IKIWISI: An Interactive Visual Pattern Generator for Evaluating the Reliability of Vision-Language Models Without Ground Truth

2025-05-28

AI Summary

Evaluating the reliability of vision-language models (VLMs) for video object recognition remains challenging in zero-annotation settings where ground-truth labels are unavailable. Method: This paper proposes the first ground-truth-free visual cognitive auditing paradigm. It introduces an interactive binary heatmap generator (green/red encoding object presence), a “spy object” mechanism to actively induce and detect model hallucinations, and a cognitive alignment analysis framework enabling user-driven adversarial inspection. Contribution/Results: In a 15-participant user study, the tool demonstrates high usability and discriminative efficacy. Human judgments correlate strongly with objective metrics (r > 0.85), and reliable assessments require inspecting only a small number of heatmap units. By formalizing human intuitive pattern recognition into a quantifiable reliability assessment methodology, this work significantly improves evaluation efficiency and enhances human-AI collaborative insight.

Technology Category

• Computer Vision: Language and Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

We present IKIWISI ("I Know It When I See It"), an interactive visual pattern generator for assessing vision-language models in video object recognition when ground truth is unavailable. IKIWISI transforms model outputs into a binary heatmap where green cells indicate object presence and red cells indicate object absence. This visualization leverages humans' innate pattern recognition abilities to evaluate model reliability. IKIWISI introduces"spy objects": adversarial instances users know are absent, to discern models hallucinating on nonexistent items. The tool functions as a cognitive audit mechanism, surfacing mismatches between human and machine perception by visualizing where models diverge from human understanding. Our study with 15 participants found that users considered IKIWISI easy to use, made assessments that correlated with objective metrics when available, and reached informed conclusions by examining only a small fraction of heatmap cells. This approach not only complements traditional evaluation methods through visual assessment of model behavior with custom object sets, but also reveals opportunities for improving alignment between human perception and machine understanding in vision-language systems.

Problem

• Evaluating vision-language models without ground truth data
• Detecting model hallucinations using adversarial spy objects
• Assessing human-machine perception alignment via visual heatmaps

Innovation

• Interactive binary heatmap for visual assessment
• Spy objects detect model hallucination instances
• Cognitive audit mechanism aligns human-machine perception