Blind-Spots-Bench: Evaluating Blind Spots in Multimodal Models

2026-07-09

AI Summary

Existing benchmarks struggle to uncover systematic blind spots in multimodal AI models on tasks humans find trivial—such as manipulating a rope or drawing a dog with five legs. To address this gap, this work introduces Blind-Spots-Bench, the first benchmark explicitly focused on “human-easy but AI-hard” challenges. It comprises 235 carefully curated and structurally annotated samples originally proposed by students, accompanied by a novel task taxonomy tailored to these failure modes. The study also establishes an automated evaluation pipeline applicable to language, vision-language, and image generation models. Experimental results reveal that leading closed-source models outperform open-source counterparts by approximately 10% on average; however, no single model dominates across all task categories, and certain challenges remain notably difficult for all evaluated systems.

Abstract

Modern AI models achieve strong performance on many established benchmarks, yet they still fail on tasks that humans find almost trivial, such as manipulating a string or drawing a dog with five legs. These examples suggest that existing benchmarks may under-measure persistent blind spots in current systems. We introduce $\texttt{blind-spots-bench}$, a benchmark designed to expose such blind spots through tasks that appear simple for humans but remain challenging for modern AI. We collect raw questions from students in an AI course, clean and annotate them with structured reference solutions, and propose a task taxonomy tailored to the resulting dataset of 235 samples. We further develop an automated grading pipeline to evaluate a wide range of models, including open-weight and closed-source language, vision-language, and image-generation models. Our analysis on $\texttt{blind-spots-bench}$ reveals that closed-source frontier models can substantially outperform open-weight models with even $\approx10\%$ gap, even when they attain comparable performance on existing benchmarks. A more fine-grained analysis shows that no single model dominates across all task types, and that some tasks remain challenging for all evaluated models. These results highlight the value of $\texttt{blind-spots-bench}$ as a diagnostic stress test for identifying concrete weaknesses in current modern models.

Problem

• blind spots
• multimodal models
• benchmark
• AI evaluation
• model weaknesses

Innovation

• blind spots
• multimodal benchmark
• automated evaluation
• task taxonomy
• model diagnostics