IC-001Vision-language models perform near chance on the NL-Eye visual abductive reasoning benchmark

Mor Ventura, Michael Toker, Nitay Calderon, Zorik Gekhman, Yonatan Bitton, Roi Reichart

SourceNL-Eye: Abductive NLI For Images

The paper evaluates several vision-language models on NL-Eye, a new benchmark for visual abductive reasoning. The task involves selecting the more plausible hypothesis image given a premise image. Humans achieve 83-85% accuracy, while most VLMs perform at or below random baselines. For example, in the triplet setup with separate images, Gemini-1.5-Pro gets 51%, GPT-4-Vision 46%, and Claude-Sonnet-3.5 49%, with the dumb pixel baseline at 50%. Performance is sensitive to hypothesis order and input format, indicating a lack of robust reasoning.

Evidence
correlational
Key metric
Human accuracy 83-85%; Gemini-1.5-Pro 51%, GPT-4-Vision 46%, GPT-4o 16%, Claude-Sonnet-3.5 49% in separate-images triplet setup; Dumb pixel baseline 50%
Caveat
The benchmark is new and synthetic, which may not fully reflect real-world visual reasoning; performance varies significantly with input strategy and model version.
Model
Gemini 1.5 / Gemini Pro 1.5 Gemini 1.5 Pro, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report GPT-4V / GPT-4 vision, GPT-4o, Claude 3.5 Sonnet, Claude 3 Opus, LLaVA-NeXT / LLaVA 1.6, Fuyu, MiniCPM-V 2.6, LLaVA-OneVision Qwen2 7B OV
Concepts
Failure mode
Datasets
NL-Eye [eval]
Related work
SNLI-VE [context]
Extraction
automatic-extraction