WORLDVIEW Benchmark Detects Cultural Bias in Prompt Revision Layers

The WORLDVIEW benchmark audits how systems like DALL-E-3 and Imagen-4 silently rewrite user prompts. Findings show that non-Western and non-Anglophone contexts undergo heavy marking and vocabulary flattening compared to US-based English baselines.