Mass-Producing Failures of Multimodal Systems with Language Models

TL;DR

This paper introduces MultiMon, a system that automatically identifies systematic failures in multimodal systems by analyzing erroneous agreement patterns using language models, revealing widespread issues in popular models like CLIP and their derivatives.

Contribution

MultiMon is a novel automated approach that detects and describes systematic failures in multimodal systems using natural language prompts and large language models.

Findings

Identified 14 systematic failure patterns in CLIP.

Failures in CLIP propagate to systems like DALL-E and Midjourney.

MultiMon can target failures relevant to specific applications.

Abstract

Deployed multimodal systems can fail in ways that evaluators did not anticipate. In order to find these failures before deployment, we introduce MultiMon, a system that automatically identifies systematic failures -- generalizable, natural-language descriptions of patterns of model failures. To uncover systematic failures, MultiMon scrapes a corpus for examples of erroneous agreement: inputs that produce the same output, but should not. It then prompts a language model (e.g., GPT-4) to find systematic patterns of failure and describe them in natural language. We use MultiMon to find 14 systematic failures (e.g., "ignores quantifiers") of the CLIP text-encoder, each comprising hundreds of distinct inputs (e.g., "a shelf with a few/many books"). Because CLIP is the backbone for most state-of-the-art multimodal systems, these inputs produce failures in Midjourney 5.1, DALL-E, VideoFusion, and others. MultiMon can also steer towards failures relevant to specific use cases, such as self-driving cars. We see MultiMon as a step towards evaluation that autonomously explores the long tail of potential system failures. Code for MULTIMON is available at https://github.com/tsb0601/MultiMon.