Researchers have discovered that asking vision-language models about a director's perspective before an action can alter the model's subsequent choices, even if the initial report of the director's view is correct. This phenomenon was observed across several GPT models and Qwen3.8-27B, where the prompt's phrasing shifted the model's interpretation from the director's viewpoint to its own. To address this, a new evaluation method is proposed, using paired prompts that differ only by the inclusion of this perspective-eliciting question, providing a baseline for accurate assessment. AI
IMPACT Highlights a subtle but significant bias in how vision-language models interpret instructions, potentially impacting their reliability in real-world applications.
RANK_REASON Academic paper detailing a novel finding about the behavior of vision-language models.
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
