A Good CREPE needs more than just Sugar: Investigating Biases in Compositional Vision-Language Benchmarks

TL;DR

This paper critically examines 17 vision-language benchmarks for compositional understanding, revealing inherent biases and limitations in current evaluation methods, and offers recommendations for more robust benchmark construction.

Contribution

It uncovers biases in existing benchmarks and demonstrates that simple heuristics can outperform models, highlighting the need for improved evaluation procedures.

Findings

Blind heuristics perform on par with CLIP models

Distribution asymmetry causes benchmark biases

Recommendations for constructing more robust benchmarks

Abstract

We investigate 17 benchmarks (e.g. SugarCREPE, VALSE) commonly used for measuring compositional understanding capabilities of vision-language models (VLMs). We scrutinize design choices in their construction, including data source (e.g. MS-COCO) and curation procedures (e.g. constructing negative images/captions), uncovering several inherent biases across most benchmarks. We find that blind heuristics (e.g. token-length, log-likelihood under a language model) perform on par with CLIP models, indicating that these benchmarks do not effectively measure compositional understanding. We demonstrate that the underlying factor is a distribution asymmetry between positive and negative images/captions, induced by the benchmark construction procedures. To mitigate these issues, we provide a few key recommendations for constructing more robust vision-language compositional understanding benchmarks, that would be less prone to such simple attacks.