AI Models Prefer Guessing Over Asking for Help When Information Missing, Research Shows

Research using ProactiveBench, a benchmark testing framework, reveals that nearly all of 22 tested multimodal language models fail to ask users for help when visual information is missing, instead defaulting to guessing answers. The testing specifically evaluated whether these AI models would proactively request needed information from users when faced with incomplete visual data. This behavior pattern represents a significant limitation in current multimodal AI systems' ability to recognize and communicate their own knowledge gaps. However, researchers found that simple reinforcement learning techniques can improve this help-seeking behavior, suggesting a pathway to making AI models more transparent about their limitations.

AI Models' Failure to Request Missing Information

• ▪Out of 22 multimodal language models tested, almost none ask for what they need when information is missing.
• ▪AI models prefer guessing over asking for help when information is missing.
• ▪Multimodal language models almost never ask users for help when visual information is missing.

ProactiveBench Testing Methodology and Results

• ▪ProactiveBench tested 22 multimodal language models.

Potential Solutions Through Reinforcement Learning

• ▪A simple reinforcement learning approach hints at a fix for multimodal language models' failure to ask for missing information.

Perspective of AI safety researchers

• ▪Reinforcement learning approaches offer a potential solution to improve help-seeking behavior in multimodal language models.
• ▪Multimodal language models' tendency to guess when visual information is missing represents a reliability problem requiring correction.

1 source