TL;DR

Q-Bench introduces two datasets — LLVisionQA and LLDescribe — that expose a systematic gap between current multimodal LLMs and human performance on low-level visual quality perception and distortion reasoning.

Summary

Multimodal LLMs perform well on high-level semantic tasks but struggle with low-level visual perception: quality assessment, distortion recognition, and comparative quality judgment. Q-Bench introduces LLVisionQA (2,990 images with open-ended and yes/no questions) and LLDescribe (499 images for quality description generation) to evaluate this capability. Evaluated MLLMs score substantially below human performance on fine-grained perceptual tasks, revealing that high benchmark scores on semantic understanding do not imply competence in perceptual quality reasoning.

Key contributions

  1. Introduces LLVisionQA, a 2,990-image dataset targeting low-level quality perception with open-ended and binary questions.
  2. Introduces LLDescribe, a 499-image dataset for evaluating quality-aware descriptive caption generation.
  3. Demonstrates a systematic gap between current MLLMs and human performance on low-level perceptual tasks.
  4. Establishes distortion recognition and comparative quality judgment as underexplored MLLM capability dimensions.

When to cite

  1. When evaluating multimodal LLMs on perceptual quality tasks rather than semantic understanding.
  2. When arguing that existing vision benchmarks fail to capture low-level visual perception capabilities.
  3. When discussing the need for quality-aware training supervision to improve MLLM perceptual reasoning.
  4. When benchmarking foundation models on image quality assessment or distortion recognition tasks.

AI LLM ML GPT