Vision language models struggle with tricky or misleading visual questions
Evaluating VQA in Vision Language Models using Cooperative Principles
Computation and Language
Summary
Some AI models that answer questions about images have trouble when the questions include confusing, irrelevant, or false details. The authors studied several popular AI systems and found that their performance drops when these tricky questions break common conversational rules. They also compared how humans and these AI systems handle such problematic questions differently. Interestingly, people find it easier to solve mistakes made by the AI than those created by humans, but the AI itself does worse on its own errors.
What this means in practice
- •For ai system developers: Improve the robustness of vision language models against ambiguous or misleading visual questions.
- •For user experience designers: Design visual question answering interfaces that minimize user confusion by detecting questions that violate conversational norms.
Authors
Monika Shah, Sudarshan Balaji, Somdeb Sarkhel, Sanorita Dey, Deepak Venugopal
Abstract
We evaluate the performance of Vision Language Models in Visual Question Answering (VQA) when questions violate Grice's maxims. To do this, we use VLMs to generate question modifiers that add non-essential, ambiguous or false information and show that in the presence of such violations, the VLMs that we evaluate (ChatGPT, Claude, Gemini and Llava) show diminished performance. Further, we empirically show the difference between how humans reason pragmatically compared to VLMs, and the difference in VLM reasoning when it resolves violations that are human-induced compared to those that are AI-generated. Finally, we show that human cognitive effort (measured through time-on-task in an experiment) is lower for resolving VLM-induced violations, but VLMs themselves perform less accurately in such cases.