HumbleBench: An Evaluation Benchmark for Cognitive Humility of Multimodal Large Language Models

HumbleBench is an evaluation benchmark for the cognitive humility of multimodal large language models (MLLMs). It fills the gap in traditional benchmarks that ignore models' self-awareness and honest expression under uncertainty, emphasizing the core value of this ability for building reliable and safe AI systems.

As MLLMs are increasingly applied in high-reliability scenarios, traditional benchmarks only focus on accuracy but ignore whether models can honestly admit their limitations when facing uncertainty or insufficient information. Cognitive humility (models' self-awareness and honest expression when confronting knowledge boundaries) has long been overlooked, and HumbleBench fills this gap.

Cognitive humility in the AI field includes three layers of meaning:

1. Self-awareness: Accurately assessing one's own level of understanding of a problem;
2. Expression of uncertainty: Appropriately expressing when information is insufficient instead of fabricating answers;
3. Boundary awareness: Clearly recognizing knowledge boundaries and not answering beyond them. This ability is key to reliable AI assistance in high-risk fields such as healthcare and law.

The core design of HumbleBench includes:

1. Multi-dimensional test scenarios: Clearly answerable questions, ambiguous/information-insufficient questions, professional knowledge questions, and multimodal information-missing scenarios;
2. Quantitative indicators: Matching degree between accuracy and confidence, rejection rate of unsolvable questions, overconfidence/underconfidence ratio, difficulty consistency;
3. Multimodal characteristics: Focus on humble expression in visual-language interaction, such as whether gaps are identified when image information is insufficient.

Cognitive humility directly affects the practicality and safety of AI:

• Avoid hallucinations: Do not fabricate false information when uncertain;
• Improve human-AI collaboration: Users can judge when manual intervention is needed;
• Risk assessment: In high-risk decisions, model reliability is more critical than accuracy;
• Continuous learning: Identifying knowledge boundaries is the foundation for targeted knowledge supplementation.

HumbleBench reflects the shift of AI research towards reliability and interpretability, which aligns with the direction of AI safety. It raises a new question: How to maintain cognitive humility while improving model capabilities? This involves technical challenges such as training data, loss functions, and post-processing calibration.

HumbleBench is an important advancement in AI evaluation, reminding intelligent systems to know when not to answer. When pursuing more powerful models, we need to pay attention to subtle yet critical abilities like cognitive humility. It provides developers with practical tools and will be more important in future applications in key fields.