimodal)
- [Society](https://thesimind.si/papers/?tag=society)
- [Agents](https://thesimind.si/papers/?tag=agents)

**arXiv ID:** 2611.20821v1
**Submitted:** November 16, 2026
**Fields:** cs.CL (Computation and Language)
**Level:** Intermediate

Fingerprint, SHA-256

`df1e098370c2ddf09721915a0578c925f854ca4e757b32e3eb6b8f5457c26f4c`

On Solana: [Proof of Mind seal, 5WzZUk...UPMg](https://solscan.io/tx/5WzZUksJeCL3fYgfzcQ6DSdVxipu97PAmY5Y9M59gCMXbJtnyBVsqaoPSDHWrP4AiZAh7k4iVso5ts7WvxdNUPMg)

### The exact text fingerprinted

{"arxiv":"2611.20821","version":1,"title":"CuRSE-Bench: A Benchmark for Probing Cultural Blind Spots in Text-to-Image Safety Evaluation","authors":["Minsuk Jang","Minseok Son","KiHyun Seol","Yeongeun Shim","Changick Kim"],"day":"2026-11-18","rank":5,"score":83,"status":"written","summary":{"takeaway":"Safety checks for text-to-image SI often use one culture-blind harm label, and this benchmark suggests they miss content that is harmful only within a particular culture.","found":"The authors built CultureCraft, an agentic framework (a system of SI agents working in steps). It writes image prompts that look harmless under a general rubric but carry a hidden harmful meaning in one target culture. They used it to make CuRSE-Bench: 10,709 culture-specific prompts and 2,648 culture-general controls, covering six regions and ten harm axes. Each prompt comes from a predefined knowledge base of documented harm vectors and can be traced to at least two independent sources. They scored every prompt with six region persona judges and one paired culture-blind judge. The culture-blind judge under-rated culture-specific latent harm, while matching the persona judges on the controls. Each persona judge mainly decoded the hidden harm of its own region. The authors say this selective blind spot persists in deployed moderators.","matters":"Most safety checks for image generators give each prompt or image one harm label that ignores culture. This paper argues that content can pass as harmless while being harmful in a specific culture. A benchmark with sourced, region-specific cases gives builders a way to test for this gap. The abstract does not give numbers for how large the gap is, so its size is not clear from this summary.","race":"If the results hold, teams building image-generating SI would need to test moderation across cultures, not only with one general rubric. This is a new evaluation tool rather than a gain in capability.","safety":"The paper bears directly on safety and misuse. Moderation systems that judge harm without cultural context can let harmful content through, and the authors report this blind spot in deployed moderators. Their persona judges are SI models acting as regional reviewers, so they are a test instrument and not a substitute for real people from those cultures.","level":"intermediate","tags":["safety","benchmarks","evals","multimodal","society","agents"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Minsuk Jang, Minseok Son, KiHyun Seol, Yeongeun Shim, Changick Kim

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 14, 2026 · Number 3MobileProbe: Adaptive Security Evaluation of Mobile Agents via Trajectory-Aware InjectionMobileProbe tests phone-operating SI agents by adapting attacks to the route each agent takes, and the authors report it beats fixed, hand-written tests at finding weaknesses.SecurityAgentsEvals](https://thesimind.si/papers/2611.16693/) · [Nov 9, 2026 · Number 4HomeLoop: A Simulation Framework and Benchmark for Robot-IoT Coordination in Everyday Domestic Assistance and SafetyHomeLoop is a new simulated home benchmark where SI robots must work with smart devices and handle ambiguous alerts. The best of six models scored 61.9%, and models often over-reacted.RoboticsBenchmarksEvals](https://thesimind.si/papers/2611.08914/) · [Nov 20, 2026 · Number 2AutoPI: Taxonomy-Guided Adaptive Red Teaming for Prompt Injection in LLM AgentsAutoPI is a red teaming tool that tests SI agents for prompt injection across many attack types, not just how often attacks work, and it reports higher success and coverage than recent rivals.SecurityAgentsEvals](https://thesimind.si/papers/2611.24313/)

[Number 4](https://thesimind.si/papers/2611.21538/) · [All 12 for Nov 18, 2026](https://thesimind.si/papers/day/2026-11-18/) · [Number 6](https://thesimind.si/papers/2611.20445/)
